Interventions · countable events

When a commit leaves its lane, extra work triggers — automatically

Every commit here gets a deterministic, LLM-free drift receipt (the mechanism). When a receipt lands out of lane (off-lane share above the 15% threshold), that is one countable event — and it triggers an intervention: a sensemaking pass explains the drift, the fixes get checklisted, and the story is published here. Countable events are what make a lane priceable; this page is the count. Each entry is a structural alert, not a scolding — a diagnosis, not a verdict — and the ledger you are reading is a loss ratio, published by the house that carries it. Stated exceptions are what surplus looks like in print.

Out-of-lane receipts per day

The interventions — why · fix · checklist

1. 2026-08-29 · 📐 architect · offPct 32 · mistagged

   fix(chat): "detected, harvested nothing" now says BY WHOM

   Why it left the lane: The diff is 100% inside scripts/chat/gemini-pulse-tick.sh and tests/chat/no-turns-receipt-names-the-harvester.test.mjs — vocabulary like BEAT_STARTED, _receipt "probe-changed-no-turns", -nt, and 'scraped-by-another-process-this-beat'. None of that touches 📐 architect's declared owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). The commit's own Relevant-Rooms trailer already names 🧭 navigator and 🎩 operator alongside architect — the pull toward the out-of-lane coords is real content pull (pipeline/detection-latency vocabulary), not noise; the receipt is reading it correctly, the lane tag is what's wrong.

   The semantic pull: prewalkOverlap=5 is very low, meaning even before the recursive walk ran, the gzip-NCD reef seed barely overlapped architect's lane vocabulary against this commit's terms — the seed placement itself already diverged. litIntent=130/walksIntent=130 and litReality=139/walksReality=139 both sit at ~100% fill, so the walk didn't discriminate further, it just carried the seed's mismatch forward at full saturation.

   Intent–reality gap: The commit message claims (and Originating-Terminal states) this is 📐 architect's work; the receipt measured the actual diff living entirely in scripts/chat/ and tests/chat/, a surface architect's owned-glob list does not include. The code itself is correct and well-guarded — the lane label on it is the thing that's wrong.

   • Add scripts/chat/** and tests/chat/** to the owned-surface glob of whichever room's .workflow/rooms/*.html actually claims gemini-pulse/elicitation infra (🧭 navigator per this commit's own Relevant-Rooms trailer) so future commits in this path stop scoring against architect's strategy-only surface.
   • Confirm via `node scripts/ops/rules-for.mjs scripts/chat/gemini-pulse-tick.sh tests/chat/no-turns-receipt-names-the-harvester.test.mjs` which room(s), if any, currently claim this surface — if none do, that's an unowned-surface gap the /recontext skill should close, not just this one commit.
   • Patch the drift-receipt's room-lane resolver (wherever it currently reads 'room' for the offPct computation) to prefer Relevant-Rooms glob-match over Originating-Terminal.
   • Add the guard named in missing_guard in the same commit as the resolver patch (CODE DEFINITION-OF-DONE).
   • Re-run this exact commit (fb307c4af9) through the patched resolver as a regression fixture and confirm offPct drops toward its navigator/operator-lane figure.

   What we learned: This room's 103 out-of-lane events in 14 days and its climbing offPct trend (14→28, 17→26 across two prior facts-only interventions) are not 103 independent misjudgments — they're one structural fact repeating: the receipt's room lane is bound to Originating-Terminal (which terminal the operator was physically in), while CLAUDE.md's own commit-trailer design already anticipated and separately records the correct multi-room mapping (Relevant-Rooms) that the lane computation never consults. Prior interventions treated each commit as its own drift event; none of them touched the resolver that keeps manufacturing this exact class.

   The missing guard (the deliverable): A guard (e.g. tests/pmu-simulator/room-lane-uses-relevant-rooms.test.mjs) that constructs a synthetic commit whose Originating-Terminal names a room owning none of the changed files but whose Relevant-Rooms trailer correctly names the owning room, and asserts the drift-receipt's lane/offPct computation credits the Relevant-Rooms room — SMOKE: fails today against the current Originating-Terminal-only resolver; NO-SMOKE: a commit whose Originating-Terminal and Relevant-Rooms agree still resolves cleanly.

   Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Saturated, not right-sized: litIntent/walksIntent = 130/130 and litReality/walksReality = 139/139 — every walked cell lit on both sides, which means the walk isn't discriminating signal from noise on a thin, two-file commit like this one. Concrete adjustment: for commits under a small diff-mass threshold (this one: 2 files, 80 lines), raise the lit-threshold or cap walk depth so a genuinely small commit can produce a partially-lit, discriminating result instead of full saturation on both lanes. Experiments: Re-run fb307c4af9's placement with the reef seed fed from the Relevant-Rooms trailer (navigator+architect+operator, weighted) instead of Originating-Terminal alone, and measure whether offPct drops below the 15 threshold. · Ablation: build a distinct chat-pipeline reef vocabulary bucket (harvest/scrape/beat/receipt/dedupe terms) separate from architect's strategy vocabulary, and measure whether future scripts/chat/ commits under any Originating-Terminal show lower cross-lane offPct once that bucket exists. · Test raising the lit-threshold specifically for small-mass commits (diff insertions below some N) and measure whether litIntent/walksIntent and litReality/walksReality stop saturating at 100% on this commit's historical bytes.

   Measured outcome (deterministic, not self-graded): room mean offPct 19→29 over 12 commits since · dictionary ablation saved 933B

   A correct, well-guarded pipeline fix charged against the wrong room's premium looks like architect's loss-ratio worsening for work architect never actually owns, and that's restored only once the lane computation prices commits against Relevant-Rooms glob-ownership instead of Originating-Terminal.

   Open the signed attestation →

2. 2026-08-29 · 📐 architect · offPct 27 · mistagged

   fix(elicit): the held lane is for anything the machine wrote to itself, not just canaries

   Why it left the lane: The diff's entire mass sits in two paths — scripts/elicit/elicit-run.mjs (MACHINE_MARKERS, isMachineAuthored, CANARY_RE, isCanary) and tests/elicit/machine-authored-turns-are-held.test.mjs — and zero bytes touch any of the room's owned globs (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, docs/01-business/strategy/, CLAUDE.md, per .workflow/rooms/vscode-architect.html's own 'Owned file/dir surface' list). This isn't a partial bleed at a boundary; it's a 100% miss between the commit's actual vocabulary (marker registry, ladder rung, refine ask) and the room's defined vocabulary (roadmap, cardinal goal, wedge). The commit's own Relevant-Rooms trailer already names the true owners — 🧭 navigator, 📐 architect, 🧪 laboratory, architect not even listed first — which is the commit disclosing the mismatch itself.

   The semantic pull: MACHINE_MARKERS / isMachineAuthored / CANARY_RE / 'elicit-run.mjs' / 'tests/elicit/' — pure Operations-side pipeline vocabulary with no lexical overlap against the Strategy-lane terms (roadmap, cardinal goal, positioning, wedge) that seed the architect room's reef. prewalkOverlap=1 shows the seed already agreed the content wasn't architect's before the walk even ran; the walk then spread across 123 lit reality cells purely inside the pipeline-code lane.

   Intent–reality gap: The commit message is entirely about an elicitation-pipeline correctness bug (a lane-detection generalisation in scripts/elicit); the receipt scored it against 📐architect's docs/strategy surface and found 27% of lit reality outside it. The gap isn't that architect did the wrong work — it's that the commit was classified by which terminal the operator was sitting in (Originating-Terminal), not by what was built, and the trailer protocol that permits this divergence is CLAUDE.md's own stated design.

   • Update .workflow/rooms/vscode-architect.html's Drift-detection hook section: implemented-but-unwired, not 'spec'd, not yet implemented'
   • Rewrite scripts/room-drift-architect-check.sh's trigger to parse the Originating-Terminal trailer instead of $TERM_PROGRAM/literal string match
   • Add a warn-only call to scripts/room-drift-architect-check.sh from hooks/pre-commit
   • grep the other 8 .workflow/rooms/*.html files for the same 'Drift-detection hook ... spec'd, not yet implemented' pattern — confirm which of the 8 have a shipped-but-unwired script like architect's, vs genuinely unbuilt
   • Add a wired-check guard (tests/ops/claimed-guards-exist.test.mjs's sibling: claimed-guards-are-called) that fails today because room-drift-architect-check.sh has zero references in hooks/
   • Check whether the PMU offPct computation itself (the pipeline that scored this commit) can be re-run scoring against each Relevant-Rooms entry instead of only Originating-Terminal, and whether that changes the room-mean trend

   What we learned: The prior lesson (ff09774e5) diagnosed 'three rooms independently spec'd the same fix and nobody shipped it' — this event proves that diagnosis was itself incomplete. Architect's fix WAS shipped, as a standalone file, five months before that lesson was even written (2026-05-12) — but it was never wired into hooks/pre-commit, and its own room's spec page still reads 'not yet implemented' today. A written-but-uncalled guard is functionally identical to no guard, except it also produces false confidence that the class is handled, because nobody who reads the spec page or greps for the filename learns it's dormant. The next escalation past 'ship the fix' is 'audit whether previously-shipped fixes actually got wired' — existence and connection are two independently-failable states, and this repo has now failed at the second one for over three months on this exact class.

   The missing guard (the deliverable): A wired-guard test — distinct from tests/ops/claimed-guards-exist.test.mjs, which only proves a file loads — that asserts hooks/pre-commit (or hooks/post-commit) actually invokes scripts/room-drift-architect-check.sh (and any sibling scripts/room-drift-<room>-check.sh). It fails today: grep across hooks/pre-commit and hooks/post-commit returns zero matches for 'room-drift'.

   Detector probe: signal attribution — gzip reef 15% · ballistic walk 85%. Right-sized but truncated asymmetrically: intent's walk (115/115, fully lit, no truncation loss) converged cleanly while reality's walk (123/123) hit the same nominal budget while covering genuinely broader ground (a registry generalisation plus its guard). Adjustment: when litReality>litIntent AND walkTruncated=true, re-run reality's walk at one additional --max-depth tier before scoring offPct, so a legitimately broader but still-converging lane isn't scored as if it had already saturated. Experiments: Recompute this commit's offPct three times, once per Relevant-Rooms entry (navigator, architect, laboratory) instead of Originating-Terminal, and record whether any scores under the 15 threshold — direct evidence for whether the receipt should read Relevant-Rooms. · Grep hooks/pre-commit and hooks/post-commit for every scripts/room-drift-*-check.sh reference (currently 1 file exists, 0 referenced); after wiring, replay the last 14 days of architect's 123 out-of-lane commits through the wired guard and count how many it would have caught before the commit landed. · Re-run --max-depth one tier higher specifically on walkTruncated=true commits where litReality>litIntent, and check whether offPct converges toward walksIntent/walksReality parity — measures how much of the 27% is truncation artifact versus real drift.

   Measured outcome (deterministic, not self-graded): room mean offPct 20→21 over 10 commits since · dictionary ablation saved 357B

   A guard that exists on disk but was never wired into the commit path reads as risk transferred when none was, which is worse for this lane's loss-ratio than having no guard at all — premium here should stay elevated against the 123-events/14-days frequency until hooks/pre-commit actually calls scripts/room-drift-architect-check.sh and the room's own spec page stops claiming the fix is unbuilt.

   Open the signed attestation →

3. 2026-08-29 · 📐 architect · offPct 24 · mistagged

   fix(elicit): merge two independent canary decompositions — I clobbered the first

   Why it left the lane: The entire diff (98 insertions/101 deletions, 1 file) is scripts/elicit/canary-legs.mjs — vocabulary like `nonce`, `sealedAt`, `tape`, `CANARY_RECEIPTS`, `mailedFor`, `sqlite3`, `Gemini ingest` — none of which touches docs/strategy/, docs/03-product/, .workflow/rooms/*.html or CLAUDE.md, the six globs architect actually owns. Confirmed on disk: .workflow/rooms/rio-navigator.html and .workflow/rooms/cursor-laboratory.html both list `scripts/elicit` under their own 2026-08-29 focus_areas, and neither string appears in vscode-architect.html's owned-surface list. So the reality mass (litReality=137, walksReality=137) landed exactly where it belongs — in navigator/laboratory territory — while the intent lane being scored (architect's owned surface) never had any of this vocabulary in it to begin with.

   The semantic pull: walksReality(137) > walksIntent(95) with prewalkOverlap=1 — the reef seed matched cleanly pre-walk, so the divergence isn't reef noise, it's the walk faithfully finding the real cluster of ops/instrumentation terms (`canary`, `nonce`, `ledger`, `tape`, `mail`, `guard`) that sit in C.Operations-flavored territory, not A2.Strategy.Goal (architect's own coordinate, per vscode-architect.html: "roadmap, cardinal goals, room architecture"). The commit message's dense incident-report prose ("WHAT HAPPENED", "NEITHER VERSION WAS THE RIGHT ONE") reads structurally like the strategy room's narrative register, which is plausibly what pulled the intent-side reef seed toward architect at all — the code itself never was.

   Intent–reality gap: The commit message claims nothing about strategy/roadmap work and is honest about what it did (merged two colliding canary-instrumentation implementations); the trailers even correctly name Relevant-Rooms as 🧭 navigator and 🧪 laboratory. The gap isn't in what the commit claims — it's that the receipt scored offPct against Originating-Terminal's (architect) owned surface instead of the rooms the commit's own trailer named as relevant, which are the rooms that actually own scripts/elicit/.

   • Add `scripts/elicit/*.mjs` to navigator's and laboratory's owned-surface arrays in `data/rooms.json` — the machine-readable file the receipt pipeline reads — since the HTML docs already list it but rooms.json (the actual source the offPct classifier consumes) has not been confirmed to match
   • Implement `scripts/room-drift-check.mjs` per code_rewrite above and retire the three duplicated per-room stub descriptions in favor of one shared script referenced by all three HTML docs
   • Close self-prompts sp-301 and sp-302 (both already queued, due 2026-08-29, both describing this exact fix) against the shipped script rather than leaving them open as separate asks
   • Add a guard test asserting the lane classifier reads Relevant-Rooms when present
   • Backfill: re-score the room's last 114 out-of-lane events under the corrected classifier to see how many were mistagged-not-drifted, since the pattern here (architect terminal, non-strategy diff, correct Relevant-Rooms trailer) may explain a large share of them

   What we learned: The room-lane drift receipt currently classifies a commit's intent lane by Originating-Terminal alone, but CLAUDE.md's own commit-trailer protocol explicitly sanctions Originating-Terminal/Relevant-Rooms divergence as normal and even requires both trailers on every commit — meaning the detector has been flagging behavior the house rules declare correct, and three separate rooms (architect, navigator, laboratory) each independently spec'd the identical missing fix (`room-drift-<room>-check.sh`) without anyone shipping or unifying it.

   The missing guard (the deliverable): `scripts/room-drift-check.mjs` (the shared implementation of the three currently-dormant `room-drift-<room>-check.sh` stubs already named in all three rooms' own HTML docs) — asserts that offPct is computed against the union of Relevant-Rooms' owned-surface globs when that trailer is present, and fails the build/test suite if a commit's staged files match zero rooms across BOTH Originating-Terminal and Relevant-Rooms (the only case that is a genuine boundary violation).

   Detector probe: signal attribution — gzip reef 15% · ballistic walk 85%. walkTruncated=true alongside a 137-cell reality spread suggests the walk hit its time/depth budget before fully resolving, and it's unclear whether continued walking would converge reality further from or back toward the intent lane. Concrete adjustment: re-run this sha's walk with `--max-depth` raised (or unbounded, offline) and diff the final lit-cell count against 137 to separate genuine cross-lane spread from an artifact of the truncation. Experiments: Re-run the lane classifier for ff09774e5 substituting Relevant-Rooms' (navigator+laboratory) owned-surface union as the intent seed instead of architect's, and confirm offPct drops toward the ≤15 threshold — validates the mistagged diagnosis quantitatively rather than by inspection · Sample the architect room's other 113 out-of-lane events from the last 14 days and compute what fraction carry a Relevant-Rooms trailer that differs from Originating-Terminal — if that fraction is large, it explains most of the 114-event count as classifier lag rather than genuine drift · Re-run this sha's walk untruncated (raise --max-depth) and compare litReality against the truncated 137 to quantify how much of the +42-cell spread is genuine vs a budget artifact

   Measured outcome (deterministic, not self-graded): room mean offPct 20→14 over 8 commits since · dictionary ablation saved 1030B

   A false-positive out-of-lane flag against a correctly-attributed, correctly-trailered commit inflates the architect lane's measured loss ratio for free — fixing the classifier to honor Relevant-Rooms (rather than adding process overhead) is what restores the premium without costing the room any actual discipline.

   Open the signed attestation →

4. 2026-08-28 · 📐 architect · offPct 31 · mistagged

   feat(elicit): the tape records arrival time — accretion's gauge becomes measurable

   Why it left the lane: Every changed identifier in the diff — seen_at, first_seen_at, linkHash, turn_hash, recordTurns, verifyChain, .thetacog/tape-sessions — is tape/elicitation-pipeline vocabulary. None of it shares a token family with architect's owned corpus (docs/strategy/, docs/03-product/, docs/01-business/strategy/, CLAUDE.md, .workflow/rooms/*.html). The reef seed (prewalkOverlap=3 against litIntent=125, ~2.4%) already found almost nothing in common between architect's declared lane and this commit's actual text, so the ballistic walk had a near-empty in-lane basin to anchor to and spread instead toward whatever coordinate the pipeline vocabulary actually clusters at — which is not architect's.

   The semantic pull: scripts/elicit/turn-identity.mjs, tests/elicit/two-times-are-not-one-claim.test.mjs, and the message's own vocabulary (gemSkill, accretion, inter-turn median, thread chain hash) pulled placement toward the elicit/tape-pipeline coordinates that this room's owned globs don't cover at all — a pure content/lane mismatch rather than a spread-vs-lane one.

   Intent–reality gap: The commit message and Relevant-Rooms trailer claim this is (also) architect's work, but neither changed file — scripts/elicit/turn-identity.mjs nor tests/elicit/two-times-are-not-one-claim.test.mjs — falls under any of architect's declared owned-surface globs. The receipt measured reality landing almost entirely off that lane because there was never any of that lane in this commit to begin with.

   • Confirm in data/rooms.json / .workflow/rooms/*.html which room's owned-surface glob should legitimately cover scripts/elicit/** and tests/elicit/** (laboratory vs navigator vs neither yet)
   • Add that glob to the correct room's .workflow/rooms/<room>.html 'Owned file/dir surface' section
   • Stop listing 📐 architect in Relevant-Rooms on future elicit/tape commits unless a docs/strategy or docs/03-product file is also touched
   • Add the glob-overlap guard described in missing_guard so a zero-overlap trailer warns at commit time
   • Audit the room's other 84 out-of-lane events from the last 14 days for the same architect-terminal-doing-elicit-work pattern, to size how much of that count is this same mistagging class
   • Re-render 0517312033's triptych scored against laboratory's (corrected) lane to confirm offPct drops under 15

   What we learned: Originating-Terminal records WHERE the operator was physically sitting, not WHAT surface the commit touches — and when the receipt's intent lane is drawn from Originating-Terminal rather than from actual owned-surface glob-overlap, any commit typed from a room's terminal while doing another room's content work is guaranteed to read as drift, independent of code quality. Architect's 84-event/14-day out-of-lane rate is plausibly dominated by this attribution axis rather than by architect's own work actually wandering.

   The missing guard (the deliverable): A pre-commit (warn-only, matching the existing 'Format violation does NOT block commit' convention) that computes glob-overlap between the changed files and Originating-Terminal's owned-surface list; if overlap is zero AND no room in Relevant-Rooms glob-matches either, print a warning naming which room's globs the files DO match, so the trailer can be corrected before it ships mistagged.

   Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized in relative spread (intent/reality walk counts and lit counts track each other closely) but walkTruncated=true on a 2-file, 105-line commit suggests the depth/time budget is tripping on small diffs before convergence rather than because the semantic content is actually large — worth raising --max-depth or the time budget specifically for small-diff commits so truncation itself doesn't inject noise into offPct. Experiments: Re-render sha 0517312033's triptych with the intent lane set to laboratory's (or navigator's) owned surface instead of architect's, and confirm offPct falls under the 15 threshold — directly tests the mistagging hypothesis · After adding scripts/elicit/** + tests/elicit/** to the correct room's owned-surface glob, re-run the last 14 days of architect out-of-lane events and measure what fraction were commits touching those same paths · Track prewalkOverlap against offPct across those 84 events; if low prewalkOverlap (<5) consistently predicts high offPct, use it as a cheap pre-walk flag for 'likely mistagged' before spending the full ballistic walk budget

   Measured outcome (deterministic, not self-graded): room mean offPct 16→32 over 3 commits since · dictionary ablation saved 1106B

   This event inflates architect's measured loss-ratio with an attribution defect rather than a genuine strategy-lane drift, so its premium correction is a re-scoring against the room that actually owns scripts/elicit — not a code fix — and architect's true loss-ratio can only be read accurately once that reassignment lands.

   Open the signed attestation →

5. 2026-08-28 · 📐 architect · offPct 26 · mistagged

   feat(elicit): the ingest mail is dated, carries per-intent progress, and buzzes Telegram

   Why it left the lane: The commit message is 100% about scripts/elicit/ingest-and-report.mjs (dated subject, per-intent progress, Telegram buzz), but 538 of 611 changed lines (88%) sit in books/tesseract/chapters-html/08-from-meat-to-metal.html, public/book/chapters/08-from-meat-to-metal.html, and public/book/chapters/06-the-sandbagging-trap.html — a book-glossary metavector insertion into 'The Compression Witness' section, plus a full chapter-6 line-anchor renumbering. Neither chapter, glossary term, nor 'metavector' is mentioned anywhere in the commit message, so the walk lights up two full domains (ops/mail pipeline + book-glossary cross-reference) while the intent lane only names one.

   The semantic pull: 'metavector:the-compression-witness', 'Compounding Verities', 'Symbol Grounding Failure', 'ShortRank Addressing', 'Trust Debt / The Scrim', 'k_E = 0.003', '/book/chapters/glossary#' links, and the wholesale re-anchoring of every 'line-N' id across ch.6 and ch.8 — dense book-craft/glossary vocabulary with zero lexical overlap against the diff's actual intent vocabulary ('Telegram', 'subject', 'ingest', 'queue.ndjson', 'TARGETS').

   Intent–reality gap: The message claims a single-domain ops feature (dated mail subject + per-intent progress + Telegram buzz); the receipt measured a commit that is mostly an unrelated book-chapter glossary edit across two chapters, none of which the message fences or even acknowledges.

   • Split this class of commit going forward: never let a `git add -A` / hook-staged commit merge scripts/** with books/tesseract/** or public/book/** unless the message names both.
   • Add a warn-only pre-commit check that flags when a commit touches both a non-book script domain and books/tesseract/**|public/book/** with no shared keyword between the diff paths and the commit message.
   • Verify chapter 6's 386-line renumbering was intentional (not a stray `./scripts/build-book-html.sh` re-run picking up a stale markdown edit) — confirm books/tesseract/chapters/06-the-sandbagging-trap.md was the source of that diff, not leftover build drift.
   • Confirm the metavector block insertion has its own generator/source (if `scripts/elicit`-adjacent tooling produced it) and document that script's path so future book-metavector commits are traceable to running code, per FIND THE RUNNING CODE FIRST.

   What we learned: A commit's owned-file surface can legitimately span zero domains the operator thinks they're claiming: this commit's Originating-Terminal (📐 architect) owns none of the files actually changed (scripts/elicit/**, books/tesseract/**, public/book/**) — the room-attribution and the file-domain-attribution are both silently detached from the one-sentence commit message, and nothing catches a book-content ride-along on an unrelated feature commit until the receipt measures it after the fact.

   The missing guard (the deliverable): A pre-commit warn (not block) that computes the set of top-level path-prefixes in the staged diff and flags when more than one prefix-domain is present (e.g. `scripts/elicit/` vs `books/tesseract/`+`public/book/`) with no shared token between the domains and the commit subject — surfacing the split-or-explain decision before the commit lands, mirroring the existing lexical/outreach precommit gates.

   Detector probe: signal attribution — gzip reef 25% · ballistic walk 75%. litReality=138/144 (96%) and litIntent=124/144 (86%) are both near-ceiling, and walkTruncated=true confirms the walk hit its time/depth budget before naturally converging — this is saturation, not under-use. Concrete adjustment: cap per-file contribution to the walk (a single 386-line HTML diff should not seed as many walks as a 73-line script diff) so one oversized file can't drag the whole reality lane toward saturation. Experiments: Re-run this same offline probe on a version of this commit with books/tesseract/**+public/book/** excluded from the reality corpus and confirm offPct drops below 15 — isolates how much of the 26% is purely the unmentioned book domain. · Add a per-file walk-budget cap proportional to LOC share and measure litReality/litIntent saturation across the last 20 architect-room commits — outcome: saturation rate should drop below 90% median. · Track prewalkOverlap vs offPct correlation across the room's 88 recent out-of-lane events — if low prewalkOverlap consistently predicts high offPct (as here), seed strength, not walk depth, is the lever to tune next.

   Measured outcome (deterministic, not self-graded): room mean offPct 15→15 over 6 commits since · dictionary ablation saved 840B

   This is a commit-hygiene defect, not a code defect — it inflates this lane's measured noise (loss-ratio) by conflating two legitimate but unrelated bodies of work under one under-specified intent claim; premium is restored by splitting such commits going forward so each one's offPct reflects a single true domain.

   Open the signed attestation →

6. 2026-08-28 · 📐 architect · offPct 22

   fix(pmu): the publication boundary hashes the span — found by planting one

   Why it left the lane: (sensemaking unavailable — facts-only intervention)

   Measured outcome (deterministic, not self-graded): room mean offPct 30→27 over 8 commits since

   Open the signed attestation →

7. 2026-08-28 · 📐 architect · offPct 31 · mistagged

   test(elicit): no NEW commit may carry a spec that does not describe it

   Why it left the lane: The diff's own prose is written in CLAUDE.md's HARD-RULE dialect, not test-file dialect: 'NON-RETROACTIVE BY CONSTRUCTION', 'the cutoff must not become a way to stop counting', 'this test permanently red, which is the same as deleting it', and the closing line '"I could not look" and "I looked and saw nothing" are different claims and this repo has paid for collapsing them all day' is a near-verbatim echo of the GUARD-SEEN-TO-FAIL and META-BULK sections already living in CLAUDE.md (architect's own owned file). gzip-NCD register-similarity locks onto that dialect match and pulls the commit's reality mass toward architect's coordinate even though every changed byte (specs/s-0155.../*, tests/elicit/*.test.mjs) sits outside architect's declared surface.

   The semantic pull: 'NON-RETROACTIVE BY CONSTRUCTION', 'permanently red ... same as deleting it', and the two-different-claims closing sentence are the pull terms — they are CLAUDE.md's own rhetorical fingerprints, reused here to justify a test guard rather than a CLAUDE.md rule, so the reef seed reads this commit as 'CLAUDE.md-adjacent' before the walk ever looks at file paths.

   Intent–reality gap: The Relevant-Rooms trailer names 📐 architect first, but 100% of the diff (specs/s-0155.../acceptance.md, snapshot.hash, spans.ndjson, spec.md, tests/elicit/spec-trailer-names-this-commit.test.mjs) falls outside architect's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md) — the receipt measured a testing/spec-infrastructure landing while the trailer claimed a strategy-doc lane.

   • Audit `.workflow/rooms/architect.html`'s owned-surface globs to confirm they do NOT accidentally match specs/** or tests/elicit/** (would explain a scorer-level, not just a manual, mismatch)
   • Add a warn-only post-commit check comparing the Relevant-Rooms trailer against glob-matched rooms from git diff --name-only, printing (not blocking) on mismatch, per the existing 'format violation does NOT block commit' convention
   • Update CLAUDE.md COMMIT MESSAGE STRUCTURE with the files-not-prose disambiguation line for Relevant-Rooms
   • Seed the reef's intent placement for specs/*/spec.md commits from the embedded `**Coordinate**` line (here 'C1 Operations.Grid') instead of relying solely on commit-message prose similarity
   • Note commit 119349dff5 as a known trailer-vs-surface mismatch in whatever downstream dashboard consumes trailers, since the commit itself can never be amended on shared main
   • Extend tests/elicit/spec-trailer-names-this-commit.test.mjs's own class (spec-vs-commit description accuracy) with a sibling check for room-trailer-vs-files accuracy — same mechanism, same file, same author already thinking about it

   What we learned: The Relevant-Rooms trailer convention (CLAUDE.md: 'the 3 highest-scoring matches against .workflow/rooms/*.html owned-surface globs') is unenforced and in practice defaults to reusing Originating-Terminal's room rather than being independently glob-computed — so any commit written in a room's own rhetorical register (here, architect's HARD-RULE dialect, reused to justify a test guard) will self-tag into that room's lane regardless of which files actually moved.

   The missing guard (the deliverable): A warn-only commit-trailer/room check: compute the true top-3 glob-matched rooms from the commit's changed files against `.workflow/rooms/*.html` owned-surface globs, and print a mismatch when the Relevant-Rooms trailer doesn't overlap — the same convention CLAUDE.md already declares but never mechanized.

   Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litReality (140) modestly exceeds litIntent (127) and the walk was truncated before natural termination on a 5-file, dual-directory (specs/ + tests/) commit — under-provisioned for a commit spanning two structurally distinct corpora. Adjustment: raise --max-depth (or the time budget) specifically for commits touching both specs/**and tests/** in one diff, so the walk resolves fully before truncation forces a partial, noisier placement. Experiments: Rerun this sha's placement with the commit message stripped from the reef seed (file-content-only) and compare offPct — a large drop confirms prose-register, not files, drove the misplacement · Rerun with the spec.md '**Coordinate**' line (C1 Operations.Grid) injected as an explicit intent anchor for specs/*/spec.md commits and measure whether prewalkOverlap rises above 13 · Rerun with 2x --max-depth on this sha and check whether offPct converges toward the 15 threshold or stays elevated, isolating truncation as a cause versus genuine drift

   Measured outcome (deterministic, not self-graded): room mean offPct 17→31 over 6 commits since · dictionary ablation saved 1383B

   A room mistagging its own trailer corrupts the per-room offPct calibration this lane's premium is priced from, so the fix that restores premium is the trailer-vs-glob guard landing before the next architect-tagged commit, not a retroactive edit to this one.

   Open the signed attestation →

8. 2026-08-28 · 🧪 laboratory · offPct 38

   fix(elicit): the conformance detector was grading a construct it could no longer see — false GAP re-aimed to computed

   Why it left the lane: (sensemaking unavailable — facts-only intervention)

   Measured outcome (deterministic, not self-graded): room mean offPct 20→10 over 3 commits since

   Open the signed attestation →

9. 2026-08-28 · 📐 architect · offPct 22

   feat(elicit): outward work runs as a DRAFT — composing is not sending

   Why it left the lane: (sensemaking unavailable — facts-only intervention)

   Measured outcome (deterministic, not self-graded): room mean offPct 16→15 over 8 commits since

   Open the signed attestation →

10. 2026-08-28 · 🎭 performer · offPct 29 · structural

    comms(external): Jason Bishara direct ask sent — attempt 3, peer channel repaired

    

    Why it left the lane: The diff's live vocabulary is comms-pipeline operational language, not git-discipline signal language: 'Three slips (11 Aug held / 19 Aug moved / 25 Aug declined)' and 'issued Tue 8 Sep 3:30pm ET' pull toward B1.Tactics.Speed (scheduling/velocity); 'pre-negotiate the trigger as an endorsement' and 'the one closed question' pull toward B2.Tactics.Deal (exchange); 'Graded via 4 independent Sonnet subagents', the grade.json's 'graded_by'/'needs' array, and the sent/.md's 'Fired by: GCal C2 event' pull toward C1/C2/C3 Operations.Grid/Loop/Flow (the send-and-grade pipeline's own topology, cadence, throughput). None of that content is a git-commit signal at all — B3.Tactics.Signal only got weight because the commit was physically typed in the 🎭 performer terminal, and the lens has nowhere else to put docs/comms/external/ content.

    Intent–reality gap: The commit message claims a performer-room git-discipline act ('comms(external): ... sent') but its own Relevant-Rooms trailer already names ☕ network, 🎩 operator, 🔒 vault — not 🎭 performer — as the rooms this touches; the receipt then measured the work landing across six Speed/Deal × Grid/Loop/Flow coordinates because docs/comms/external/ has no owning glob for the lens to place it against, so it fell back to the terminal of origin.

    • Add `docs/comms/**` to ☕ network's owned-surface glob in .workflow/rooms/network.html
    • Sweep other CLAUDE.md-named-but-unglob'd trees the same way: docs/01-business/deals/ prep briefs, data/self-prompts.json, data/room-bifurcations.json (per prior lesson f08ec5875)
    • Add a room-ownership coverage test (see missing_guard) covering both the JSON-state class and the docs/comms/ class in one assertion
    • Re-run this sha's placement after the glob lands and record the before/after offPct in the intervention ledger
    • Document in CLAUDE.md's sovereign-outreach section that docs/comms/** is network-owned regardless of originating terminal

    What we learned: The ownerless-surface defect from f08ec5875 is not confined to bookkeeping JSON — it recurs for any directory CLAUDE.md describes in prose (the whole 'Channels = folders' comms architecture) but never registers as an .workflow/rooms/*.html glob; a prose-only ownership declaration is invisible to the lens the same way a prose-only rule is invisible to reseed, and the fix is the same shape: give it a glob, not a paragraph.

    The missing guard (the deliverable): tests/rooms/comms-surface-has-owner.test.js — asserts every path CLAUDE.md's Sovereign Outreach / Canonical Communication Protocol sections name as a channel folder (docs/comms/**, docs/outreach/**) resolves to exactly one .workflow/rooms/*.html owned-surface glob; extend it to also cover the data/room-bifurcations.json class from f08ec5875 so both known instances of this defect are guarded in one place.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Under-utilized for content this heterogeneous: a comms send legitimately spans scheduling+deal+pipeline-status in one document, but a 2-ply walk can't distinguish 'this is one coherent multi-topic artifact' from 'this drifted.' Concrete adjustment: re-run this sha with --max-depth 3 and compare whether the six out-of-lane coordinates collapse toward fewer, more confident cells or stay equally flat — that tells you whether shallow-ply truncation, not real drift, produced the spread. Experiments: Add the docs/comms/** → network glob, then re-run this exact sha's placement and diff offPct before/after (single-variable test of the ownership fix) · Re-walk this sha at --max-depth 3 and compare encircled-region count/spread against the current 2-ply result to test the walk_utilization hypothesis · Seed the gzip baseline dictionary with the compression_dictionary_adds strings below and ablation-test prewalkOverlap on the last 20 docs/comms/ commits with/without them

    Measured outcome (deterministic, not self-graded): room mean offPct —→30 over 6 commits since · dictionary ablation saved 638B

    This event is a false-positive loss signal, not a real one — the send itself was correctly graded and rule-compliant, so the fix that restores this lane's loss ratio is removing the ownership gap (the glob addition), not disciplining the content.

    Open the signed attestation →

11. 2026-08-28 · 📐 architect · offPct 45 · structural

    feat(elicit): the committed spec record, sealed at graduation

    Why it left the lane: No encircled-region narrative was cached for this sha, so I'm reasoning from file paths + diff vocabulary directly. The entire diff is two new files — scripts/elicit/spec-record.mjs and tests/elicit/spec-record.test.mjs — both under scripts/elicit/ and tests/elicit/. Architect's declared lane is docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md — a docs-only surface with zero glob coverage for any scripts/ or tests/ path. Every identifier the diff actually contains — writeSpecRecord, verifySpec, snapshotOf, trailerOf, the four filenames spec.md/spans.ndjson/acceptance.md/snapshot.hash — is executable JS + its test harness, which the walk correctly places in code-territory cells outside architect's docs lane. The commit reads like an architecture decision (it literally is one: sealing a promise at graduation) but every byte of it landed as code.

    The semantic pull: 'spec.md','spans.ndjson','acceptance.md','snapshot.hash' (the four sealed files), 'writeSpecRecord'/'verifySpec'/'snapshotOf'/'trailerOf' (the exported functions), and the path prefixes scripts/elicit/ + tests/elicit/ themselves — none of these strings intersect architect's owned glob set at all, so the reality walk has nowhere in-lane to land once it reaches this subsystem.

    Intent–reality gap: The commit message frames this as an architecture act — 'the committed spec record, sealed at graduation,' a promise/contract concept squarely in architect's conceptual territory — while the receipt measured 100% of the changed bytes landing in scripts/elicit/ and tests/elicit/, paths with no entry anywhere in architect's owned-surface declaration.

    • Add scripts/elicit/**, tests/elicit/**, .thetacog/elicit/** to architect's owned-surface glob in .workflow/rooms/architect.html
    • Add a `// Lane: 📐 architect` header comment to scripts/elicit/spec-record.mjs
    • Audit .thetacog/elicit/ (probe-state.json, pulse-receipts.ndjson, tcc-backoff — all already modified per git status) for the same missing-lane pattern before more commits land there
    • Write a guard test asserting every scripts/**/ and tests/**/ directory prefix appears in at least one room's owned-surface glob list
    • Given 76 off-lane events in 14 days for architect, evaluate whether 'elicit' should graduate into its own tracked coordinate rather than being folded into architect's surface
    • Mirror the glob addition into any CLAUDE.md room-ownership table entry for architect if one lists path globs directly

    What we learned: The 'elicit' subsystem (scripts/elicit/, tests/elicit/, .thetacog/elicit/) has been accumulating files for multiple commits with NO owned-surface glob in any room's declaration anywhere in the reef — a genuine new coordinate has been growing unclaimed, so every commit into it is structurally guaranteed to read as off-lane no matter how careful the work is. Prior lessons for this room never named a specific missing-surface subsystem; this is the first event that traces the recurring 76-in-14-days pattern to one identifiable orphan directory rather than to diffuse drift.

    The missing guard (the deliverable): A test that walks every top-level scripts/**/ and tests/**/ directory and asserts each prefix appears in at least one room's owned-surface glob declaration (`.workflow/rooms/*.html`) — it would have caught scripts/elicit/ the moment the first file landed there, instead of after dozens of off-lane commits accumulated against architect.

    Detector probe: signal attribution — gzip reef 25% · ballistic walk 75%. Saturated, not right-sized: lit/walked ≈ 1.0 on BOTH intent (122/122) and reality (119/119) means the walk is lighting essentially every cell it visits on both sides — it isn't discriminating in-lane from out-of-lane, it's just filling whatever ply-radius it's given. Concrete adjustment: cap --max-depth at 2 for this commit class (spec/contract-shaped commits) and check whether litReality drops below walksReality — if it does, the walk was over-extending into code-territory plies that a shallower budget would have pruned. Experiments: Re-run this commit through pmu-onchip --ballistic --start <architect-anchor> --max-depth 2 (vs default) and compare offPct — tests whether shallower plies reduce saturation-driven leakage into scripts/elicit/ · Seed the architect reef corpus explicitly with scripts/elicit/**, tests/elicit/** content and re-measure prewalkOverlap on the next 5 elicit-subsystem commits — tests whether reef overlap and offPct move together once the glob exists · Ablation-recompress the last 20 architect commits' reality corpus with and without 'spec.md'/'spans.ndjson'/'acceptance.md'/'snapshot.hash' in the reef dictionary, and measure whether offPct on elicit-subsystem commits specifically drops

    Measured outcome (deterministic, not self-graded): room mean offPct 34→17 over 3 commits since · dictionary ablation saved 760B

    76 off-lane events in 14 days reads as a systemic misclassification risk in architect's lane definition rather than operator negligence, which inflates the lane's apparent loss-ratio on every elicit-subsystem commit; premium is restored the moment scripts/elicit/**, tests/elicit/**, and .thetacog/elicit/** are registered against an owning room so the receipt prices these commits against a lane that actually claims them.

    Open the signed attestation →

12. 2026-08-28 · 📐 architect · offPct 26

    fix(elicit): span leakage — a name in a spec reached the published commit page, measured with a canary

    

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 28→30 over 9 commits since

    Open the signed attestation →

13. 2026-08-28 · 📐 architect · offPct 19

    docs(elicit): s-0153 sealed, then superseded by measurement — not executed

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 18→21 over 5 commits since

    Open the signed attestation →

14. 2026-08-28 · 🎩 operator · offPct 16

    feat(club): the free half names its cargo — the full book rides in the repo, ask your own AI

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 15→30 over 13 commits since

    Open the signed attestation →

15. 2026-08-28 · 🎩 Kitty Operator · offPct 34 · mistagged

    fix(voice): three mangled nouns that each cost a round-trip today

    Why it left the lane: The diff is 31 lines of pure additions to data/voice/glossary.json — three new alias entries whose VALUES are themselves other domains' vocabulary: 'encircled panel' (an Operations.Grid/receipt-pipeline term), 'double-entry' (an accounting/ledger term, i.e. Operations.Loop/Flow territory), and 'sticker print job' (production/physical-goods, Tactics.Deal-adjacent). The reef's gzip-NCD seed grips on the STRINGS being added, not the mechanism doing the adding, so the placement mass followed the referenced domains (B1-B3 Tactics × C1-C3 Operations, 31 cells, off-lane 100%) instead of staying inside the room's actual owned surface, which is dashboard/pipeline/accounting infra files data/trust-debt/, docs/01-business/pipeline/, docs/01-business/accounting/ — none of which data/voice/glossary.json is a member of.

    The semantic pull: 'encircled panel' and 'double-entry' are the two strings doing the dragging — they are the load-bearing nouns of this whole deck's OTHER domains (receipt pipeline, accounting), so a term-level, domain-blind reef seed reads their presence as 'this commit is about Operations.Grid and the ledger' rather than 'this commit is about the dictionary that maps them.'

    Intent–reality gap: The commit message claims a narrow, bounded voice-glossary fix ('three mangled nouns... add the entry the moment it costs a round-trip') scoped entirely to the dictation pipeline — but the file it touches, data/voice/glossary.json, is not on 🎩 Kitty Operator's owned surface at all (no data/voice/ glob anywhere in the room's file list), and the commit's own Relevant-Rooms trailer already names 🎤 voice as more relevant than 🎩 operator. The receipt is correctly reporting that this work does not belong to the room it was filed under.

    • Add data/voice/*.json to 🎤 voice room's owned-surface glob in the rooms config so glossary commits stop scoring as off-lane for whichever terminal happened to type them
    • CLAUDE.md → VOICE DICTATION AUTO-CORRECT: state that glossary.json commits are voice-room surface by content, not by originating terminal
    • Add a lightweight domain tag to new glossary entries whose canonical form names another lane's vocabulary (encircled panel, double-entry) so ingest can down-weight the alias value string
    • Cross-check Relevant-Rooms trailer against the room the receipt actually assigns lane-ownership to at commit time, and warn (not block) when they disagree
    • File a one-line note in the pipeline room's docs that 'double-entry' and 'encircled panel' are voice-glossary aliases, not new pipeline vocabulary, to pre-empt a future reef reseed re-learning the same pull

    What we learned: A glossary of OTHER domains' terms is structurally guaranteed to drift out of whatever lane it's filed under, because the reef grips on string content and a dictionary's entire content IS other domains' strings — this is a distinct failure shape from code or docs drift (which at least describe their own lane) and needs its own placement rule, not just 'add the guard that caught the last thing.'

    The missing guard (the deliverable): tests/rooms/voice-glossary-owned-surface.test.mjs — asserts every commit touching data/voice/glossary.json carries either Originating-Terminal: 🎤 Voice or a Relevant-Rooms entry naming 🎤 voice first, so a glossary edit filed from a different terminal is flagged before the receipt fires, not after.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Saturated-but-truncated: walkTruncated=true with lit≈walked on both sides suggests the walk filled its budget window completely rather than terminating early for lack of connectivity. Concrete adjustment: rerun this sha through pmu-onchip --ballistic --max-depth 5 (current run implies depth 2) to see whether a deeper budget resolves more of the 31 off-lane cells back toward B3.Tactics.Signal, or whether they're genuinely reachable-and-lit regardless of depth (which would confirm the seed, not the walk, is the fixable half). Experiments: Re-run this sha's reef seed with the two alias VALUES (encircled panel, double-entry) masked out of the gzip-NCD corpus and diff prewalkOverlap before/after — if overlap rises materially above 4, the domain-tag code_rewrite is validated as the right lever · Compare offPct for the last 5 glossary-only commits (data/voice/glossary.json, no other file changed) against the room they were filed under vs 🎤 voice — if all 5 show the same B1-B3×C1-C3 spray, this is a repeatable class worth a standing exemption rule, not a one-off patch · Measure whether prewalkOverlap<5 is itself a leading indicator of eventual offPct>15 across the last n120 series — if it correlates, surface prewalkOverlap in the commit-time warning before the receipt ships, not after

    Measured outcome (deterministic, not self-graded): room mean offPct 16→28 over 9 commits since

    A mistagged-lane commit costs this room's loss-ratio nothing in code quality but inflates its offPct trend with false-positive noise (3 events in 14 days, none of them real cross-lane work), which erodes the signal the premium is priced on — restored by moving data/voice/*.json onto 🎤 voice's owned surface so future glossary commits stop being scored against 🎩 operator's lane at all.

    Open the signed attestation →

16. 2026-08-28 · 🎤 voice · offPct 26

    feat(deck): dentry v1.2-d3 — the patent line on the tollbooth (US 19/637,714, Track One)

    

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 10→25 over 4 commits since

    Open the signed attestation →

17. 2026-08-28 · 📐 architect · offPct 56 · mistagged

    feat(elicit): turn identity survives a re-render — the tape becomes a record

    Why it left the lane: The diff is dense with cryptographic/ledger vocabulary — `createHash('sha256')`, `turn_hash`, `linkHash`, `appendFileSync`, `readFileSync`, `mkdirSync`, `verifyChain`, `cursorFor` — none of which shares vocabulary with 📐 architect's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, CLAUDE.md). The reef seeded from those docs-lane anchors and the recursive walk (walksReality=139, litReality=139 — fully saturated, no unlit cells) filled almost entirely on the implementation/operations side of the lattice rather than the strategy-prose side the architect lane anchors to. prewalkOverlap=3 confirms this split existed before the walk even started spreading: only 3 cells overlapped between the two corpora at the gzip-NCD seed stage.

    The semantic pull: Terms like `sha256`, `linkHash`, `appendFileSync`, `mkdirSync`, `process.argv`, `fileURLToPath` are mechanism/infrastructure vocabulary — they pull toward the Operations axis (grid/loop/flow — deterministic pipeline machinery), not the Strategy axis architect's docs corpus anchors. No region narrative is cached for this sha, so the exact [row,col] anchors aren't nameable here, but the direction of pull is unambiguous from the diff's own vocabulary versus the architect owned-surface glob list.

    Intent–reality gap: The commit message and trailers claim this is architect-room work (Originating-Terminal: 📐 architect) but the commit's own Story line and Relevant-Rooms trailer already name 🧪 laboratory and 🔨 builder as the rooms that actually recognize this content — the receipt's intent lane is computed from Originating-Terminal's owned surface (docs-only) while the diff is 305 lines of new code + tests under scripts/elicit/ and tests/elicit/, which architect's glob list does not cover at all.

    • Add a `scripts/elicit/` + `tests/elicit/` owned-surface glob to 🔨 builder's and/or 🧪 laboratory's room HTML (whichever currently owns scripts/*.mjs implementation work), so future elicit commits land in-lane by construction.
    • Locate and patch the intent-lane resolver (likely in scripts/pmu/*.mjs where Originating-Terminal/Relevant-Rooms trailers are parsed) to union Relevant-Rooms globs as a fallback.
    • Add a pre-commit or post-commit check: if a commit's diff paths match zero globs from Originating-Terminal's owned surface but match a glob from a Relevant-Rooms entry, tag the receipt row as `terminal-divergence` rather than raw drift.
    • Re-run this sha's receipt against the widened lane (Relevant-Rooms union) as a one-off experiment to confirm offPct drops toward the 15% threshold — this validates the mistagged diagnosis before trusting the fix broadly.
    • Audit the room's other 73 out-of-lane events in the last 14 days for the same terminal-divergence signature (code diff, docs-only originating room) to see what fraction of the chronic 74-event count this single fix would reclassify.

    What we learned: Architect is being used as the terminal an operator happens to be sitting in while doing cross-room implementation work, and CLAUDE.md already blesses this divergence as normal — but the receipt's lane computation only reads Originating-Terminal's owned surface, so every code-heavy commit typed from architect will register as high offPct 'drift' even when the commit's own trailers correctly name the real owning rooms. The 74-events-in-14-days figure for this room is plausibly inflated by this exact structural gap, not by 74 genuine cross-lane mistakes.

    The missing guard (the deliverable): A guard asserting: for any commit whose Originating-Terminal room's owned-surface globs match zero changed files, the commit's Relevant-Rooms trailer must list at least one room whose owned-surface globs DO match — and the receipt pipeline must be provably reading that trailer when computing offPct (not silently ignoring it).

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Saturated, not right-sized: 100% lit-to-walked ratio on BOTH sides while walkTruncated=true suggests the walk hit its time/depth budget while still finding every visited cell lit — a sign it's flooding a dense region rather than converging on discriminative structure. Concrete adjustment: lower --max-depth for this corpus pairing (or add an early-stop when the lit-ratio holds at 100% for 2 consecutive plies) so the walk reports a spread number that means something rather than a uniform full-fill. Experiments: Recompute this commit's receipt with the intent lane widened to the union of Relevant-Rooms (architect+laboratory+builder) owned-surface globs and record whether offPct falls under 15 — a direct test of the mistagged diagnosis. · Re-run prewalkOverlap for the last 5 scripts/elicit/* commits against docs/strategy/ vs scripts+tests owned-surface seeds; if overlap stays near 0 across all of them, that confirms this is a structural corpus-disjunction (docs prose vs. crypto/hash code) rather than a one-off. · Seed the gzip baseline dictionary with the boilerplate strings below and re-measure prewalkOverlap on this sha; if overlap rises meaningfully that means the low overlap was partly dictionary-starvation on common Node.js idioms rather than genuine topical distance.

    Measured outcome (deterministic, not self-graded): room mean offPct 33→36 over 22 commits since · dictionary ablation saved 1229B

    A 56%-off-lane receipt on what is actually correctly-scoped, well-tested cross-room engineering work looks like a loss event but is a labeling artifact — leaving the lane computation as-is means this room's premium keeps pricing in false drift; fixing intent-lane resolution to read Relevant-Rooms restores the loss ratio to what the room's actual (non-terminal-divergent) drift rate is.

    Open the signed attestation →

18. 2026-08-28 · 🎩 operator · offPct 46

    feat(club): the email says what shipped — and the Meta vector follows the prose

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct —→11 over 4 commits since

    Open the signed attestation →

19. 2026-08-28 · 🎩 Kitty Operator · offPct 34

    spec(deck): the sticker goes on frame 10, the QR does not — and both specs get their envelope

    

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 21→9 over 3 commits since

    Open the signed attestation →

20. 2026-08-28 · 📐 architect · offPct 28 · mistagged

    feat(elicit): the index rebuilds from the tape — and the coverage ships with the verdict

    Why it left the lane: The commit's every touched file (scripts/elicit/rebuild-index.mjs, scripts/elicit/turn-identity.mjs, tests/elicit/rebuild-reports-its-coverage.test.mjs, .thetacog/tape-sessions/gemini-web/turns.ndjson) sits entirely outside 📐 architect's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, docs/01-business/strategy/, CLAUDE.md). The room label came from Originating-Terminal (the physical terminal the operator happened to be typing in), not from where the diff's mass actually lands, so litReality (141/144) necessarily scattered outside architect's declared lane.

    The semantic pull: Vocabulary like 'row-content identity', 'AXIOM 1', 'projection was not SUFFICIENT', 'sqlite', 'ingestScrapedFiles', 'contentHash', 'verifyRebuild' is data-pipeline/sufficiency-proof language — it pulled the walk toward Operations/Tactics coordinates (index rebuild, tape ingestion, measurement-of-coverage) with no term anywhere touching 'strategy', 'law', 'goal', or 'fund', which is what architect's owned docs/strategy/ and docs/01-business/strategy/ corpora are built from. litIntent saturating to 144/144 (walkTruncated=true) shows the abstract cross-cutting AXIOM-1 language broadened the walk past any single lane rather than anchoring it in architect's.

    Intent–reality gap: The message claims (and the trailers assert) this is 📐 architect work; the receipt shows 28% of the lit reality mass landed off that lane because the actual work — a tape→sqlite rebuild verifier and a dropped-column fix in the elicit pipeline — has no legitimate home inside architect's owned-surface globs at all.

    • Add scripts/elicit/**, tests/elicit/**, .thetacog/tape-sessions/** as an owned-surface glob to whichever room's .workflow/rooms/<room>.html should hold the elicit pipeline (🧪 laboratory is the strongest Relevant-Rooms match already on this commit)
    • Update the drift-receipt room-attribution script to prefer the highest-glob-match Relevant-Room over Originating-Terminal when they diverge, per the spec_rewrite text
    • Re-render this commit's receipt after the glob update and confirm offPct drops for the newly-owning room
    • Audit the other 85 out-of-lane events for 📐 architect in the last 14 days for the same zero-glob-match signature before assuming they're all genuine drift

    What we learned: No room's owned-surface glob currently claims scripts/elicit/**, tests/elicit/**, or .thetacog/tape-sessions/** — so any commit touching the elicit pipeline registers as 'out of lane' against whichever room's terminal happened to be open, independent of code quality, because there is structurally no lane it could land inside.

    The missing guard (the deliverable): A pre-commit or post-commit check that compares the changed-file set against the Originating-Terminal room's owned-surface glob; if zero files match while a different Relevant-Room's glob matches the majority, print a warning that the drift receipt will score against a lane none of the diff touches.

    Detector probe: signal attribution — gzip reef 30% · ballistic walk 70%. Saturated and truncated: litIntent=144/144 (total lattice) and litReality=141/144 with walkTruncated=true is the AR 'unguided flood' failure mode — the walk isn't discriminating a lane, it's covering nearly everything. Adjustment: lower --max-depth for commits whose message corpus is dominated by cross-cutting axiom/rule language, or add an early-stop once lit fraction exceeds ~60% and flag the result as 'diffuse' instead of continuing to truncation. Experiments: Re-run this commit's walk with --max-depth capped one ply lower and measure whether litIntent/litReality drop below saturation while offPct stays non-zero (i.e. still informative) · After adding the scripts/elicit/** glob to a room, re-render this commit's receipt and measure whether offPct drops now that the reality mass has a legitimate lane to land in · Pull prewalkOverlap across the other 85 architect out-of-lane events from the last 14 days; if most sit in single digits like this one (10), the architect reef corpus itself is too thin to hold discriminating signal and needs enrichment, not walk tuning

    Measured outcome (deterministic, not self-graded): room mean offPct 15→10 over 4 commits since · dictionary ablation saved 472B

    An elicit-pipeline commit correctly filed under architect's terminal but scored against a lane it structurally cannot occupy inflates this room's loss-ratio on paper; premium is restored by giving the elicit surface a real owning room so the same code stops registering as drift.

    Open the signed attestation →

21. 2026-08-28 · 📐 architect · offPct 34 · structural

    docs(elicit): s-0157 — the spec for the note-correction commit, sealed after the fact

    Why it left the lane: The diff is a spec-authoring artifact (spec.md, acceptance.md, spans.ndjson, snapshot.hash) about the correction/trailer detector itself — its Targets list names 'scripts/elicit/spec-trailer-matches.mjs' and 'tests/elicit/spec-trailer-names-this-commit.test.mjs', and its body is entirely about 'the Spec trailer', 'the detector', 'git note', 'trailer'. Those terms live in B3.Tactics.Signal (signal/detector/trailer semantics) and C1/C2/C3 Operations (grid/loop/flow — the elicit pipeline's own machinery), not in a single fenced coordinate — so the walk lights up broadly across B1/B2/B3 × C1/C2/C3 because 'spec', 'commit', 'detector', 'trailer' are exactly the vocabulary those cells are keyed on. There's no code change at all (docsOnly=true), so the reality walk has nothing narrow to grip and spreads across every cell those words touch.

    The semantic pull: Strings 'Spec trailer', 'detector', 'git note', 'correction channel', and the two Targets paths ('scripts/elicit/spec-trailer-matches.mjs', 'tests/elicit/spec-trailer-names-this-commit.test.mjs') are generic pipeline/signal nouns that the reef recognizes across the whole Operations×Tactics.Signal block rather than pinning to one cell — hence litReality=131 against litIntent=114 and 31 out-of-role coords.

    Intent–reality gap: The commit message claims a narrow, disciplined act — sealing one spec for one prior correction, explicitly declining a trailer because 'it was not spec-driven' — but the receipt measured the work landing broadly across 31 out-of-role Tactics×Operations cells, because the spec's own prose (about detectors, trailers, signal channels) is intrinsically about the machinery that spans those cells.

    • Reword spec.md's Intent line to front-load 'seals the record of a completed correction' before any mention of 'detector' or 'channel'
    • Split the Targets list: keep 'tests/elicit/spec-trailer-names-this-commit.test.mjs' (record/acceptance-scoped) but drop 'scripts/elicit/spec-trailer-matches.mjs' from Targets unless this spec actually changes that script's behavior — naming a live detector script as a Target for a docs-only sealing commit is itself part of the pull
    • Add a one-line 'Non-goals' section to spec.md: 'Does not change spec-trailer-matches.mjs behavior' — non-goals fence the walk the same way targets expand it
    • Add scripts/elicit/spec-trailer-matches.mjs (or a sibling doc) a header comment distinguishing 'detector logic' (B3.Signal) from 'sealed-after-the-fact record specs' (C1.Grid) so future docsOnly commits about corrections don't reuse detector vocabulary

    What we learned: A docs-only spec commit that documents a decision ABOUT the detector (using the detector's own vocabulary — 'trailer', 'channel', 'detector') gets placed AS IF it were detector code, even with zero code changes — the walk grips prose semantics, not diff type. docsOnly=true does not shrink the lane; it can widen it, because prose is freer to use broad nouns than a code diff naming one file.

    The missing guard (the deliverable): A pre-commit check for specs/*/spec.md: if the spec's Targets list names a file already owned by an existing B3.Tactics.Signal-tagged script (e.g. anything under scripts/elicit/*-matches.mjs), require either (a) a code change to that file in the same commit, or (b) an explicit 'Non-goals' section stating the spec does not alter that script's behavior — this forces the mistagged-vs-intentional distinction to be declared at write time instead of discovered at receipt time.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. litReality(131) > litIntent(114) with walkTruncated=true suggests the walk is running too shallow relative to this diff's vocabulary breadth — it's spreading into many cells but getting cut before any of them dominate/converge. Concrete adjustment: raise --max-depth by one ply specifically for docsOnly commits under specs/, since prose has no file-path anchor to narrow it the way code diffs do. Experiments: Re-run this exact commit with --max-depth+1 and compare offPct — if it drops, truncation (not genuine breadth) explains most of the 31 out-of-role cells · A/B the walk seeded from Targets-file-paths only vs seeded from full spec.md prose — if the Targets-only seed produces a materially tighter offPct, that's evidence the prose vocabulary (not the actual named files) is driving the spread, confirming semantic_pull · Compute prewalkOverlap for the last 10 specs/*/spec.md commits and correlate with offPct — if low prewalkOverlap consistently precedes high offPct for docsOnly specs, gate docsOnly commits on a minimum seed overlap before accepting the walk's placement as final

    Measured outcome (deterministic, not self-graded): room mean offPct 19→21 over 14 commits since · dictionary ablation saved 839B

    A docsOnly commit landing at offPct=34 in a room already carrying 95 out-of-lane events in 14 days erodes 📐 architect's loss-ratio on paper-only work — the fix is cheap (narrower spec prose + a Non-goals fence) and restores premium the moment specs stop naming live detector files as Targets without a code change or explicit scope fence.

    Open the signed attestation →

22. 2026-08-28 · 🎩 operator · offPct 30

    fix(club): a code comment is never a chapter — fences suspend structure parsing

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 22→13 over 3 commits since

    Open the signed attestation →

23. 2026-08-28 · 📐 architect · offPct 28

    feat(elicit): a graduated spec becomes an attested entry on the physics tape

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 29→15 over 4 commits since

    Open the signed attestation →

24. 2026-08-28 · 🧪 Cursor Laboratory · offPct 31 · mistagged

    fix(elicit): the consent checker was its own TCC client — the grant was there all along

    Why it left the lane: Every changed path in the diff — scripts/chat/aeperm.jxa, scripts/chat/browser-lease.mjs, scripts/chat/browser-lease.sh, scripts/chat/gem-doctor.mjs, scripts/chat/gemini-pulse-tick.sh, tests/chat/safari-consent-gate.test.mjs, tests/chat/receipt-ndjson-is-one-line.test.mjs — sits under scripts/chat/ or tests/chat/. Cursor Laboratory's declared owned surface (swarm-comms/, .thetacog/orchestrator/, thetacoach-crm-mcp/, .claude/settings*.json, .mcp.json, scripts/claude-room, scripts/init-claude-flow-with-crm.sh, scripts/check-mcp-health.sh, punch-list*) contains not one of them. 100% of the diff's mass landed in a directory the room never claims, which is exactly what a 31%-over-threshold offPct measures.

    The semantic pull: The pull is toward scripts/chat/*.mjs and tests/chat/*.mjs — the elicit/TCC/Safari-consent subsystem — because that is where the real code change is (the aeperm.swift→aeperm.jxa migration, the osascript re-routing in browser-lease.mjs/.sh, gem-doctor.mjs, gemini-pulse-tick.sh). The commit is coherent and lands exactly where its own subject matter lives; the divergence is that the commit's Originating-Terminal trailer says 🧪 Cursor Laboratory, whose glob list is CRM/MCP/orchestrator infrastructure and has zero intersection with scripts/chat/.

    Intent–reality gap: The commit claims to be Cursor Laboratory work (Originating-Terminal: 🧪 Cursor Laboratory), but the receipt measured 142 lit reality cells against a 122-cell intent lane with 31% of the reality mass outside it — because the actual work (the aeperm/osascript TCC-client fix) has no code-surface overlap with anything Cursor Laboratory owns. The work is correct; the lane claim is wrong.

    • Audit .workflow/rooms/*.html for all 9 rooms and confirm exactly which room's owned-surface glob is meant to include scripts/chat/** and tests/chat/** — today none of the 9 listed for Cursor Laboratory does, and this is the second time (per PRIOR LESSONS trend) this room's offPct has swung wildly.
    • Add scripts/chat/**, tests/chat/** to that room's glob in the same commit that fixes the routing.
    • Grep past 14 days of Cursor-Laboratory-tagged commits for other scripts/chat/ or tests/chat/ touches to size how often this mistag has already happened silently.
    • Add a room-glob-overlap sanity check (see missing_guard) so a zero-overlap Originating-Terminal is caught before merge, not after the receipt fires.
    • Update COMMIT MESSAGE STRUCTURE guidance in CLAUDE.md to note Originating-Terminal should be cross-checked against the room's owned-surface glob, not just the terminal app the operator happened to be in.

    What we learned: This room's prior three lessons were all facts-only (sensemaking unavailable) and all showed the SAME symptom — offPct swinging wildly commit to commit (24→54, 0→65, 98→33) — which this pass now explains structurally: Cursor Laboratory's owned-surface glob has no path into scripts/chat/, so any commit whose Originating-Terminal is tagged 'laboratory' but whose real work is chat/consent infrastructure will always read as far out-of-lane, regardless of how correct the code is. The wild swings were never content drift — they were the same routing gap firing repeatedly.

    The missing guard (the deliverable): A guard (e.g. tests/ops/room-glob-overlap.test.mjs) that computes, per commit, the intersection between the Originating-Terminal trailer's room glob (from .workflow/rooms/<room>.html) and the actual changed-file list, and fails/warns when the intersection is empty — catching a mistagged room BEFORE the receipt renders instead of after.

    Detector probe: signal attribution — gzip reef 60% · ballistic walk 40%. Fully saturated on both sides (100% of walked nodes lit) with walkTruncated=false — this is not a depth problem, the walk is right-sized but uninformative here because it can't discriminate partial pull once saturated. Adjustment: don't deepen or shallow the walk; fix the reef seed precision (room-glob ownership) instead, since a saturated walk just amplifies whatever the seed already committed to. Experiments: Re-render this exact sha's receipt after adding scripts/chat/**,tests/chat/** to the correct room's glob and confirm offPct drops materially — a direct before/after test of the routing-gap hypothesis. · Backtest the last 14 days of Cursor-Laboratory-tagged commits: measure what fraction of their changed files fall inside vs. outside the room's glob, to check whether the room's 3 prior wild-swing lessons share this same root cause. · Ablation-test seeding 'aeperm', 'osascript -l JavaScript', 'AEDeterminePermissionToAutomateTarget' into the reef dictionary tied to the correct owning room, then re-measure prewalkOverlap on this and future chat/-domain commits.

    Measured outcome (deterministic, not self-graded): room mean offPct —→24 over 3 commits since · dictionary ablation saved 738B

    A commit that is entirely correct code shipped under a room tag with zero surface overlap reads as unbounded risk to this lane's loss ratio until the glob is fixed, because every future scripts/chat/ commit tagged Cursor Laboratory will re-trip the same false-positive; premium restores once scripts/chat/ has a real owning room and the room-glob-overlap guard exists to catch the next mistag before it renders.

    Open the signed attestation →

25. 2026-08-28 · 📐 architect · offPct 35

    fix(elicit): a Spec: trailer must name a spec that describes THIS commit — 6 of 12 did not

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 17→31 over 6 commits since

    Open the signed attestation →

26. 2026-08-28 · 📐 architect · offPct 25 · mistagged

    feat(panel): the delegation email says what the spec claimed and what the commit did

    

    Why it left the lane: The diff's actual vocabulary is pipeline arithmetic — `execSync(`git show --numstat --format= ${sha}`)`, the `TARGETS` regex capture, `esc()`, `\u2212`, `toLocaleString()` — plus a brand-new 109-line test file under `tests/pmu-simulator/`. None of that is docs/strategy prose; it's the exact code-and-test vocabulary of C.Operations pipeline mechanics. The claimed lane came from `Originating-Terminal: 📐 architect`, whose declared surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, CLAUDE.md) is pure A.Strategy documentation. The reef seeded the intent lane from that room-tag, not from the file paths actually touched, so a quarter of the lit reality mass (offPct=25) landed on scripts/pmu/commit-triptych.mjs and tests/pmu-simulator/ — real code coordinates the intent lane never named.

    The semantic pull: Room-tag terms ('architect', docs/strategy vocabulary implicit in the claimed lane) pulled the intent seed toward A.Strategy, while the diff's own strings — `git show --numstat`, `TARGETS`, `SPEC CLAIMED`/`COMMIT DID`, the test harness's brace-walking `extractBlock()` — are unmistakably C.Operations pipeline/execution vocabulary. The two vocabularies don't overlap, so the walk fanned out from a seed that never covered the file paths it needed to.

    Intent–reality gap: The commit message and trailers describe correct, well-tested, additive pipeline work — the code is fine and the guard proves it. What's wrong is upstream: the commit was filed under 📐 architect, a room whose owned-file-surface is docs/strategy/CLAUDE.md, while the actual work is scripts/pmu/ + tests/pmu-simulator/, surface that plainly belongs to 🔨 builder or 🎩 operator (both of which the commit's own Relevant-Rooms trailer already names).

    • Add the scripts/pmu/**, tests/pmu-simulator/** exclusion to .workflow/rooms/architect.html's owned-surface block
    • Wire the drift receipt's claimed-lane input to the same glob-scoring function that generates the Relevant-Rooms trailer, not to --print-label output
    • Backfill-check the other 71 architect out-of-lane events from the last 14 days: what fraction touch scripts/pmu or tests/ paths (same mistagging) vs genuine docs/strategy drift
    • Add a pre-commit warn (not block) when Originating-Terminal's owned-surface glob matches zero staged files, suggesting the Relevant-Rooms top pick as the correct Originating-Terminal instead
    • Document in CLAUDE.md's room table that Originating-Terminal is physical-presence metadata, and Relevant-Rooms (content-matched) is the lane the receipt should score against

    What we learned: The receipt's claimed-lane input (Originating-Terminal, detected by which terminal app is physically open) and the mechanism that actually knows which room a commit's files belong to (the glob-scoring that produces the Relevant-Rooms trailer) are two different computations that were never wired together — so any commit authored from a terminal doing work outside that terminal's declared surface registers as drift even when the code and the tagging inside the commit message are both correct.

    The missing guard (the deliverable): A test asserting that for every commit, the Originating-Terminal's owned-surface globs (from `.workflow/rooms/<room>.html`) intersect at least one changed file path — or, if not, that Relevant-Rooms names a different top room, proving the mismatch was detected and recorded rather than silently measured as drift.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Saturated, not under-used: litIntent==walksIntent and litReality==walksReality (100% fill on both sides), with walkTruncated=true. A fully-lit walk that was also truncated suggests it hit its ply/time budget before it could differentiate in-lane from off-lane mass by depth. Adjustment: raise --max-depth or the ply budget specifically for commits that mix prose-heavy strings (HTML email templates) with code-heavy strings (execSync/regex) in one diff, since that span is wider than the walk's default budget assumes. Experiments: Recompute this commit's claimed-lane using Relevant-Rooms glob-scoring instead of Originating-Terminal, re-run the receipt offline, and measure whether offPct drops below 15 · Re-score all 71 architect-room out-of-lane events from the last 14 days against Relevant-Rooms glob-matching and measure what fraction were mistagged (terminal ≠ file-surface) vs genuinely off-lane content · Raise --max-depth for commits whose diff mixes HTML/email-string prose with execSync/regex code (detect via a simple heuristic: >2 distinct file extensions with >30% HTML-tag density in one) and measure whether litIntent/litReality diverge from 100% saturation

    Measured outcome (deterministic, not self-graded): room mean offPct 28→36 over 15 commits since · dictionary ablation saved 953B

    A mistagged-not-miscoded event is a labeling loss, not a code loss — it inflates this lane's apparent loss-ratio without any actual defect, so premium is restored by fixing the claimed-lane computation (Relevant-Rooms glob-scoring) rather than by disciplining the room's engineering output, which was already correct and tested.

    Open the signed attestation →

27. 2026-08-28 · 📐 architect · offPct 23 · mistagged

    test(elicit): guard the two silent-success defects the port surfaced

    Why it left the lane: The diff is a pure test file — 'ingest-cli-does-not-eat-flags.test.mjs' — testing arg-parsing/CLI-flag-eating behavior in scripts/elicit/turn-identity.mjs (an elicit/ingest concern, not Operations.Grid/Loop/Flow content). The strings 'execFileSync', 'CLI', '--ingest --full', and 'mkdtempSync/rmSync' pull toward process-orchestration and pipeline-throughput vocabulary that the reef maps onto B1.Tactics.Speed (throughput/velocity), C1.Operations.Grid (topology/exchange), C2.Operations.Loop, and C3.Operations.Flow — the 31 encircled cells are exactly the operations/tactics lattice a CLI-args-and-directory-scanning test superficially resembles, even though the commit's real content is a silent-success guard for one script, not operational infrastructure.

    Intent–reality gap: The message claims narrow, surgical debt-payment — 'guard the two silent-success defects the port surfaced' in one named module (turn-identity.mjs) — but the receipt measured the work landing broadly across 31 out-of-lane Operations/Tactics cells with 100% off-lane share, because a test file about CLI argument handling and directory scanning reads, lexically, like generic pipeline/throughput content.

    • Add a 'Scope:' trailer line to elicit/ingest commits naming the exact module under test, so intent-walk seeding has a tight anchor
    • Add tests/elicit/ to a room-owned-surface glob (see mechanical_intervention) so future elicit-CLI test commits seed correctly instead of defaulting to Operations vocabulary
    • Cross-check scripts/elicit/turn-identity.mjs's own header/doc-comment states its lane (elicit/ingest, not pipeline-ops) so the walk's reef seed picks it up
    • Re-run this commit's receipt after the Scope trailer lands (on a follow-up commit) to confirm offPct drops for elicit-test commits specifically

    What we learned: A test file whose PROSE is dominated by process/CLI/pipeline nouns (execFileSync, --full, directory scanning, exit codes) drifts onto the Operations/Tactics lattice by lexical resemblance even when its actual subject is a single narrow module's argument-parsing bug — the reef cannot distinguish 'testing a CLI' from 'building a CLI pipeline' without an explicit scope fence in the message.

    The missing guard (the deliverable): A commit-message linter for tests/elicit/**/*.test.mjs that requires a 'Scope:' trailer naming the guarded module/function, so the intent walk has a named anchor narrower than generic CLI-argument vocabulary — absence of this trailer should itself be flagged (warn-only) in hooks/post-commit alongside the existing drift receipt.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. Right-sized in spread (litIntent==litReality==125, fill 0.98% well under the 70% saturation ceiling) but arguably too shallow in disambiguation power at 2 plies — a 2-ply walk from a CLI-testing seed can't distinguish 'CLI test for one narrow module' from 'CLI/pipeline infrastructure work', both of which share row/column anchors on Operations/Tactics. Concrete adjustment: for commits whose diff is 100% under tests/elicit/**, seed the walk from the tested module's own doc-comment lane (turn-identity.mjs's header) in addition to the diff text, rather than from the test file's prose alone. Experiments: Run this exact commit (9a6e1535c7) through the walk with prewalkOverlap forced to include scripts/elicit/turn-identity.mjs's header comment as an extra seed source, and measure whether offPct drops below 15 — tests whether module-doc seeding fixes the mistag without a Scope trailer · Ablate: recompute offPct for the last 10 tests/elicit/*.test.mjs commits with and without 'execFileSync'/'CLI'/'--ingest' tokens masked from the intent-walk seed, and compare mean offPct — tests whether CLI-process vocabulary is the specific driver of the Operations/Tactics pull · Add a synthetic control commit (a tests/elicit/ test with zero process/CLI vocabulary, pure assertion logic) and compare its offPct to this one — isolates whether the lattice-lattice mistag is vocabulary-driven or path-driven (tests/elicit/ itself pulling toward ops)

    Measured outcome (deterministic, not self-graded): room mean offPct 24→9 over 9 commits since · dictionary ablation saved 927B

    A 100%-off-lane, 31-cell mistag on a debt-paying guard commit reads as noisy signal rather than true architectural drift, which dilutes this lane's loss-ratio precision; premium is restored by the owned-surface glob addition, which lets future elicit-test commits seed correctly and keeps genuine Operations/Tactics breaches distinguishable from lexical false positives.

    Open the signed attestation →

28. 2026-08-28 · 📐 architect · offPct 45 · mistagged

    test(elicit): assert the subject's PROPERTY, not its phrasing

    Why it left the lane: The diff is entirely inside tests/elicit/tape-fills-from-the-beat.test.mjs — a single assert.match call replaced by two. The commit message and Spec trailer (s-0155.no-new-commit-may) name a general META-RULE about guard-writing discipline ('assert the subject's PROPERTY, not its phrasing'), which reads as governance/testing-discipline prose, not a Signal/Grid-lane change. The lit reality mass (142 cells, litReality=142 vs litIntent=98) spread across Tactics.Speed/Deal and Operations.Grid/Loop/Flow blocks because the commit's own language ('THIRD TIME TODAY', 'the pattern is mine and worth naming', 'checks what the subject must actually DO') is abstract self-diagnosis vocabulary that grips broadly across topology/velocity/throughput/exchange terms rather than concentrating in one cell — a one-line regex fix produces a wide, diffuse semantic footprint because the JUSTIFICATION prose is much larger and more general than the CODE change it defends.

    The semantic pull: Strings like 'the pattern is mine and worth naming', 'checks what the subject must actually DO', and the trailer's 'Story: The one red test in 351 was a guard punishing a change the operator asked for' are generalized process/meta-commentary — they read as touching topology (B1,C1), velocity (B1,C2), throughput (B1,C3), and exchange (B2,C1/C2/C3) because that vocabulary is domain-agnostic ('this changes how we verify things everywhere') rather than scoped to the one test file it edited.

    Intent–reality gap: The commit claims a narrow, mechanical fix ('assert the subject's PROPERTY, not its phrasing' — one regex swapped for two) but the receipt measured reality mass 45% outside the declared lane because the commit MESSAGE itself is written as a broad structural lesson ('third time today', 'the pattern is mine') rather than as a scoped changelog entry, so the ingest reads the prose as staking a claim across many coordinates the actual 12-line diff never touches.

    • Move the 'THIRD TIME TODAY...pattern is mine' generalization out of the commit-message body into a durable location (e.g. a `docs/incidents/` note or a CLAUDE.md addendum under THE PROXY IS NOT THING class 2) so future commits citing this lesson link to it instead of restating it in prose each time.
    • When a commit fixes ONE test file but the message narrates a REPO-WIDE pattern, split into two commits: the scoped fix, then a separate `docs(...)`/`chore(...)` commit that graduates the lesson (per DELEGATE-TO-SUBTASKS's own rule about locked coordinates graduating into CLAUDE.md).
    • Add a lightweight commit-message lint (warn-only) that flags phrases like 'third time today', 'the pattern is mine', 'worth naming' as signals the message is making a repo-wide claim disproportionate to a single-file diff.
    • Cross-reference specs/s-0155 in the test file itself (a one-line comment pointing to the spec) so the intent-lane and reality-lane share an explicit anchor beyond the commit trailer.

    What we learned: A commit whose CODE diff is one file but whose MESSAGE narrates a repo-wide meta-pattern ('third time today') produces high offPct not because the work drifted, but because the message's claim-scope is larger than its diff-scope — the receipt is measuring the MESSAGE's breadth as much as the code's landing, and THE PROXY IS NOT THE THING rule (hard-coded strings vs properties) applies recursively to commit messages themselves: a message that asserts a broad LITERAL lesson from a narrow diff is the same proxy-for-construct error one level up.

    The missing guard (the deliverable): A commit-message lint (warn-only, added to hooks/pre-commit) that computes diff file-count vs message breadth-signal (phrases like 'third time', 'the pattern is', 'always', 'every guard') and warns when a single-file/single-assert diff pairs with repo-wide claim language — prompting the author to split into a scoped commit + a separate graduated-lesson commit.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. litReality (142) exceeding litIntent (98) by a wide margin on a 12-line single-file diff suggests the walk is over-saturating on the reality side relative to a narrowly-scoped commit — adjust by capping walk depth/fanout when diff file-count is 1 and diff line-count is under ~20, so a small mechanical fix doesn't recruit as much lattice mass as a structural one. Experiments: Experiment: re-run the ballistic walk on this same commit with --max-depth reduced by 1 ply and measure whether litReality drops toward litIntent (98) without losing the true out-of-lane signal — if it converges, small diffs should get a shallower default depth. · Experiment: strip commit-message meta-commentary (paragraphs matching 'third time', 'the pattern is', 'worth naming') from the reality-corpus ingest for single-file diffs and re-measure offPct on this sha and its siblings — if offPct drops significantly, the message-breadth-vs-diff-size mismatch is confirmed as the dominant driver. · Experiment: track offPct against diff-file-count across the last 120 commits (the n120 series already computed) bucketed by 1-file vs multi-file commits — if 1-file commits systematically show higher offPct when the message is long, that's the reef signature to correct for.

    Measured outcome (deterministic, not self-graded): room mean offPct 19→21 over 14 commits since · dictionary ablation saved 1028B

    A single-file test-assertion fix carrying a repo-wide diagnostic claim inflates this lane's measured loss-ratio without any real cross-lane work having occurred — premium is restored by scoping commit-message claims to their actual diff breadth, which the warn-only lint above nudges without blocking the (correct, small) fix.

    Open the signed attestation →

29. 2026-08-28 · 🧪 laboratory · offPct 26

    feat(elicit): close four gemSkill ringer/constants gaps — preheader, sentence gate, spec TTL 4h, staleness hold

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 26→17 over 17 commits since

    Open the signed attestation →

30. 2026-08-28 · 📐 architect · offPct 23

    test(elicit): reach the WRITER, not just the backfilled tape

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 15→10 over 4 commits since

    Open the signed attestation →

31. 2026-08-28 · 🎩 operator · offPct 43

    fix(book): the fishbone locates its chapter by content — the URL is a door, not a locator

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 22→18 over 3 commits since

    Open the signed attestation →

32. 2026-08-28 · 📐 architect · offPct 55 · mistagged

    test(elicit): the conformance harness may not count itself as evidence

    Why it left the lane: The diff is entirely a test file about the conformance harness (gemskill-conformance.mjs), self-referential guard logic, grepFiles/SELF-exclusion internals, and the C1.Operations.Grid coordinate the commit itself claims ('tier expensive, C1 Operations.Grid'). The message and code are saturated with operational/testing/pipeline-signal vocabulary — 'harness', 'grep', 'checker', 'verdict', 'evidence', 'gate' — which is exactly the lexicon that scatters across Tactics.Speed/Deal/Signal × Operations.Grid/Loop/Flow: velocity, throughput, exchange, topology. A commit whose whole content IS a meta-test about a measurement instrument reads, to the reef, as touching every operational-signal axis at once because it names the instrument's own vocabulary (grep, verdict, evidence, harness) rather than domain content that would anchor it to one lane.

    The semantic pull: Strings like 'grep -r', 'evidence', 'verdict', 'built', 'walk', 'signal' inside the new test file are generic pipeline/measurement terms that also happen to be the reef's own lexicon for Tactics.Signal (B3) and Operations.Grid/Loop/Flow (C1/C2/C3) — the encircled regions. 'grepFiles', 'SELF', 'recursive grep', 'raw' pull toward Operations.Grid (topology/structure-scanning); 'defined in N file(s)', 'reports', 'evidence line' pull toward Tactics.Signal (measurement/reporting) and Operations.Loop/Flow (velocity/throughput of the check itself running end-to-end).

    Intent–reality gap: The commit claims a narrow, single-purpose test addition ('the guard asserts the SELF exclusion is present... and runs the harness') scoped to elicit/conformance testing, but the receipt measured the work landing 100% out of its declared lane, spread across 31 orthogonal Tactics×Operations cells — the claim was a tight unit test, the measured placement reads as a broad cross-cutting infrastructure change.

    • Add the two-coordinate trailer (C1.Operations.Grid + B3.Tactics.Signal) to future commits touching tests/elicit/conformance-excludes-itself.test.mjs
    • In scripts/elicit/gemskill-conformance.mjs, tag the SELF-exclusion block with a coordinate comment (e.g. '// coord: C1.Operations.Grid') so future greps/tests inherit the correct lane without re-deriving it
    • Extend tests/ops/rules-for.mjs's directory-rule lookup to cover tests/elicit/ so a future self-referential-harness test surfaces this drift class before commit, not after
    • Cross-link this test file's header comment to CLAUDE.md's THE PROXY IS NOT THE THING section (class 2: token standing in for a construct) since that is the literal failure this guard targets

    What we learned: A commit that writes a test ABOUT a measurement/reporting instrument (grep, verdict, evidence, harness) will itself be scored by that same reef as touching the operational-signal axes it's merely describing in prose — self-referential test commits are structurally prone to this drift class because their vocabulary IS the reef's lexicon, independent of actual file scope.

    The missing guard (the deliverable): A pre-commit check that flags when a test file's own prose (comments + assertion messages) reuses ≥N reef-lexicon terms (grep/verdict/evidence/walk/signal/harness) without a matching multi-coordinate trailer — i.e. detect vocabulary-lane mismatch before the receipt does, the same class this commit's own guard was built to catch but pointed at itself.

    Detector probe: signal attribution — gzip reef 30% · ballistic walk 70%. Right-sized in ply count (fill 0.98%, well under the 70% saturation ceiling) but the litReality/litIntent gap (120 vs 117, only 3 cells wider) combined with 31 out-of-lane hits suggests the walk is finding real orthogonal signal, not noise — no adjustment to depth needed; the fix is upstream (spec precision), not walk tuning. Experiments: Re-run this same commit's ballistic walk with the proposed two-coordinate trailer added and confirm offPct drops materially — an ablation test of the spec_rewrite hypothesis · Measure prewalkOverlap across the last 20 architect-room commits touching tests/elicit/*.mjs to see if self-referential test files systematically show overlap<5 (a detectable pre-signal for this drift class) · Add 'harness', 'verdict', 'evidence' as a tracked reef-lexicon-overlap metric emitted alongside offPct in the receipt row, so vocabulary-pull can be distinguished from actual file-surface pull in future diagnoses

    Measured outcome (deterministic, not self-graded): room mean offPct 31→56 over 4 commits since · dictionary ablation saved 953B

    One 55%-off-lane self-referential-test event nudges this lane's loss ratio up but doesn't threaten the premium on its own (73 prior events already priced into the room's baseline); restoring it means shipping the vocabulary-lane mismatch pre-check so the next self-referential-harness commit is caught before the receipt, not after.

    Open the signed attestation →

33. 2026-08-28 · 📐 architect · offPct 23

    feat(elicit): the compiled spec IS the panel's intent corpus — no change to the panel, the seam existed

    

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 27→30 over 6 commits since

    Open the signed attestation →

34. 2026-08-28 · 🎩 operator · offPct 26 · structural

    fix(club): repair the stale 'The ask' anchor in the statement guard — red since bf2ab0044d

    Why it left the lane: The commit's Originating-Terminal trailer declares 🎩 operator as the intent lane, so the receipt checks the reality walk against operator's owned surface (data/trust-debt/, pipeline/, accounting/, dashboard*, pmu-demo). But the entire diff is a one-line repair inside tests/ops/bookclub-statement.test.mjs — pure book-club/newsletter comms vocabulary. The ballistic walk seeded off strings like 'P.S. — the quiet part', 'The physics is not in this email', 'one-door streamline', and 'the rest of the email must be unaffected', all of which cluster in the comms/newsletter coordinate region the Relevant-Rooms trailer itself already names (🔨 builder, 🧪 laboratory) — not in operator's trust-debt/dashboard region. 26% of the 127 lit reality cells landed in that comms cluster, outside the operator-only lane the receipt checked.

    The semantic pull: 'bookclub-statement', 'the quiet part', 'the physics is not in this email', 'one-door streamline', 'the ask' — every load-bearing noun in the diff is comms/email vocabulary, none of it trust-debt/pipeline/dashboard vocabulary, so the walk's mass naturally seeds toward the builder/laboratory comms coordinates rather than operator's owned cells.

    Intent–reality gap: The commit's Originating-Terminal claims the operator lane, but 100% of the changed content is a book-club email-guard test fix with zero relationship to trust-debt/pipeline/dashboard work — the only thing 'operator' about this commit is which terminal happened to be open when it was typed.

    • Add a regression fixture: a commit with Originating-Terminal=operator and Relevant-Rooms including builder/laboratory touching a builder-owned file must NOT trip the lane threshold once the union fix lands.
    • Audit the room's other 6 out-of-lane events in the last 14 days for the same Originating-Terminal-vs-Relevant-Rooms mismatch pattern before assuming each needs its own diagnosis.
    • Update `.workflow/rooms/operator.html` (or wherever owned-surface globs live) with a comment noting operator is expected to fire cross-room test/hotfix commits and that Relevant-Rooms, not Originating-Terminal, is the lane authority.
    • Add a `Lane-Note: cross-room fix, expected` trailer convention for the rare case a fix is genuinely undocumented cross-room work, so the receipt can distinguish 'known divergence' from 'true drift' even before the union fix ships.

    What we learned: The repo's own commit convention explicitly permits Originating-Terminal to diverge from content ownership ('that's fine and worth recording'), but the drift-receipt's lane check was never updated to read Relevant-Rooms instead of Originating-Terminal — so the receipt is structurally guaranteed to flag every legitimate 'fixed someone else's test from my terminal' commit as drift, independent of the operator room specifically.

    The missing guard (the deliverable): A test asserting the drift-receipt's intent-lane computation includes the Relevant-Rooms-named rooms' owned-surface globs, not just the Originating-Terminal room's — fired against a synthetic commit with intentionally divergent Originating-Terminal/Relevant-Rooms to prove it stays green.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Saturated, not shallow: every walked cell lit on both sides (121/121 and 127/127), for a one-line, single-file diff. That's disproportionate spread for the size of the actual change — the walk radius/ply budget isn't scaling down for small diffs, so a tiny fix gets the same full-lattice reach as a large one, diluting lane precision. Adjustment: scale max-depth/ply budget to diffstat size (e.g., single-file/<10-line diffs get a reduced-ply walk) so small fixes don't spread mass across unrelated coordinate regions. Experiments: Re-run this commit's ingest restricted to only the changed lines (not whole-file context) and measure whether offPct drops below 15 — isolates whether whole-file ingestion is the source of the spread. · Recompute offPct for the room's last 7 flagged events using the Relevant-Rooms-union rule and report how many resolve below threshold — validates the code_rewrite's expected impact before shipping it broadly. · Correlate prewalkOverlap against final offPct across this room's 14-day history to test whether prewalkOverlap<5 is itself a cheap leading indicator of offPct>15, enabling a pre-walk early-exit flag.

    Measured outcome (deterministic, not self-graded): room mean offPct 15→28 over 15 commits since · dictionary ablation saved 809B

    A structural false-positive in the lane check inflates this room's apparent loss ratio without any real scope creep, so premium should stay flat pending the Relevant-Rooms-union fix rather than pricing in a mechanism that doesn't reflect actual drift risk.

    Open the signed attestation →

35. 2026-08-28 · 📐 architect · offPct 46

    feat(elicit): the gate gets its missing four checks — six, as specified

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 34→16 over 5 commits since

    Open the signed attestation →

36. 2026-08-28 · 🎩 operator · offPct 16 · structural

    book(ch8): The Compression Witness reads clean cold — seven edits from a sentence-level ghost-read

    Why it left the lane: The diff adds a literal citation of the operator room's own instrument — "`scripts/pmu/sigma-localize.mjs` in `github.com/wiber/thetacog-mcp`" — plus dense repeated PMU vocabulary ("hundred and forty-four" anchors/zones repeated 6+ times, "compression witness", "matched aperture", "gzip", "sigma 8.85", "the family law"). None of that lives in books/tesseract/ conceptually — it IS the vocabulary and even the literal path of data/pmu/ and scripts/pmu/, which sit inside 🎩 operator's owned surface. gzip-NCD compresses shared strings regardless of which file they sit in, so a passage that quotes the instrument's own path and re-explains its own significance bands will always collide with the coordinate the instrument itself occupies.

    The semantic pull: literal string `scripts/pmu/sigma-localize.mjs` + `github.com/wiber/thetacog-mcp`, plus repeated "hundred and forty-four"/"anchor"/"sigma"/"compression witness"/"gzip" terms pulling reality mass toward the PMU-instrument coordinate lane that 🎩 operator's data/pmu/pipeline and data/pmu/claudbridge occupy.

    Intent–reality gap: The commit claims pure book-prose work (Persona-Intent: a cold reader clearing undefined nouns; Relevant-Rooms: voice/builder/network, no operator) but the receipt measured 16% of reality mass landing outside that lane — because the prose it edited is a technical description of the operator room's own PMU instrument, quoting its real path.

    • Add a commit-message lint that extracts literal repo-relative paths (regex over the diff body, e.g. `scripts/[\w-]+/[\w.-]+`, `data/[\w-]+/`) and cross-checks each path's owning room (from `.workflow/rooms/*.html` owned-surface globs) against the Relevant-Rooms trailer
    • Document the citation-bleed pattern in CLAUDE.md's RESEARCH → BOOK → BLOG process: quoting a live instrument path in prose is REQUIRED (per 'name an argument, link its post') and will predictably register 10-20% reality mass in that instrument's room lane — this is expected, not a defect, once declared
    • Re-run this exact commit's offPct after seeding the gzip dictionary with the two literal strings (see compression_dictionary_adds) to confirm the citation, not unrelated drift, is the driver
    • Add the trailer-completeness check as a warn-only pre-commit gate alongside the existing lexical/Six-Needs checks, matching the non-blocking posture of the other Relevant-Rooms convention rules

    What we learned: Book prose that names/links a live instrument's own file path (a HARD RULE requirement — 'Name an Argument, Link its Post') structurally imports part of that path's coordinate lane into the reality mass; the repo currently has no mechanism to declare that tension as expected, so every future PMU-documenting book passage will keep tripping this exact threshold unless the citation is pre-fenced in the trailer rather than caught after the fact.

    The missing guard (the deliverable): A commit-msg/pre-commit check that scans the diff for literal repo-relative paths cited in prose and asserts the Relevant-Rooms trailer names the room owning each cited path's glob (cross-referenced against `.workflow/rooms/*.html`), warning (not blocking) when a citation's owning room is missing from the trailer.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. litIntent (134) and litReality (128) are close and both well below saturation (144), so the walk is neither too shallow nor saturated — it's right-sized for this commit. Concrete adjustment: bump --max-depth by one ply on this specific commit and re-measure offPct to test whether the PMU-citation cells converge back toward the book lane at greater depth (would indicate a shallow-walk artifact) or hold steady (would confirm genuine content-driven placement). Experiments: Re-run the ballistic walk on sha c1bd8fab82 with --max-depth+1 and diff offPct against this run — tests whether the 16% is a depth artifact or stable placement · Seed the gzip baseline dictionary with 'scripts/pmu/sigma-localize.mjs' and 'github.com/wiber/thetacog-mcp', recompress this commit's reality corpus, and compare offPct before/after — isolates whether the literal path string alone drives the collision · Build a held-out corpus split (book chapters citing a PMU/reef path vs. chapters with no such citation) and measure mean offPct per group across the last 20 book commits — turns the citation-bleed hypothesis into a measured rate instead of a one-commit inference

    Measured outcome (deterministic, not self-graded): room mean offPct 21→29 over 6 commits since · dictionary ablation saved 967B

    A citation-driven 16% off-lane reading against a 15% threshold is a cheap, self-inflicted claim on 🎩 operator's lane (12 events in 14 days already) that a one-line trailer convention resolves for near-zero cost, restoring the room's loss ratio without touching the prose itself.

    Open the signed attestation →

37. 2026-08-28 · 📐 architect · offPct 39 · mistagged

    fix(elicit): a guard was demanding the leak it should have caught

    

    Why it left the lane: The diff is a single test-file edit (tests/elicit/spec-is-panel-intent.test.mjs) that touches assertion strings naming env vars, file-write calls, and a security narrative ('leak', 'redacted', 'publishes operator quotes onto public/commit/'). None of that vocabulary is git/commit/push/iteration-bundle terms — it's elicit-panel/security-seam terms, so the reef's gzip-NCD seed and the walk both pull toward B3.Tactics.Signal (signal/leak/redaction reads as 'signal integrity') and C1/C2/C3 Operations.Grid/Loop/Flow (writeFileSync, env.PMU_SPEC_INTENT, panelPath — pipeline plumbing terms) rather than staying in the C,B3 git-discipline lane the room claimed via commit trailers.

    The semantic pull: Strings 'env.PMU_SPEC_INTENT = panelPath', 'writeFileSync(specPath', 'writeFileSync(panelPath', 'the redacted copy is what the panel reads' are pipeline-plumbing/data-flow vocabulary — they describe routing a value through a panel/executor, which is exactly what Operations.Grid/Loop/Flow encode (routing, exchange, throughput). The commit message's own words 'guard', 'leak', 'security fix', 'red suite' pull toward Signal (B3) as an integrity/detection concept. Nothing in the diff or message contains git/commit/push/hook vocabulary, so C,B3 (git-discipline) has zero organic pull and is only the room's SELF-CLAIMED lane via the trailer, not a lane the content itself supports.

    Intent–reality gap: The commit message frames this as a git-discipline story ('I committed the leak fix before running the full suite... The suite is the check on me') and the room trailers claim 📐 architect/git-discipline, but the actual work is a one-line test-assertion fix in the elicit/panel security seam — the receipt correctly measured the content landing in Operations.Grid/Loop/Flow and Tactics.Signal, not in the git-workflow lane the message narrates around.

    • Rewrite the commit's Story trailer (per spec_rewrite) so future drift receipts for this file don't get read against git-discipline vocabulary
    • Add tests/elicit/**.test.mjs and scripts/elicit/**.mjs to a room's owned-surface glob in .workflow/rooms/*.html so elicit-panel work has a home lane other than the default git-discipline fallback
    • Cross-reference CLAUDE.md's 'THE RECEIPT IS LLM-FREE' section — this commit is exactly the class of fix that section exists to protect (a test that encoded leaking behaviour); add a one-line pointer there so elicit/panel test-drift is recognized as its own recurring class
    • File a follow-up: audit other tests under tests/elicit/ for the same 'test encodes the insecure behaviour' pattern now that one instance was caught by a canary rather than review

    What we learned: A commit's own narrative voice ('I committed X before running the suite... the suite is the check on me') can be pure git-process prose glued onto a payload that is 100% pipeline-internals work — the trailers get the room right by convention (whoever's terminal was open) but the receipt is scoring the DIFF, not the prose, so self-narrated process commentary is a durable source of false git-discipline placement distinct from the three prior lessons here (which were facts-only, no mechanism named).

    The missing guard (the deliverable): A pre-commit check that flags when a commit's Relevant-Rooms trailer names a room whose owned-surface globs (.workflow/rooms/*.html) do NOT include any of the staged file paths — i.e. verify the trailer's room claim against actual glob ownership of the diff before allowing the commit, rather than trusting the terminal-detected default.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. litReality (123) exceeding litIntent (111) with only 2 plies and walkTruncated=true suggests the walk is UNDER-budgeted for this content class relative to how far the reality terms actually propagate through the Operations sub-lattice — one concrete adjustment: raise --max-depth by 1 specifically when prewalkOverlap<5 (a proxy for 'seed barely matched the claimed lane'), since a low-overlap seed needs more plies to find whether reality genuinely diverges or just needed one more hop to reconverge with intent. Experiments: Re-run this exact commit's ingest with --max-depth+1 and diff offPct against the current 39 to test whether under-budgeting (not genuine drift) explains part of the spread · Compute prewalkOverlap distribution across the room's last 70 out-of-lane events; if most cluster <5 like this one, that's evidence the git-discipline reef seed is too narrow for test-file diffs specifically, testable by adding 3-5 test-file-diff exemplars to the git-discipline seed corpus and re-measuring mean prewalkOverlap · A/B the trailer-vs-glob guard proposed in missing_guard against the last 70 events: replay each commit's Relevant-Rooms trailer against .workflow/rooms/*.html globs and count how many would have been flagged pre-commit, giving a measurable baseline for how many of the 70 this class of guard actually catches

    Measured outcome (deterministic, not self-graded): room mean offPct 27→25 over 11 commits since · dictionary ablation saved 793B

    An out-of-lane security-seam fix mistagged as git-discipline inflates this lane's apparent breach rate (47.5% n120) with false-positive noise that isn't really git-process risk, diluting the signal underwriters would price against real git-discipline failures; restoring it means the trailer-vs-glob guard so future elicit/panel fixes stop landing in — and skewing — this lane's loss ratio.

    Open the signed attestation →

38. 2026-08-28 · 🧪 laboratory · offPct 24

    test(bookclub): the generator now has a test that runs it — deleting a helper turns the suite red

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct —→26 over 4 commits since

    Open the signed attestation →

39. 2026-08-28 · 📐 architect · offPct 17

    fix(pmu,chat): the spec is placed (not pinned), and a starved probe is not a missing tab

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 17→8 over 4 commits since

    Open the signed attestation →

40. 2026-08-28 · 📐 architect · offPct 32 · structural

    feat(elicit): funnel three — readiness is a compiler, and the funnel mails itself as text

    

    Why it left the lane: The diff is 447 insertions across scripts/chat/gem-receipt-email.mjs, scripts/chat/gemini-pulse-tick.sh, scripts/elicit/autonomy.mjs, scripts/elicit/elicit-execute.mjs, scripts/elicit/spec-compiles.mjs, scripts/elicit/spec-emit.mjs and tests/elicit/* — every touched path sits outside 📐 architect's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). Strings like `DENY = [...]`, `class 1/2/3`, `liveClasses()`, `autonomy.ndjson`, `spec-compiles.mjs` are runtime gate/ledger/compiler logic — pure Operations-lane engineering, not strategy or business-doc prose. The commit's own Relevant-Rooms trailer already names 🧪 laboratory as the better fit, so the room's own metadata disagrees with the room the receipt was scored against.

    The semantic pull: Terms like 'compiler', 'TTL'd receipt', 'ledger', 'DENY/ALLOW', 'class 1/2/3 autonomy', 'spec compiles', 'executor' are all control-plane/automation vocabulary that pulls placement toward Operations.Loop/Operations.Grid-style coordinates — the row→column recursion lights up cells describing gates and execution pipelines, not the Strategy/Product-doc cells 📐 architect's owned surface anchors to. No region narrative is cached for this sha, so the exact [row,col] pair can't be quoted, but the diff's own vocabulary is unambiguously engine-build, not strategy prose.

    Intent–reality gap: The commit message frames the work through architecture-adjacent language ('one lattice, both halves', 'tesseract on both ends') and fires under Originating-Terminal: 📐 architect, but the receipt measured zero mass in any file architect actually owns — the room label came from which terminal the operator was physically sitting in, not from what the commit built.

    • Locate the receipt-generation script that reads Originating-Terminal for the room field and confirm it is the drift source (grep for 'Originating-Terminal' across scripts/pmu/)
    • Add scripts/elicit/**, tests/elicit/**, scripts/chat/gem-*.mjs, scripts/chat/gemini-pulse-tick.sh to 🧪 laboratory's owned-surface glob in data/rooms.json / .workflow/rooms/laboratory.html
    • Add the CLAUDE.md clarification sentence to the COMMIT MESSAGE STRUCTURE section (quoted in spec_rewrite)
    • Ship tests/pmu-simulator/room-lane-from-surface-not-terminal.test.mjs using cf601fdc9 as the fixture: Originating-Terminal=architect, Relevant-Rooms=[architect,laboratory,voice], changed files all outside architect's globs — assert the computed lane is laboratory, not architect
    • Backfill-check the room's other 65 out-of-lane events in the last 14 days for the same terminal-vs-surface pattern before assuming a single fix explains all of them

    What we learned: The receipt's per-room offPct is being scored against Originating-Terminal (a proxy for 'where the operator physically sat') rather than against owned-surface glob matches on the changed files (the construct for 'whose lane this is') — and CLAUDE.md's own commit-trailer convention explicitly documents that these two fields are allowed to diverge, which means the current scoring guarantees a false-positive-shaped offPct spike every time an operator builds pipeline code while sitting in a room whose lane doesn't own that surface. Prior interventions for this room (offPct 14→28, 17→26, one success 14→1) never named this because none inspected which field drives the room attribution itself.

    The missing guard (the deliverable): tests/pmu-simulator/room-lane-from-surface-not-terminal.test.mjs — a fixture commit (this sha) with Originating-Terminal and Relevant-Rooms deliberately diverging from the changed-file owned-surface globs, asserting the receipt's room/lane field is computed from surface-glob matches on the diff, never from the Originating-Terminal trailer alone.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized, not mis-tuned — litIntent(132) and litReality(137) are within 4% of each other and walkTruncated=false, so depth/breadth are not the lever here. No depth adjustment warranted; the fix is upstream of the walk, in what seeds it. Experiments: Re-seed the reef for this commit from owned-surface glob matches (Relevant-Rooms union) instead of raw commit-message text, re-run cf601fdc9 offline, and check whether offPct drops below the 15 threshold · Pull prewalkOverlap for the last 20 architect-room commits and correlate against final offPct — if prewalkOverlap<10 reliably predicts offPct>25, surface that as an early warning on the receipt itself before the walk even runs · Run a controlled ablation: strip Originating-Terminal from room-selection logic entirely (use only Relevant-Rooms + surface-glob match) for the last 66 architect out-of-lane events and diff the resulting offPct distribution against the shipped one

    Measured outcome (deterministic, not self-graded): room mean offPct 27→27 over 3 commits since · dictionary ablation saved 970B

    An offPct=32 mistagged to a chronically-drifting room (66 events/14d) inflates that lane's apparent loss ratio with false-positive noise; restoring it means re-scoring room attribution off owned-surface matches rather than physical terminal, so the premium reflects real cross-lane work instead of where the operator happened to be sitting.

    Open the signed attestation →

41. 2026-08-28 · 📐 architect · offPct 49

    test(elicit): measure gemSkill.txt against the running code, check by check

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 30→56 over 4 commits since

    Open the signed attestation →

42. 2026-08-28 · 📐 architect · offPct 22

    feat(elicit): the panel resolves its spec from git, not from a live env var

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 33→21 over 18 commits since

    Open the signed attestation →

43. 2026-08-28 · 📐 architect · offPct 25

    fix(ingest): scraped turns carry seen_at, so the store can report its own cutoff

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 31→17 over 13 commits since

    Open the signed attestation →

44. 2026-08-28 · 🎤 voice · offPct 44

    fix(email): one PERMANENT thread per alert type — the date moves out of the subject into the preview line

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct —→5 over 4 commits since

    Open the signed attestation →

45. 2026-08-28 · 📐 architect · offPct 39

    feat(elicit): the executor — work implied by a Gemini round now runs, behind a deny-first gate

    

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 27→31 over 5 commits since

    Open the signed attestation →

46. 2026-08-28 · 📐 architect · offPct 27 · mistagged

    feat(elicit): one thread per spec, four terminal states, and LEARNED is a result

    

    Why it left the lane: The diff's new file scripts/elicit/spec-threads.mjs and the touched scripts/chat/gem-receipt-email.mjs and scripts/elicit/spec-emit.mjs are dense with policy-grade narrative prose — 'LEARNED IS A FIRST-CLASS OUTCOME', 'EVERY THREAD TERMINATES OUT LOUD', 'BLOCKED BY is the cross-spec check the funnel cannot catch on its own', operator quotes in blockquote style — the exact register that normally lives in docs/strategy/ or docs/03-product/. The gzip-NCD reef seed matches that mass against strategy/policy vocabulary even though the four changed files (scripts/elicit/spec-threads.mjs, scripts/elicit/spec-emit.mjs, scripts/chat/gem-receipt-email.mjs, tests/elicit/spec-threads.test.mjs) sit entirely under scripts/ and tests/ — none of which the 📐 architect room's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md) actually covers.

    The semantic pull: litReality=110 vs litIntent=102 with prewalkOverlap only 6 — the two walks share almost no seed cells before diverging, so the ~8-cell breadth difference cannot explain offPct=27 (~30 off-lane cells out of 110). The pull is a seed-placement mismatch, not a spread mismatch: the policy-prose comment blocks (state names SHIPPED/PREEMPTED/MERGED/EXPIRED/LEARNED, the 'thinking out loud is thinking, not instruction' argumentative style) place the reality walk's ballistic seed in strategy-adjacent lattice territory that the intent lane (built from the commit message + spec, itself also prose-heavy) never anchored to the same coordinates.

    Intent–reality gap: The commit's own Relevant-Rooms trailer claims '📐 architect, 🧪 laboratory, 🎤 voice' — architect listed first, implying architect's owned surface is most touched — but the receipt's reality lane is 100% scripts/elicit + scripts/chat + tests/elicit, which architect's owned-surface glob list matches zero times. The commit message text is a strategy-grade design narrative; the actual landed work is a pipeline/engineering module.

    • Add scripts/ops/relevant-rooms.mjs: glob-match staged/changed files against every room's owned-surface list in .workflow/rooms/*.html, print the ranked top-3 — this becomes the thing a session consults before writing the Relevant-Rooms trailer instead of guessing.
    • Declare an owned-surface glob for scripts/elicit/** and tests/elicit/** in whichever room's .workflow/rooms/*.html is meant to hold the elicitation-pipeline engineering surface (currently unowned by any of the 9 rooms' listed globs, which is itself part of why this landed on the nearest prose-heavy room, architect).
    • Extract the state-taxonomy narrative (SHIPPED/PREEMPTED/MERGED/EXPIRED/LEARNED, the BLOCKED-BY rule) out of spec-threads.mjs's header comment into a real doc under docs/03-product/, leaving a one-line pointer comment in the code — this is architect-owned content by right, not code-comment content.
    • Add tests/ops/relevant-rooms-glob-check.test.mjs asserting the guard script returns a non-empty ranked list for the files this commit touched, and that architect does not rank #1 for scripts/elicit/** paths.
    • Update CLAUDE.md COMMIT MESSAGE STRUCTURE section per spec_rewrite above.

    What we learned: This repo's own house style — dense narrative/policy prose living inside .mjs comment headers rather than in docs/ — is what makes the gzip-NCD reef mistake pipeline code for strategy documentation. The prior two out-of-lane events for this room were logged facts-only with no mechanism named; this is the first time the mechanism is visible: the Relevant-Rooms trailer is hand-typed prose with no glob-verification step, so a commit can assert architect relevance while touching zero architect-owned files, and nothing catches that assertion being wrong.

    The missing guard (the deliverable): A commit-trailer check (pre-commit warn, escalating to block if repeat offenses continue) that glob-matches the staged file list against every room's owned-surface list and fails/warns when the commit message's declared Relevant-Rooms has zero overlap with the top-1 computed match.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Right-sized in breadth but walkTruncated=true means saturation could be a budget artifact rather than genuine convergence — full lit==walked on both sides looks identical whether the walk finished naturally or was cut off mid-expansion. Adjustment: emit the ply-decay curve alongside the truncation flag so a future receipt can distinguish 'walked to completion and saturated' from 'hit --max-depth while still expanding.' Experiments: Re-run this commit's reef placement with the policy-prose comment block stripped from spec-threads.mjs and measure whether offPct drops — isolates whether narrative-comment mass is the actual driver. · Seed the reef dictionary with the recurring elicit-pipeline vocabulary (SHIPPED/PREEMPTED/MERGED/EXPIRED/LEARNED/BLOCKED BY) and measure the change in prewalkOverlap on the next elicit-pipeline commit. · Compute prewalkOverlap across this room's last 67 out-of-lane events and test whether prewalkOverlap<10 predicts offPct>20 reliably enough to become a pre-commit early-warning threshold.

    Measured outcome (deterministic, not self-graded): room mean offPct 29→15 over 6 commits since · dictionary ablation saved 870B

    An architect-tagged commit that touches zero architect-owned files erodes the trailer's value as a routing signal, raising this lane's loss-ratio on every downstream process that trusts Relevant-Rooms for triage; restoring it means the trailer is glob-computed rather than asserted, so 'architect' on a commit means architect-owned files actually moved.

    Open the signed attestation →

47. 2026-08-28 · 📐 architect · offPct 20

    fix(elicit): the reality half is semantic, and the spec is bulked to the aperture

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 30→6 over 21 commits since

    Open the signed attestation →

48. 2026-08-28 · 🎩 operator · offPct 23

    fix(club): the metavector pipeline holds end-to-end — door-vs-locator closed at all three joints

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 21→26 over 8 commits since

    Open the signed attestation →

49. 2026-08-28 · 📐 architect · offPct 27

    fix(pmu): the placement delta is calibrated — 2/3 at n=3, so it gates nothing

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 17→25 over 8 commits since

    Open the signed attestation →

50. 2026-08-28 · 📐 architect · offPct 32

    feat(elicit): q8 answered — the narrator is local, by requirement not preference

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 15→17 over 15 commits since

    Open the signed attestation →

51. 2026-08-28 · 📐 architect · offPct 52

    fix(pulse): the receipt writer wrote nothing for 21 hours — and every Gemini round now mails a receipt

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 26→29 over 6 commits since

    Open the signed attestation →

52. 2026-08-27 · 📐 architect · offPct 35

    comms(newsletter): lock A8 — the gate now measures the RENDERER's output, not a file we hand-write

    

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 23→34 over 8 commits since

    Open the signed attestation →

53. 2026-08-27 · 🔨 iTerm Builder · offPct 33

    chore(book): rebuild HTML (detached pre-push auto-heal)

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 11→22 over 14 commits since

    Open the signed attestation →

54. 2026-08-27 · 📐 architect · offPct 18 · structural

    content(blog): the flip — our argument needs no sign, the rebuttal needs a specific one

    Why it left the lane: The diff is pure essay prose in src/content/blog/2026-08-27-take-the-exponent.mdx, and the vocabulary it adds is dense with strategy/legal-argument register — 'burden-shift', 'the whole argument', 'structural feature', 'a far stranger thing to assert', 'load-bearing asymmetry', 'measurement nobody has' — the same register that fills docs/strategy/ and docs/01-business/strategy/, which are 📐architect's actual owned surface. That vocabulary overlap is what gzip-NCD-seeds and then the ballistic walk spreads toward architect's coordinate space, even though the touched file sits nowhere in architect's glob list. No cached region narrative exists for this sha, so the coordinate-level pull can't be pinned further than: rhetoric-about-argument-structure reef-matches strategy-doc prose.

    The semantic pull: 'burden-shift', 'structural feature', 'a negative exponent, and it is a far stranger thing to assert', 'the whole argument' — abstract argument-scaffolding language that is stylistically identical to how docs/strategy/ frames a position, pulling reality mass toward architect's lane despite the file being content(blog).

    Intent–reality gap: The commit's Originating-Terminal/Relevant-Rooms trailers claim 📐architect, but the diff touches only src/content/blog/ — a surface CLAUDE.md's own COMMIT MESSAGE STRUCTURE section explicitly says may diverge from the room ('a commit's originating terminal and its file-relevance to rooms can diverge — that's fine and worth recording'). The receipt's lane-check doesn't encode that documented exception, so a sanctioned divergence gets scored as drift.

    • Audit .workflow/rooms/*.html to find which room's glob (if any) actually claims src/content/blog/**; if none does, assign it explicitly rather than letting it default to whichever terminal committed
    • Trace the receipt pipeline (scripts/pmu/pipeline.mjs or wherever room-lane is derived) to confirm whether it keys off Originating-Terminal or Relevant-Rooms, since that's the crux of the structural gap named above
    • Pull the last 14 days of architect's 60 out-of-lane events and tag how many are content(blog)-typed commits vs true architect-surface work — this determines whether the fix above resolves most or only some of them
    • Add the CLAUDE.md addendum from spec_rewrite in the same commit as any code change
    • Add the missing_guard test below before closing this out

    What we learned: The drift receipt currently treats Originating-Terminal as the lane-claim for every commit type, but CLAUDE.md itself already carves out an exception for terminal/file divergence — the detector and the doc it's supposed to enforce disagree, and architect's unusually high 60-events/14-days rate is the visible symptom of that disagreement, not of 60 separate real mistakes.

    The missing guard (the deliverable): A test asserting: any commit typed `content(blog): ...` touching only src/content/blog/** must be lane-scored against the room whose .workflow/rooms/*.html glob actually claims src/content/blog/**, never against the commit's Originating-Terminal trailer alone.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. Both sides are fully lit (no unlit slack) yet walkTruncated=true — the walk hit its budget while still saturating every cell it reached, meaning the truncation may be cutting off exactly the further-out intent-lane cells that would have explained more of the reality spread as in-lane. Adjustment: re-run this class of philosophy/strategy-adjacent blog commit with a higher --max-depth to see if offPct drops once the intent lane is allowed to extend as far as the walk already wants to go. Experiments: Re-score the last 30 days of content(blog) commits by Relevant-Rooms-glob-match instead of Originating-Terminal and measure whether offPct systematically drops for those commits · Tag a distinct 'rhetoric/argument-structure' reef cluster separate from 'strategy-decision' vocabulary and measure whether architect's out-of-lane rate on blog commits falls once the two are no longer conflated · Run guard-seen-to-fail against a synthetic content(blog)-typed commit tagged architect to confirm the proposed lane-check guard goes red on today's code and green after the fix

    Measured outcome (deterministic, not self-graded): room mean offPct 25→28 over 6 commits since

    This tick is most likely a false-positive against architect's loss-ratio — a documented terminal/file divergence being scored as drift — so it inflates the room's premium without pricing a real defect; keying the lane-check to file surface instead of Originating-Terminal for content(blog) commits restores accurate pricing.

    Open the signed attestation →

55. 2026-08-27 · 🎤 voice · offPct 34 · mistagged

    fix(cook): three rounds of one cook are one thread — the subject carried the four things that vary

    Why it left the lane: The diff's own comment block is doing signal work that reads as topology/velocity/throughput/exchange language rather than plain messaging-fix language: 'threading rides on References, not on Gmail's same-subject heuristic', 'construct', 'daily-thread.ts' as a cross-reference, plus the CLI flags --references and --preheader (protocol-level email routing verbs). Those terms cluster toward B1.Tactics.Speed/B2.Tactics.Deal/C1-3.Operations.Grid/Loop/Flow (routing, exchange, flow-control vocabulary) because the fix is genuinely ABOUT message routing/threading semantics, not just a string edit — the encircled regions (B1,C1 topology · B1,C2 velocity · B1,C3 throughput · B2,* exchange) are exactly the reef's shorthand for 'this changed how something is routed/sequenced', which the subject/preheader/References rewiring literally is.

    The semantic pull: '--references', '--preheader', 'threading rides on References, not on Gmail's same-subject heuristic', and the explicit cross-file citation 'same construct as src/lib/email/daily-thread.ts' are the pull terms — they're protocol/exchange vocabulary (References header = RFC 5322 threading protocol) that the reef reads as Operations.Grid/Loop/Flow and Tactics.Deal/Speed rather than as 'voice content authoring', which is this room's declared lane.

    Intent–reality gap: The commit message frames this as a voice-room content fix ('the subject carried the four things that vary') scoped to one script, but the actual change implements email-protocol-level routing logic (RFC 5322 References threading, preheader injection) that is architecturally an Operations/messaging-infrastructure concern being authored from the 🎤 voice room's file (scripts/blog/cook-round.mjs is not in voice's owned-surface list at all).

    • Add scripts/blog/cook-round.mjs to a 🔨 builder or 🎩 operator owned-surface glob in .workflow/rooms/*.html so threading-protocol edits there are correctly attributed
    • Extract shared References/preheader logic into scripts/email/thread-headers.mjs, referencing src/lib/email/daily-thread.ts as the canonical prior art (the commit message already names this construct — make it code, not a comment)
    • Update CLAUDE.md voice owned-surface bullet to carve out cook-round.mjs's content-only scope (see spec_rewrite)
    • Add a guard asserting scripts/blog/cook-round.mjs's --subject argument never re-grows a round-varying interpolation (the commit's own stated deliverable — verify tests/blog/cook-round-threads.test.mjs exists and covers this)
    • Cross-link src/lib/email/daily-thread.ts and the new thread-headers.mjs module in a comment so the next threading fix finds the shared helper instead of re-deriving References logic a third time

    What we learned: A commit can be textually and functionally correct, scoped to exactly one file, and STILL mistag its room — because the room-ownership signal isn't 'which directory did I touch', it's 'what kind of problem did I actually solve'. cook-round.mjs lives nowhere in any room's owned-surface glob, so any fix to it defaults to whichever room's terminal happened to be open, and when the fix is protocol-shaped (email threading) rather than content-shaped (blog prose), the reef correctly flags the mismatch even though the file and message both look like 'voice' work.

    The missing guard (the deliverable): A pre-commit check that cross-references the room implied by Originating-Terminal against .workflow/rooms/*.html owned-surface globs for every changed file, and WARNS (not blocks) when a changed file matches no room's glob at all — surfacing 'unowned file, defaulting to terminal-of-origin' the moment it happens rather than letting the drift receipt discover it after the fact.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. litIntent=141/walksIntent=141 and litReality=133/walksReality=133 (near-100% lit-to-walked ratio both sides) indicates the walk is saturated, not shallow — it's lighting almost every cell it visits, which for a 2-file, 12-line diff (per diffstat) is disproportionate spread. Concrete adjustment: for commits under a diffstat threshold (e.g. <20 lines changed), cap walk plies at 1 instead of the current 2, so a small, single-concept diff doesn't recursively bleed across 6 Operations/Tactics sub-coordinates before intent has a chance to anchor. Experiments: Re-run this exact commit (17ac3cc75) with --max-depth 1 instead of the current 2-ply walk and diff offPct — if it drops materially, the ply-2 default is the mechanism, not the diff content · Add scripts/blog/cook-round.mjs to a room's owned-surface glob (per checklist) and re-run the receipt on the NEXT cook-round.mjs commit to measure whether prewalkOverlap rises now that the file has an intent-lane anchor · Ablation: strip '--references'/'--preheader'/'threading' tokens from the diff text fed to the reef seed and recompute prewalkOverlap — if it jumps materially, those three tokens are the concrete drag terms and belong in a room-glob comment, not just this diagnosis

    Measured outcome (deterministic, not self-graded): room mean offPct 13→12 over 6 commits since · dictionary ablation saved 766B

    A correct, well-guarded fix that nonetheless lands 34% out-of-lane raises this room's loss-ratio on paper without any real defect — the premium signal needs the owned-surface glob fix (mistagged, not misbehaved) before the next few voice commits, or the lane's breach rate keeps climbing on false-positive misattribution rather than actual drift.

    Open the signed attestation →

56. 2026-08-27 · 🔨 builder · offPct 34

    chore(receipts): commit the accumulated telemetry, panels and lens signals

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct —→23 over 6 commits since

    Open the signed attestation →

57. 2026-08-27 · 📐 architect · offPct 48

    feat(deck): PDF a keyboard-stepped React deck route — the old tool silently rendered frame 1 only

    

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 26→35 over 3 commits since

    Open the signed attestation →

58. 2026-08-27 · 🎤 voice · offPct 22 · structural

    content(blog): r3 revision — first FULL read (55/55), and the course that refuses to close scored highest

    Why it left the lane: The commit message's intent lane is entirely a grading/process narrative — per-course scores (A:70 B:68 C:82 D:64 E:85 F:55 G:62), paragraph-level ghost-read quotes ('P35', 'this honest number vs the mirror one'), and process forensics about the r2/r3 tooling contention. The diff itself, though, is not process prose — it inserts new substantive domain content: a restated legal-forum paragraph ('nobody accepts the employee's own summary as the record', 'timesheets, approvals, systems of account', 'power of attorney', 'a contractor has a spec and an inspector who does not work for them') and expanded theory citations ('Bennett 1973', 'the data processing inequality', 'record conservation... a state cannot change without something being either thrown away or kept', 'Rice 1953'). None of that legal/insurance/theory vocabulary is named anywhere in the message. The intent lane the receipt built from the message never lit those terms, so when the walk traced the diff's actual content it landed 22% of its mass in cells the message-derived lane never reached — walksIntent=143 vs walksReality=130 with prewalkOverlap only 8, i.e. the reef seed itself barely overlapped before the walk even ran.

    The semantic pull: Concrete pull terms: 'timesheets, approvals, systems of account kept by someone with no stake in the answer', 'power of attorney', 'a contractor has a spec and an inspector who does not work for them', 'the only forum that has to decide these questions with money attached', 'Bennett 1973', 'the data processing inequality', 'Rules out the obvious escape' / 'Rules out the product everybody else is selling' — a legal-doctrine and citation-heavy cluster the grading-score message never anchors.

    Intent–reality gap: The message claims this commit is about GRADING OUTCOMES (which courses scored high/low and why, per the r3 ghost-read pass); the receipt measured the diff as substantial NEW DOCTRINAL CONTENT (a legal-forum paragraph, restated physics/CS citations) that the grading narrative never names or fences.

    • Add a 'Content-Domains:' trailer line to r3-style ghost-read revision commits (alongside the existing Persona-Intent trailer) naming any new proper-noun/citation/legal clusters introduced by the diff, not just the grading deltas.
    • Update the ghost-read revision commit template (wherever r1/r2/r3 commit messages are composed) to require one domain-tag per course-fix bullet when the fix adds new source material rather than just clarifying existing prose.
    • Audit the other 2 courses fixed this round (D, F) for the same gap — D's fix ('record conservation... glossed in place') and F's fix (citation-framing rewrite) both add theory vocabulary the message summarizes but does not tag as new domain mass.
    • Backfill a one-line domain tag on the last few r1/r2/r3 commits in this post's history so the reef's intent corpus for this file catches up retroactively.

    What we learned: This room's out-of-lane events are not (only) content sprawling outside its owned globs — the file is correctly owned blog content. The class revealed here is a COMMIT-MESSAGE GENRE mismatch: the ghost-read grading convention narrates SCORES, and scores are a meta-layer one level removed from the prose being scored, so a message can be a perfectly accurate account of the grading pass while still failing to name the actual domain vocabulary the diff introduces. Prior interventions treated offPct as a room-ownership or content-sprawl problem; this event shows it can also be a report-vs-content genre gap within a single, correctly-owned file.

    The missing guard (the deliverable): A pre-commit check (scripts/voice-commit-domain-check.sh, firing on src/content/blog/*.mdx and books/tesseract/chapters/*.md diffs authored by the ghost-read/r-round workflow) that extracts capitalized multi-word terms and citation patterns (e.g. 'NAME 19NN', 'the X inequality', quoted legal-doctrine phrases) newly added in the diff hunks, and warns if none of them appear anywhere in the commit message body — forcing every grading-narrative commit to also name what content it actually planted.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized but asymmetric: litReality (130) trails litIntent (143) by about 9%, suggesting the reality-side walk saturates slightly shallower than the intent-side walk on this docsOnly commit. Adjustment: for docsOnly commits, run both sides at the same --max-depth explicitly (rather than letting reality's smaller diff naturally terminate earlier) so the comparison isn't penalizing reality for legitimately having less text to walk. Experiments: Re-run reef ingest for sha 43c0fd335 with the commit message ONLY vs message+diff-quoted-citations (Bennett 1973, Rice 1953, data processing inequality) and compare prewalkOverlap before/after — tests whether citation-inclusion in intent corpus closes the seed gap. · Ablation: force walksReality to the same max-depth as walksIntent for this sha and re-measure offPct — isolates how much of the 22% is walk-asymmetry vs seed-mismatch (should confirm the 65/35 split above or correct it). · Once the Content-Domains trailer guard ships, re-measure offPct on the NEXT r-round commit to this same post and compare against this baseline (22%) as the direct before/after test of the fix.

    Measured outcome (deterministic, not self-graded): room mean offPct 13→5 over 13 commits since · dictionary ablation saved 779B

    A structural (message-genre) miss rather than a content-sprawl miss reads as a lower-severity loss for this lane's premium — it's fixable by tightening the reporting convention, not by re-scoping the room's owned surface — and premium is restored once the Content-Domains trailer + guard ship and the next r-round commit on this post measures a lower offPct against this 22% baseline.

    Open the signed attestation →

59. 2026-08-27 · 🎩 Kitty Operator · offPct 38 · structural

    fix(deck): rebuild grip-deck.pdf from the v1.1 print HTML — the download was still v1.0

    Why it left the lane: The diff is a single binary asset swap (public/decks/grip-deck.pdf, 569519→571380 bytes, 0 insertions/0 deletions text-wise) with zero source files touched — no .html, no .ts/.tsx, no route, no script. The commit message's own vocabulary ('grip rail deck', 'headless', 'print HTML', 'route', 'download button') names UI/routing/build-pipeline concepts that live at C1.Operations.Grid, C2.Operations.Loop and C3.Operations.Flow (the deck-serving surface, the render loop, the asset-refresh flow), and B1/B2 Tactics.Speed/Deal pick up 'rebuild', 'regenerated', 'verified 10 pages' as tactical-execution language — but the walk has nothing IN the intent lane (B3.Tactics.Signal, the actual coordinate this room owns) to anchor against, because the diff itself carries no code identifiers. With litIntent=124 vs litReality=103 and prewalkOverlap=1, the reef seeded almost entirely off the commit-message prose rather than any changed-file content, so the walk fanned out across every coordinate whose keywords resemble 'deck/build/route/print/download' — 31 of them, spread near-uniformly (B3.Tactics.Signal ×11 is the single largest pull, everything else is 3-6 hits each, i.e. no real concentration, just message-noise diffusion).

    The semantic pull: Terms 'rebuild', 'regenerated', 'the route showed the new one', 'download button', 'headless' in the commit body — none of which appear in the diff itself (the diff is opaque PDF bytes) — dragged the walk toward Operations.Grid/Loop/Flow build-pipeline coordinates instead of staying inside B3.Tactics.Signal (git-discipline/asset-provenance, this room's actual lane).

    Intent–reality gap: The commit message claims a scoped, verified asset-rebuild (a discipline/provenance action: check the leak history, regenerate, verify byte/page count) — that IS a B3.Tactics.Signal action. But the receipt has nothing in the diff to place against that claim; a binary PDF has no semantic content the walk can read, so litReality (103) gets seeded almost entirely from the prose describing OTHER subsystems (route, deck rail, print HTML, headless build) rather than from the artifact itself.

    • Add `Asset-Scope:` trailer convention to CLAUDE.md COMMIT MESSAGE STRUCTURE section
    • Update scripts/pmu/pipeline.mjs ingest to detect binary-only diffstat (`Bin \d+ -> \d+ bytes`, zero text insertions/deletions) and cap prose-seed length
    • Add a docs/decks/ or public/decks/ note (or a comment atop grip-deck.pdf's generating script) naming the canonical B3.Tactics.Signal ownership of deck-artifact provenance commits, distinct from the deck-BUILD pipeline itself which is C1/C2
    • Add the binary-only-commit case to tests/pmu-simulator (a fixture commit with a Bin-only diffstat and a multi-subsystem-mentioning message) asserting seed truncation kicks in
    • Cross-link this lesson into the deck skill's own doc (`deck` skill invoked for docs/decks/ builds) so future deck-only asset commits use `Asset-Scope:` from the start

    What we learned: A binary-asset-only commit (Bin diffstat, 0 text insertions/deletions) is a receipt-ingest blind spot distinct from the docs-only case already flagged (docsOnly=true here) — the ingest has NO changed-file content to ground against, so it silently full-walks the ENTIRE commit-message prose including sentences that narrate other subsystems for context/justification (here: the route, the print HTML source, and a prior leak incident), and every subsystem named for narrative color becomes a false out-of-lane pull. This is a new structural gap not covered by the room's prior three lessons, which were facts-only and didn't diagnose an ingest-seeding mechanism.

    The missing guard (the deliverable): tests/pmu-simulator/binary-only-commit-seed-cap.test.mjs — assert that for a commit whose diffstat contains only `Bin` lines (no text hunks), the intent-corpus seed derived by the ingest step is bounded to ≤1 sentence + any Asset-Scope trailer, and that a synthetic multi-subsystem commit message (mentioning 3+ unrelated coordinate-keyword clusters) does NOT produce litReality spread across more than N (e.g. 10) distinct out-of-lane coordinates.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Right-sized in depth (2 plies, 0.98% fill, well under the 70% ceiling) but starved on INPUT quality — the walk is doing its job correctly on bad seed data. Adjustment: don't add plies; fix upstream seeding (the code_rewrite above) so the walk starts from a tighter, more accurate seed rather than compensating with fewer plies or truncated walks. Experiments: Recompress the last 20 binary-only commits (grip-deck.pdf, other PDF/PNG rebuilds) with the seed-cap applied vs not applied, and measure whether mean offPct for that commit class drops below the 15 threshold · Add a synthetic fixture: a binary-only commit with a message mentioning 4 unrelated subsystems, run it through the current ingest, confirm litReality spread >20 coordinates reproduces this incident before shipping the cap, then confirm it drops post-fix · Track offPct specifically for `Asset-Scope:`-tagged commits vs untagged binary commits over the next 20 occurrences to validate the trailer actually narrows placement

    Measured outcome (deterministic, not self-graded): room mean offPct 8→27 over 7 commits since · dictionary ablation saved 1294B

    An asset-only commit whose receipt shows 100% off-lane on a room with 2 prior out-of-lane events in 14 days reads as a repeat-offender pattern and would raise this lane's loss-ratio on a naive read; the restoring fact is that the underlying WORK was correct (verified page count, byte size, leak-history check) and the drift is a measurement artifact of prose-only seeding on binary diffs — the premium is restored by shipping the seed-cap guard, not by re-litigating the commit.

    Open the signed attestation →

60. 2026-08-27 · 📐 architect · offPct 18

    bookclub(one-door): the menu is deleted, iamfim.com is the pipeline — and a 200 was never a route

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 22→16 over 4 commits since

    Open the signed attestation →

61. 2026-08-27 · 📐 architect · offPct 25

    club(skill): the lock is a LEDGER of confirmed decisions — and I relocked it from the wrong passage

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 25→17 over 4 commits since

    Open the signed attestation →

62. 2026-08-27 · 🎤 voice · offPct 16

    content(blog): r2 revision — the inversion closed, and the falsifier I wrote was rhetoric

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 15→22 over 6 commits since

    Open the signed attestation →

63. 2026-08-27 · 🔨 builder · offPct 29

    chore(public): 247MB of the same bytes were tracked twice

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct —→23 over 6 commits since

    Open the signed attestation →

64. 2026-08-27 · 🎤 voice · offPct 27

    fix(cook): three rounds of one cook are one thread — the subject carried the four things that vary

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 14→12 over 6 commits since

    Open the signed attestation →

65. 2026-08-27 · 📐 architect · offPct 21 · structural

    fix(chat): node can no longer raise the Safari consent modal — gate on a non-prompting TCC probe

    Why it left the lane: Every substantive line in the diff is macOS TCC/AppleEvents plumbing — `AEDeterminePermissionToAutomateTarget`, `safari_may_automate` (8 call sites in browser-lease.sh), `mayAutomate()` (4 sites in browser-lease.mjs), `askUserIfNeeded=false`, `swiftc -O -o`, `osascript` — none of which are docs/strategy prose. The gzip-NCD reef seed for 📐 architect is built from `docs/strategy/`, `docs/03-product/`, `.workflow/rooms/*.html`, `.workflow/cognitive-dashboard.html`, `CLAUDE.md`; the reality corpus for this commit is 100% `scripts/chat/*.{swift,mjs,sh}` + `tests/chat/*.mjs`. There is no lexical overlap for the walk to anchor on inside architect's lane, so all 136 lit reality cells necessarily fall in adjacent operational/security coordinates outside the declared 21%-over lane.

    The semantic pull: Terms 'TCC probe', 'consent gate', 'AEDeterminePermissionToAutomateTarget', 'watchdog', 'launchd-spawned', 'unattended caller' pull toward operations/infra coordinates (the kind scripts/chat/*.sh and tests own), not toward the strategy/product-doc coordinates architect's owned-surface glob seeds the reef with.

    Intent–reality gap: The commit message and trailers claim this is architect-room work (Originating-Terminal: 📐 architect), but the receipt measured zero mass in any file architect's owned-surface glob covers — the entire diff is `scripts/chat/` + `tests/chat/`, which CLAUDE.md's own commit-trailer rule already predicts can diverge from Originating-Terminal ('a commit's originating terminal and its file-relevance to rooms can diverge — that's fine and worth recording... `thetacog_overview` attributes by glob ownership'). The receipt's lane check evidently still keys offPct off Originating-Terminal alone rather than the glob-ownership attribution CLAUDE.md says is canonical.

    • Confirm `.workflow/rooms/builder.html` (or whichever room's owned-surface glob) explicitly lists `scripts/chat/**` and `tests/chat/**` — grep for it and add if missing.
    • Add an explicit negative note to `.workflow/rooms/architect.html`'s owned-surface block: architect owns strategy/product docs only, never `scripts/**` or `tests/**`.
    • Locate the script that derives the receipt's evaluated `room` field and patch it to prefer glob-match over Originating-Terminal (see code_rewrite).
    • Add the CLAUDE.md sentence from spec_rewrite to the COMMIT MESSAGE STRUCTURE section.
    • Backfill: re-score the room's other 51 out-of-lane events from the last 14 days against glob-match instead of Originating-Terminal to see how many were mistagged rather than truly drifted.
    • Add `tests/pmu-simulator/lane-uses-glob-not-terminal.test.mjs` (see missing_guard).

    What we learned: CLAUDE.md already specifies the correct attribution mechanism (glob-ownership, same as `thetacog_overview`) for exactly this divergence, but the drift-receipt pipeline apparently doesn't consume it — it scores offPct against Originating-Terminal's reef even when the commit's own Relevant-Rooms trailer names the content-correct room. The gap isn't a missing rule, it's an unwired one: a documented design that the receipt pipeline never implements is functionally the same as no rule at all.

    The missing guard (the deliverable): tests/pmu-simulator/lane-uses-glob-not-terminal.test.mjs — construct a synthetic commit whose Originating-Terminal names room A but whose changed files match room B's owned-surface glob with zero overlap on A's; assert the receipt's evaluated room/lane is B, not A. Fails today against this exact shape of commit.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Over-saturated, not shallow: 100% of both walksIntent and walksReality are lit AND walkTruncated=true, meaning the walk hit its depth/time budget while still fully lit rather than settling into a stable partial-overlap region. Adjustment: raise --max-depth for chat/infra-code commits specifically (or extend the time budget) and re-run to see whether offPct converges to a lower, non-truncated steady state instead of the truncation-inflated 21%. Experiments: Re-run this commit's receipt with the intent-lane recomputed via glob-match (builder instead of architect) and measure whether offPct drops below the 15 threshold — direct test of the structural diagnosis. · Ablation-test seeding 'safari_may_automate' and 'mayAutomate()' (each independently repeated 8x/4x in this diff) into the gzip baseline dictionary and recompress the last 20 chat-room commits with/without them; measure prewalkOverlap delta. · Re-run this walk with --max-depth+1 (or 2x current budget) on this sha and check whether litReality/walksReality drops below 100% (i.e. whether the current truncation is itself inflating the measured spread).

    Measured outcome (deterministic, not self-graded): room mean offPct 19→31 over 6 commits since · dictionary ablation saved 742B

    A mistagged-lane event inflates architect's apparent loss ratio on work it never actually did, which either overprices the room's premium or (worse) trains the underwriting signal to distrust a room that is in fact clean — restored by re-scoring against glob-ownership rather than Originating-Terminal, which is the cheap, mechanical fix here.

    Open the signed attestation →

66. 2026-08-27 · 📐 architect · offPct 20 · structural

    bookclub(pipeline): the box stops moving — it is declared, not spliced wherever </body> lands

    

    Why it left the lane: No region-coordinate narrative is cached for 8760d8b4b, so the positional signal has to be read off the file-path/room-ownership mismatch instead. Every touched file — scripts/bookclub/render-email.mjs, scripts/lib/callback-box.mjs (new CALLBACK_MARKER export), tests/comms/callback-box-placement.test.mjs — sits entirely outside 📐architect's declared owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). The mass that lit reality cells (litReality=119) is bookclub-email-pipeline vocabulary — CALLBACK_MARKER, signalHtml, proofBlock, linkClaims, RFC 8058 mailto — none of which has a lexical home in a strategy-doc reef, so it necessarily lit orthogonal blocks.

    The semantic pull: The commit's own trailers pull in two different directions at once: 'Originating-Terminal: 📐 architect' (physical terminal at commit time) vs 'Relevant-Rooms: 🎤 voice, 📐 architect, 🎩 operator' (the file-relevance ranking, architect listed but not first). The reef lane is being scored against the physical-terminal trailer while the commit's own Story line ('Every defect in this commit is the same shape — a position that was inferred instead of declared') is itself about render-email.mjs template plumbing, not strategy.

    Intent–reality gap: The commit claims 📐architect as its lane via Originating-Terminal, and CLAUDE.md's own COMMIT MESSAGE STRUCTURE section already states this is expected to diverge from file-relevance ('a commit's originating terminal and its file-relevance to rooms can diverge — that's fine and worth recording') — but the receipt pipeline still scores offPct against Originating-Terminal's owned-surface glob, so a documented-as-fine divergence trips as a 20%-over-threshold drift event.

    • Add scripts/bookclub/**, scripts/lib/callback-box.mjs, tests/comms/** to the owned-surface glob of whichever room's .workflow/rooms/*.html actually covers bookclub/email-pipeline work (voice or builder, not architect) — pragmatic stopgap while the structural fix lands.
    • Add the 'Reef-Lane scored from Relevant-Rooms[0], never Originating-Terminal' rule to CLAUDE.md's COMMIT MESSAGE STRUCTURE section (the spec_rewrite text above).
    • Update the receipt pipeline's room-resolution step to parse and prefer Relevant-Rooms over Originating-Terminal.
    • Add a guard test asserting the receipt's room-lookup reads Relevant-Rooms when present.
    • Backfill-audit the room's 50 out-of-lane events over the last 14 days: what fraction have a Relevant-Rooms trailer that would have scored in-lane under the new rule — this validates or kills the structural theory before committing to the code_rewrite.

    What we learned: CLAUDE.md itself already documents that Originating-Terminal and file-relevance are allowed to diverge — the receipt pipeline was never told that, so it is penalizing a divergence the repo's own commit protocol pre-authorized. This is a config-vs-architecture gap invisible to file-diff review: the fix isn't in the diff at all, it's in which trailer the scorer reads.

    The missing guard (the deliverable): A check (warn-only, matching the trailer's 'convention not gate' status) that fails when NEITHER Originating-Terminal's nor any Relevant-Rooms room's owned-surface glob covers the changed files — flagging genuinely unclaimed work, while treating an Originating-Terminal/Relevant-Rooms split (this commit's actual shape) as expected, not a drift event.

    Detector probe: signal attribution — gzip reef 25% · ballistic walk 75%. walkTruncated=true while every reached cell lit (134/134, 119/119) means the walk hit its depth/time budget before it could explore further, not that it wandered unproductively — it's under-provisioned, not misused. Concrete adjustment: bump --max-depth (or the time budget) for architect-tagged commits specifically, since this room's commits (bookclub template surgery) tend to touch large multi-hundred-line diffs that need more plies to fully resolve. Experiments: Re-run this commit's ballistic walk with --max-depth increased by 2 plies and compare litIntent/litReality growth — if they saturate before the old truncation point, the current depth was already sufficient and offPct=20 is a real signal, not a truncation artifact. · Backtest the Relevant-Rooms-first scoring rule against the room's last 50 out-of-lane events: recompute offPct using each commit's Relevant-Rooms[0] room instead of Originating-Terminal, and measure what fraction drop below the 15% threshold. · Seed the voice/builder room's reef corpus with scripts/bookclub/** and tests/comms/** file content, then re-walk this exact commit under that room's lane and measure whether offPct falls under 15 (validates that the mass genuinely belongs there, not that it's diffuse everywhere).

    Measured outcome (deterministic, not self-graded): room mean offPct 19→31 over 6 commits since · dictionary ablation saved 1203B

    An out-of-lane trip caused by scoring the wrong trailer field reads as unpriced drift when it is really a measurement-boundary bug, so this lane's loss-ratio stays honest only once Relevant-Rooms is read — otherwise every physical-terminal/file-owner split in this repo's own documented convention keeps manufacturing false premium load on 📐architect.

    Open the signed attestation →

67. 2026-08-27 · 📐 architect · offPct 47 · mistagged

    fix(chat): chatgpt-scrape carried the same dead-session output path — third and last instance

    Why it left the lane: The diff's only real content is a one-line variable rename cluster: SCRATCH_DEFAULT, REPO_ROOT, CHATGPT_SCRAPE_HOME, OUT_DIR, plus the comment prose 'Same dead-session scratchpad the Gemini scraper carried' and 'may not live in another session's temp directory'. None of this is git/commit-discipline vocabulary (B3.Tactics.Signal) or grid/pipeline-topology vocabulary (C1-C3.Operations.*) — it's a single bash path fix in a chat-scraper script. The reef has no tight home for 'scratch dir path bug in a scraper', so the gzip-NCD placement smeared broadly across all 9 Tactics×Operations off-lane cells rather than concentrating — the encircled pull (B3×11, C1×6, C2×4, C3×4) reads as diffuse noise, not a real cross-lane claim.

    The semantic pull: Words like 'session', 'scratch', 'dir', 'temp', 'path', 'sweep', and the trailer's own 'Story:'/'guard' vocabulary compressed against the Signal/Grid/Loop/Flow reef cells because those cells are thin on 'file path lifecycle' content and the walk had nothing better to lock onto — a one-file six-line diff has too little mass to anchor a tight lane, so it scattered.

    Intent–reality gap: The commit message claims a narrow, correct scope (fix one hardcoded dead-session path in chatgpt-scrape.sh, third instance of a known bug class) and the diff matches that claim exactly — six lines, one file. The receipt's 47% off-lane is not evidence the WORK drifted; it's evidence the LANE (C.Operations × B3.Tactics.Signal, 'git-discipline') has no real content for 'scraper temp-path bugs', so a correctly-scoped tiny commit has nowhere tight to land and reads as scattered across nine cells instead.

    • Add a short CLAUDE.md line naming scripts/chat/*-scrape.sh + transcript-ingest as C2.Operations.Loop territory, not git-discipline, so the reef has a seed for this class
    • Grep scripts/chat/ for any remaining hardcoded /private/tmp/claude-*/ paths beyond the three already fixed (gemini, transcript-ingest, chatgpt) — this is explicitly called 'third and last instance' in the message, worth one grep to confirm
    • Add a one-line guard (shellcheck-style grep in a pre-commit or CI check) that fails if any scripts/chat/*.sh contains a literal /private/tmp/claude- string, so a fourth instance can't reappear
    • Seed the gzip reef with a small 'scraper path lifecycle' corpus fragment (a few sentences describing SCRATCH_DEFAULT/REPO_ROOT/session-scoped-tmp-dir pattern) so future commits in this class have real mass to compress against instead of diffusing

    What we learned: A correctly-scoped, correctly-committed, single-file six-line bugfix can still show 47% off-lane not because the work drifted but because the REEF has no cell shaped for its actual domain (scraper output-path hygiene) — the walk doesn't fail loud, it fails by spreading thin across whatever's nearby. This is different from the three prior architect lessons (which were facts-only, no sensemaking available): this is the first time the room's own lane taxonomy, not the commit, is shown to be the gap.

    The missing guard (the deliverable): A pre-commit or CI grep across scripts/chat/*.sh (and any future ingest scripts) that fails the build if a literal '/private/tmp/claude-' or similar session-scoped temp path is hardcoded as a default — this is the actual recurring bug (three instances), and it's mechanically checkable independent of the lane-routing issue.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Right-sized but starved: 132 reality-lit cells vs 111 intent-lit cells is a normal spread for a 2-ply walk, not saturation or shallowness. The problem is upstream of the walk — the seed (prewalkOverlap=3) gave it almost nothing to lock onto, so 'more walk depth' won't fix this; concentrate on reef content (spec_rewrite above), not walk parameters. Experiments: Run the current 190-rule reef corpus through a coverage check: for each scripts/chat/*.sh file, confirm at least one reef rule/sentence mentions its filename or function — if zero, that's a structural gap explaining low prewalkOverlap for this whole subsystem · Ablation-test adding a 2-3 sentence 'scraper output-path lifecycle' seed fragment to the reef, then re-run this same commit's receipt offline (post-drift-history.sh against sha a1180d1db) and measure whether offPct drops — if it doesn't move, the seed wasn't the bottleneck · Track prewalkOverlap distribution across the last 20 architect commits; if scripts/chat/* commits systematically show prewalkOverlap<5 while scripts/pmu/* commits show >20, that confirms an ingest-pipeline reef blind spot worth a dedicated CLAUDE.md section

    Measured outcome (deterministic, not self-graded): room mean offPct 21→32 over 4 commits since · dictionary ablation saved 913B

    A mistagged-lane event on a correctly-scoped bugfix should cost this lane nothing in loss-ratio terms once flagged as mistagged rather than defective — but until the reef seed for scripts/chat/* is fixed, every future scraper-path commit will keep tripping the same false-positive breach and inflating this room's 54-events/14-days count with noise instead of signal, which is itself a cost (alert fatigue) worth restoring via the spec_rewrite.

    Open the signed attestation →

68. 2026-08-27 · 🔨 builder · offPct 21 · mistagged

    feat(elicit): COS was not broken, it was starved — the ladder now feeds it

    

    Why it left the lane: No encircled-region narrative was cached for this sha, so the coordinate-level pull can't be named directly — but the owned-surface facts are unambiguous: every changed path (scripts/chat/gemini-pulse-tick.sh, scripts/elicit/elicit-cos-enqueue.mjs, tests/elicit/cos-enqueue.test.mjs) falls OUTSIDE 🔨 builder's declared glob list (src/app/, src/components/, src/lib/, public/, book-build scripts, hooks/pre-commit, voice-*-check.sh, decentralizedtalk/, thetacoach-crm-mcp/, vercel.json). The commit's own vocabulary — 'cos_overnight', 'elicit-run.mjs', 'elicit-dispatch.mjs', 'gemini-pulse-tick.sh', 'controller', 'the ladder' — is pipeline/ops language that has no seed anywhere in builder's reef, which is why prewalkOverlap sits at only 5 against 140 walked cells: the reef barely recognized this commit as builder's before the walk even started.

    The semantic pull: Literal strings dragging placement off-lane: 'cos_overnight', 'elicit-run.mjs / elicit-calendar.mjs / elicit-dispatch.mjs / elicit-cos-enqueue.mjs', 'gemini-pulse-tick.sh', 'bifurcate-close.mjs', 'delegation-digest.mjs --send', 'sqlite3', 'INSERT OR IGNORE INTO cos_overnight' — none of these tokens exist in builder's owned-surface vocabulary (app/component/book/CRM terms), so every one of them is a positive match for a different room's reef, not builder's.

    Intent–reality gap: The commit's Originating-Terminal trailer claims 🔨 builder, and the message is a coherent, well-scoped account of real work — but the receipt measured the work landing entirely in scripts/elicit/ and scripts/chat/, paths builder's owned-surface glob doesn't cover at all. The work is correctly built; the room claiming it is simply the wrong one.

    • Audit .workflow/rooms/*.html owned-surface globs to confirm no room currently claims scripts/elicit/**, scripts/chat/**, tests/elicit/** — verify this rather than assume
    • Add scripts/elicit/**, scripts/chat/**, tests/elicit/**, scripts/bifurcate*.mjs, scripts/comms/delegation-digest.mjs to the owned-surface glob of whichever room's reef genuinely recognizes elicitation/pipeline vocabulary (Relevant-Rooms names 🧭 navigator first — verify against navigator's declared surface before assigning)
    • Fix scripts/open-room-session.sh --print-label to compute overlap-weighted room match instead of a pure $TERM_PROGRAM → room lookup
    • Re-run this commit's receipt after the glob fix to confirm offPct drops for a mechanically-explained reason, not a re-render
    • Bucket the room's 43 prior off-lane events in 14 days by whether the diff touches scripts/elicit|chat|comms (mistag class) vs genuine builder-lane drift, to size how much of the 43 is this exact bug

    What we learned: Room attribution here is terminal-app-based (whichever $TERM_PROGRAM was physically open), not content-based — so any cross-domain backend/pipeline work authored from a terminal whose room owns none of that surface will ALWAYS misattribute, silently, every single time, regardless of how correct the code is. This is a structural gap in the attribution layer itself, distinct from and upstream of the three prior facts-only interventions for this room (which never diagnosed a mechanism at all).

    The missing guard (the deliverable): A commit-time check that computes the staged diff's path-glob overlap against the detected Originating-Terminal room's owned surface; if overlap is 0%, the trailer generation warns (or substitutes the argmax-overlap room) instead of silently asserting a room the diff shares nothing with.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/walksIntent ≈ 1.00 and litReality/walksReality ≈ 1.00 — the walk is saturated, not shallow: essentially every cell it reaches gets lit on both sides. A fully-saturated walk has little discriminating power left; concrete adjustment: when lit/walked ratio exceeds ~0.95 on both lanes, treat offPct from this pass as reef-seed-dominated (per the estimate above) and don't credit the walk's spread as the diagnostic signal. Experiments: Add an explicit pipeline/ops owned-surface glob (scripts/elicit/**, scripts/chat/**, tests/elicit/**) to whichever room actually owns it, then re-run this exact commit's receipt and measure whether offPct drops toward the 15 threshold · Change Originating-Terminal generation to overlap-weighted room selection, then re-score the room's last 43 off-lane events and measure what fraction were mistags (0% overlap with the assigned room) vs genuine drift · Across the room's last 14 days of off-lane events, correlate prewalkOverlap with offPct — if low prewalkOverlap (~5) consistently precedes high offPct, that confirms reef-seed weight should be calibrated higher than walk-spread weight generally, not just for this commit

    Measured outcome (deterministic, not self-graded): room mean offPct 12→32 over 3 commits since · dictionary ablation saved 866B

    A mistagged-not-broken drift event doesn't raise this lane's true loss ratio, but 43 unexplained off-lane events in 14 days reads as unmodeled risk to an underwriter until the attribution mechanism itself carries a guard — the fix here is what converts an opaque room-level anomaly rate back into a priced, explained one.

    Open the signed attestation →

69. 2026-08-27 · 📐 architect · offPct 17

    fix(pulse): I corrupted 20% of the receipt ledger this morning — aeperm's exit code IS its verdict

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 26→47 over 3 commits since

    Open the signed attestation →

70. 2026-08-27 · 📐 architect · offPct 33

    bookclub(pipeline): stop mixing versions — A7's tail lands in the renderer, ten doors become one

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 19→28 over 3 commits since

    Open the signed attestation →

71. 2026-08-27 · 📐 architect · offPct 34

    content(blog): the truncation route and the permission envelope — plus C26, the argument must land before course A

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 24→18 over 8 commits since

    Open the signed attestation →

72. 2026-08-27 · 🔨 builder · offPct 19 · structural

    feat(iamfim): the fork's shared checkout was prose — both doors now reach it, and a guard reads the routes instead of the page

    

    Why it left the lane: The commit's mass leaked into A,A (Strategy⊕Strategy) and A,C2 (Strategy⊕Operations.Loop) rather than staying in the Tactics/Operations lane the UI wiring itself occupies (A1,A/A2,C3, B1..C3, B,B). The pull traces to the new guard file's comment blocks, which restate CLAUDE.md-level doctrine almost verbatim rather than staying scoped to the click-path assertion: 'the same class CLAUDE.md names as "a control wired to nothing is fabrication wearing a UI"', 'a destination named in prose is not a route', and the 40-line 'THE CONSTRUCT, NOT THE TOKEN' preamble that re-derives the PROXY-IS-NOT-THE-THING argument in full. The A,C2 lane text ('substrates of strategy underwrite the long-horizon lattice for every iterative feedback loop and hypothesis test cycle') echoes almost literally the fact that this commit's payload IS a new hypothesis-test-cycle file (tests/surfaces/iamfim-fork-ends-at-checkout.test.mjs), and A,A ('the lattice we build today is what next year's underwriter inherits... the substrate that downstream pricing rides on') echoes the checkout/pricing-substrate framing repeated three times in the diff's own comments. The actual UI edits (the <a> tags, the under-rail <p>, the data-lane strings) sit cleanly in the in-lane green regions (A1,A▸A2,C3 quarterly-goal/target-position; B3,A1▸B2 signal-bandwidth) — it is specifically the essay-density of the guard's rationale that drags the placement upward into Strategy.

    The semantic pull: Doctrinal restatement inside a guard-test comment ('THE CONSTRUCT, NOT THE TOKEN', 'a control wired to nothing is fabrication wearing a UI', 'a destination named in prose is not a route', 'the page grading its own homework') carries the same vocabulary shape as the reef's Strategy-lane definitions (substrate, lattice, underwrite, hypothesis-test-cycle), so gzip-NCD grips it there even though the code it documents is tactical UI wiring.

    Intent–reality gap: The commit message claims a scoped, tactical fix — 'both doors now reach it' — entirely within 🔨 builder's UI surface (src/app/iamfim-landing, an href, a rail). The receipt measured 19% of the lit reality mass landing in Strategy⊕Strategy and Strategy⊕Operations.Loop coordinates, because the guard test that CODE DEFINITION-OF-DONE mandates ship in the same commit is itself doctrine-dense prose about the repo's own architecture principles, not tactical assertions about one page.

    • Add `tests/surfaces/**` (and ideally `tests/**`) to 🔨 builder's owned-surface glob in the rooms/bifurcation config.
    • Trim the doctrinal preamble in tests/surfaces/iamfim-fork-ends-at-checkout.test.mjs to a short CLAUDE.md pointer; keep the concrete assertion helpers.
    • Audit tests/surfaces/iamfim-landing-absolute-links.test.mjs (the sibling guard cited in this diff) for the same comment-density pattern before it produces the next out-of-lane event.
    • Re-score this commit's offPct against the widened glob to confirm it drops under 15 — if it doesn't, the pull is content-driven (doctrine density) not attribution-driven, and only the code_rewrite fixes it.
    • Consider extracting the recurring construct-vs-token rationale into one shared doc referenced by short comment from every guard that needs it, rather than re-deriving it per file.

    What we learned: Two hard rules collide structurally for this room: CODE DEFINITION-OF-DONE mandates a guard ship in the same commit as any fix, and PROXY-IS-NOT-THE-THING / SELF-IMPROVING FIRES mandate that guard explain WHY in real architectural terms — together they guarantee that a well-written builder-room guard test will contain CLAUDE.md-doctrine-shaped prose, and the room's owned-surface glob was never extended to claim the tests/ directory where that doctrine now physically lives. This is a new, nameable mechanism, distinct from the prior three facts-only interventions for this room (which had no sensemaking to escalate past).

    The missing guard (the deliverable): A test asserting that every room's owned-surface glob includes the guard-test directories implied by files it owns under src/ (i.e. if a room claims src/app/**, its glob must also claim the matching tests/surfaces/** or tests/** path) — so CODE DEFINITION-OF-DONE compliance never silently produces unowned/mis-lane content.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/walksIntent and litReality/walksReality are both ~100% (138/138, 139/139) with walkTruncated=true — the walk is fully saturating its budget on both sides, which given truncation could mean it's cut off before genuine convergence rather than a clean stop. Adjustment: re-run this commit with --max-depth+2 and check whether litIntent/litReality grow past 138/139 (real unconverged spread) or hold flat (confirms saturation was truncation-driven, not signal). Experiments: Re-run pmu-onchip --ballistic on this sha with --max-depth+2 and compare litIntent/litReality against the current 138/139 to test whether walkTruncated=true is masking real convergence. · Recompute prewalkOverlap for this sha using the trimmed guard-file variant (doctrinal preamble stripped, per code_rewrite) vs the current file, to measure how much of offPct=19 traces to comment-prose mass specifically. · A/B the reef with 🔨 builder's owned-surface glob widened to include tests/surfaces/**, re-score this commit's offPct, and record whether it crosses back under the 15 threshold.

    Measured outcome (deterministic, not self-graded): room mean offPct —→12 over 8 commits since · dictionary ablation saved 975B

    A 19%-off-lane builder commit whose actual code is tactically sound but whose mandatory guard-test prose reads as Strategy-substrate doctrine is a low-severity miss-tag, not a defect — premium on this lane holds if the glob and comment-density fixes land before the next 46-event window, since the underlying UI work (the checkout is now genuinely reachable) is exactly what the room is chartered to deliver.

    Open the signed attestation →

73. 2026-08-27 · 📐 architect · offPct 17

    fix(chat): the TCC gate was inert — allowed() returned a truthy object, and a denial-backoff cannot stop a FIRST prompt

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 26→35 over 5 commits since

    Open the signed attestation →

74. 2026-08-27 · 📐 architect · offPct 35 · mistagged

    fix(elicit): the last hop dropped every self-assigned todo and retried it forever

    Why it left the lane: The commit message itself is thick with vocabulary that CLAUDE.md's own DELEGATION NEVER GIT-BRANCHES section owns and defines — 'bifurcate.mjs', 'a room cannot delegate to itself', 'ledger', 'receipt', 'the OPAQUE shape bifurcate-close.mjs can never retire' — all terms that live in the CLAUDE.md file 📐architect owns. That vocabulary walks the intent side straight into architect's docs/strategy coordinates. The reality diff, meanwhile, is pure runtime plumbing: `dispatchAll`, `runBifurcate`, `selfAssigned`, `appendNdjson(receiptsFile, receipt)`, `delegation.fromRoom === delegation.toRoom` — none of it in `docs/strategy/`, `docs/03-product/`, `.workflow/rooms/*.html`, or `CLAUDE.md`. prewalkOverlap=6 against litIntent=112/litReality=114 is the tell: only 6 cells of reef mass were shared before the recursion even started, meaning the two seeds (architect-flavored prose vs. dispatch-script identifiers) never had much in common to begin with.

    The semantic pull: The commit narrates the delegation/bifurcate architecture in loving, rule-quoting detail while the actual patch is a two-file elicit-dispatch bugfix — the prose is what seeds the intent walk, and the prose reads like a CLAUDE.md excerpt.

    Intent–reality gap: The commit claims architect provenance via the Originating-Terminal trailer, and CLAUDE.md explicitly says that trailer may diverge from file relevance ('that's fine and worth recording') — but the receipt's lane math scores Originating-Terminal as if it WERE the lane claim, so a legitimate terminal-of-convenience commit reads as 35% out-of-lane for a room that touched none of its own files.

    • grep -rl "Originating-Terminal" scripts/ to find the trailer-reading lane resolver used by the receipt pipeline
    • Change that resolver to score against Relevant-Rooms glob ownership, not Originating-Terminal, when they diverge
    • Land the CLAUDE.md clarification sentence above in the same commit (in-lane for architect — CLAUDE.md is on its owned surface)
    • Add a note to .workflow/rooms/architect.html's owned-surface block stating it explicitly excludes scripts/** and tests/**, so an auditor reading the glob doesn't wonder why a code commit ever hit architect
    • Add the missing guard (below) in the same commit
    • Re-run the last 14 days of architect's 57 events bucketed by (Originating-Terminal==file-owner) vs (Originating-Terminal!=file-owner) to confirm this mechanism explains the bulk of them before declaring the fix sufficient

    What we learned: CLAUDE.md's own commit-trailer convention sanctions Originating-Terminal drifting from file-ownership ('that's fine and worth recording') — but nothing tells the drift receipt which trailer is authoritative for lane scoring, so any room whose terminal doubles as a general commit desk will accumulate offPct from work that is entirely legitimate. This is the first out-of-lane event traced to a documented, intentional convention rather than a mistake in the commit itself.

    The missing guard (the deliverable): A test (e.g. tests/pmu-simulator/lane-uses-relevant-rooms-not-terminal.test.mjs) that feeds the lane resolver a synthetic commit with Originating-Terminal=architect but Relevant-Rooms=builder,laboratory and a diff confined to scripts/+tests/, and asserts the computed lane is scored against the Relevant-Rooms glob union, not solely against Originating-Terminal's owned surface.

    Detector probe: signal attribution — gzip reef 75% · ballistic walk 25%. walkTruncated=true with near-identical, wide fill on both sides but almost no shared seed suggests the walk is spreading broadly and getting cut off before it can converge on overlap — closer to saturated-then-truncated than shallow. Concrete adjustment: re-run this sha's walk with --max-depth+1 (or +2) and compare offPct; if it drops meaningfully, truncation is inflating the apparent drift rather than real placement. Experiments: Re-run this commit's ballistic walk at --max-depth+2 and record offPct delta — isolates truncation-inflation from genuine placement drift · Ablation: tag CLAUDE.md's DELEGATION NEVER GIT-BRANCHES vocabulary ('bifurcate', 'ledger', 'receipt', 'mesh') to both architect AND whichever room's files call bifurcate.mjs, then re-measure prewalkOverlap on this sha and similar delegation-themed commits · Bucket all 57 of architect's last-14-day events by (Originating-Terminal==file-glob-owner) vs (!=), and compare mean offPct between the two buckets — confirms or kills the mistagged-terminal hypothesis at the room-mean level, not just this one commit

    Measured outcome (deterministic, not self-graded): room mean offPct 22→35 over 4 commits since · dictionary ablation saved 894B

    Scoring provenance as if it were a lane claim inflates 📐architect's apparent loss ratio with events that are pure noise — work the room never actually owned — so its premium should not move on this one; it's restored the moment the resolver scores Relevant-Rooms instead of Originating-Terminal, at which point only genuine cross-lane commits count against the room.

    Open the signed attestation →

75. 2026-08-27 · 📐 architect · offPct 30 · structural

    content(blog): Section A was telling the retraction story and stopping — the full argument now lands there

    Why it left the lane: The diff is pure blog-prose surgery inside src/content/blog/2026-08-27-take-the-exponent.mdx — new paragraphs built around 'npx -y thetacog-mcp@latest attest-demo', 'the retraction has a date', 'We need no sign. The rebuttal needs a specific one', and the kicker-line rewrite. None of those strings, and no path in the diff, touches 📐architect's declared owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). The receipt's intent lane was seeded from the Originating-Terminal trailer (📐architect), so every lit reality cell for this blog-voice work necessarily fell outside that lane — prewalkOverlap of only 16 against litIntent=142 shows the seed mismatch existed BEFORE the walk even ran; the walk (135 vs 142 lit, both fully saturated, walkTruncated=false) just carried that mismatch forward faithfully.

    The semantic pull: Blog/voice vocabulary ('Section A', 'the argument', 'the retraction', 'attest-demo', 'kicker' div strings) has zero overlap with architect's strategy/ops vocabulary, so the reef placed reality in the content/voice cluster while the intent lane stayed anchored to architect's owned-surface cluster — a clean bimodal split, not a blur.

    Intent–reality gap: The commit message claims editorial/content work (rewriting Section A of a blog post) and the trailers even correctly name 🎤voice and 🧭navigator as Relevant-Rooms — but the lane-gate that produced offPct=30 evidently keyed off Originating-Terminal (📐architect) alone, so a commit CLAUDE.md itself says is fine to diverge on got scored as drift.

    • Patch the lane-resolution function to prefer changed-path glob match, then Relevant-Rooms, then Originating-Terminal (see code_rewrite)
    • Add the enforcement clause to CLAUDE.md's COMMIT MESSAGE STRUCTURE section (see spec_rewrite)
    • Backfill-recompute offPct for the room's other 61 out-of-lane events in the last 14 days using the corrected resolver, to see how many were the same Originating-Terminal-vs-content-lane artifact rather than real drift
    • Add a `content(` / `club(` / `mdx-cook(` commit-type → default-room mapping table (blog/book/comms prefixes → voice/navigator, not architect) so future seeds don't need a full glob scan
    • Write the regression test named in missing_guard in the SAME commit as the resolver fix

    What we learned: Originating-Terminal and the drift-lane gate are two different instruments answering two different questions (physical seat vs. content ownership), and CLAUDE.md already says so in prose — but the receipt pipeline apparently only reads one of the two trailers when it seeds the intent lane, so every content commit typed from the architect terminal is structurally guaranteed to register as out-of-lane regardless of how good the prose is. This is a fresh failure class relative to the prior three (which were plain accumulation, not a documented-and-ignored trailer contract).

    The missing guard (the deliverable): A test that constructs a commit with Originating-Terminal=📐architect, Relevant-Rooms=🎤voice,🧭navigator, and changed files entirely inside src/content/blog/**, then asserts the lane resolver used for offPct returns the voice/navigator-owned-surface lane (or at minimum does NOT return architect's docs/strategy glob) as the intent lane — i.e. a guard that the trailer-divergence CLAUDE.md declares 'fine' is actually treated as fine by the drift gate, not silently overridden.

    Detector probe: signal attribution — gzip reef 75% · ballistic walk 25%. Right-sized, not shallow or saturated-clipped — both lit counts equal their walk counts exactly and walkTruncated=false, so the walk explored everything it was given without hitting a depth ceiling. No depth adjustment needed; the fix is upstream of the walk, at the seed. As a control, re-running this sha with --max-depth+1 should leave litReality/litIntent essentially unchanged if this diagnosis is right. Experiments: Re-seed this sha's intent lane using Relevant-Rooms (🎤voice,🧭navigator) instead of Originating-Terminal and recompute offPct — expect it to drop under the 15 threshold, confirming the seed-not-walk diagnosis · Recompute offPct for all 62 of 📐architect's last-14-day out-of-lane events under the corrected resolver and measure what fraction were Originating-Terminal-vs-content-lane artifacts vs. genuine drift · Build the commit-type→default-room table from historical glob-overlap data and re-run the ablation to confirm room mean offPct trend improves without loosening real drift detection

    Measured outcome (deterministic, not self-graded): room mean offPct 24→17 over 4 commits since · dictionary ablation saved 922B

    This event and its 61 siblings inflate 📐architect's apparent loss ratio for a risk that isn't real — documented, sanctioned Originating-Terminal/content divergence misread as drift — so premium is restored not by tightening the room's actual work but by re-pointing the lane resolver at Relevant-Rooms/path-glob-match, a fix that lives entirely inside files 📐architect already owns (.workflow/rooms/*.html, CLAUDE.md).

    Open the signed attestation →

76. 2026-08-27 · 📐 architect · offPct 28 · mistagged

    fix(chat): gemini scrapes were landing in a dead session's temp dir

    Why it left the lane: The diff's entire mass sits in scripts/chat/gemini-web-scrape.sh (SCRATCH_DEFAULT, REPO_ROOT, GEMINI_SCRAPE_HOME, OUT_DIR) plus a one-line .gitignore add — none of it touches 📐architect's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). A grep of data/rooms.json and .workflow/rooms/*.html for 'scripts/chat' returns nothing — the path is unowned by ANY room's declared glob. The commit message's prose register ('a mechanism that reports success while the thing it exists to produce goes nowhere', the Story trailer's reflective diagnosis) reads exactly like architect's analytical voice, which is almost certainly what pulled the low prewalkOverlap(4) seed toward the architect lane before the walk ever ran — the walk then correctly lit 133 reality cells in scripts/chat/ territory that the seed never bounded.

    The semantic pull: Strings 'SCRATCH_DEFAULT', 'REPO_ROOT', 'GEMINI_SCRAPE_HOME', 'data/transcripts/gemini-web' are infra/path terms with no home in architect's five owned globs; the commit message's diagnostic-essay language ('this matters more than an untidy path', 'the entire reason the Safari puppetry exists') is the only thing in the commit that resembles architect's strategy-doc register, and it is prose about the fix, not the fix itself.

    Intent–reality gap: The commit's Originating-Terminal trailer claims 📐architect, and CLAUDE.md itself says Originating-Terminal (physical room) and file-relevance (Relevant-Rooms) are allowed to diverge — but the receipt appears to score offPct against architect's owned-surface glob alone, and 100% of the diff sits outside it, so a legitimate divergence the convention explicitly permits gets counted as a 28% lane violation.

    • Add an owned-surface glob for scripts/chat/** to data/rooms.json under the room (☕network or 🧭navigator) whose vector_keywords match chat/transcript tooling — it currently matches none
    • Change the lane/offPct scorer to union Relevant-Rooms' globs, not just Originating-Terminal's, per the spec_rewrite above
    • Add a regression fixture: a commit with Originating-Terminal=architect but Relevant-Rooms listing a room that owns the touched files — assert it does NOT trip the lane threshold
    • Backfill-check the room's other 51 out-of-lane events in the last 14 days for the same Originating-Terminal-vs-Relevant-Rooms mismatch pattern before assuming they're all real drift
    • Audit data/rooms.json for other zero-owner path prefixes under scripts/ (chat/, ops/, pmu/) the same way this grep found scripts/chat/ empty

    What we learned: The commit-trailer convention explicitly permits Originating-Terminal and Relevant-Rooms to diverge ('that's fine and worth recording' — CLAUDE.md, COMMIT MESSAGE STRUCTURE), but nothing in the receipt pipeline was found to read Relevant-Rooms at all for lane-scoring; it appears to score against the single physical-location room, which converts an explicitly-sanctioned divergence into a false drift event every time an operator fixes tooling from a room whose desk doesn't own that tooling's surface.

    The missing guard (the deliverable): A test asserting the offPct lane calculation is computed against the union of a commit's Relevant-Rooms owned-surface globs (not solely Originating-Terminal's), exercised against this exact commit (df3abae3b: Originating-Terminal=architect, files in scripts/chat/, Relevant-Rooms includes navigator+network) and asserting it does not classify as out-of-lane once scripts/chat/** is assigned an owner.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized, not shallow or saturated: litIntent(105)/litReality(133) both sit well under the 144-cell ceiling, walkTruncated=false, and the ~27% reality/intent asymmetry is modest for a genuine cross-domain commit. No depth adjustment needed on the walk itself — the fix belongs upstream, at the seed/lane-scoring stage. Experiments: Re-run prewalkOverlap for df3abae3b using only the diff's identifier tokens (SCRATCH_DEFAULT, REPO_ROOT, GEMINI_SCRAPE_HOME, OUT_DIR) with the commit-message prose excluded from the seed corpus, and measure whether overlap rises above 4 — isolates whether reflective Story-trailer prose is diluting code-term seed matching · Recompute offPct for the room's other 51 recent out-of-lane events using Relevant-Rooms-union globs instead of Originating-Terminal-only globs, and measure how many drop below the 15 threshold — quantifies what fraction of architect's 52-event/14-day count is this exact mistagging class · After assigning scripts/chat/** an owning room, re-run the receipt for df3abae3b and confirm offPct falls below 15 with reefSha/libSha otherwise unchanged

    Measured outcome (deterministic, not self-graded): room mean offPct 21→34 over 4 commits since · dictionary ablation saved 514B

    A false-positive drift event inflates this room's apparent loss ratio without any real cross-lane work having occurred, so premium should not move on this event alone — it restores only once the lane scorer unions Relevant-Rooms and scripts/chat/** gets an owning glob, at which point the room's true offPct trend (not this artifact) becomes the honest signal to price against.

    Open the signed attestation →

77. 2026-08-27 · 📐 architect · offPct 21

    fix(chat): the Gemini scraper could not be aimed — it wrote whichever thread was on screen

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 22→32 over 4 commits since

    Open the signed attestation →

78. 2026-08-27 · 📐 architect · offPct 32

    feat(gem): /gem — one command that proves the Gemini scrape chain is locked, and a guard that keeps it so

    

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 22→21 over 10 commits since

    Open the signed attestation →

79. 2026-08-27 · 📐 architect · offPct 39 · mistagged

    fix(pulse): the backoff receipt printed a literal $_STATE instead of the state it saw

    Why it left the lane: The diff touches only scripts/chat/gemini-pulse-tick.sh — a bash operational/signal-handling script (backoff timer, consent gate, _receipt() diagnostic line). None of its vocabulary ('backoff', 'consent', 'permission-backoff', 'not re-prompting') maps to 📐 architect's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, docs/01-business/strategy/, CLAUDE.md). The encircled regions pull hard into B1-B3 Tactics × C1-C3 Operations (topology/velocity/throughput/exchange language) — the tactical-operations quadrant where a signal-emitting bash fix naturally lands — while zero mass lands in Strategy. offPct=39 is the receipt correctly reporting that the reality (a one-line shell interpolation fix) never touched the lane the room claims.

    The semantic pull: '_receipt "permission-backoff"', 'consent $_STATE', 'not re-prompting; run scripts/chat/grant-safari-automation.sh' are diagnostic/signal strings — they compose toward B3.Tactics.Signal (the pixel itself, C,B3) and C1.Operations.Grid, not toward any Strategy-lane term.

    Intent–reality gap: The commit message correctly and narrowly describes what happened (a printf-escaping bug in a receipt line) — the claim is accurate. The mismatch is that this operational bash fix was committed from, and trailer-attributed to, 📐 architect, a room whose owned-surface glob contains zero operational scripts. The receipt is right; the room attribution is wrong.

    • Add scripts/chat/*.sh (or a scripts/chat/ prefix) to 🔨 builder's or 🧭 navigator's owned-surface glob in .workflow/rooms/*.html — it currently belongs to no room explicitly, so it defaults to whichever terminal happens to touch it.
    • Note in CLAUDE.md under the architect room definition: 'scripts/chat/*.sh operational fixes are NOT architect-owned even when committed from this terminal — route via bifurcate to 🔨 builder.'
    • Wrap all `_receipt "label" "..."` calls in scripts/chat/gemini-pulse-tick.sh through one helper so a shell-quoting escape bug (missing/extra backslash before $VAR) can't recur silently — grep the file for other `_receipt` call sites and check each for the same escaping class.
    • Add a shellcheck pass (flagging SC2016: variables not expanding inside single/escaped quotes) to any pre-commit or CI step that already lints scripts/chat/*.sh, so this exact bug class is caught before commit, not after an hour of silently-wrong receipts.

    What we learned: Room ownership here is glob-based and the architect glob (docs/strategy, docs/03-product, .workflow/rooms/*.html, docs/01-business/strategy, CLAUDE.md) contains zero shell-script paths — so ANY bugfix to scripts/chat/*.sh committed from the architect terminal is structurally guaranteed to register out-of-lane, independent of whether the fix is correct. Prior lessons for this room were facts-only (no sensemaking) and moved room-mean offPct up then down erratically (14→28, 17→26, 14→1) without ever naming a mechanism; this is the first pass to identify that the room's own owned-surface declaration is the actual gap, not commit discipline.

    The missing guard (the deliverable): A pre-commit check that compares the changed-file set against the CURRENT terminal's room owned-surface glob (from .workflow/rooms/<room>.html) and warns/blocks with 'no owned-surface match — route via bifurcate to <best-glob-match room>' when zero files in the diff match any glob for the originating room.

    Detector probe: signal attribution — gzip reef 30% · ballistic walk 70%. walkTruncated=true on a commit this small (1 file, 1 line changed) is a red flag — truncation on a trivial diff suggests the walk hit its ply/time budget before fully resolving the reality side, which could be inflating litReality's spread as an artifact of where it got cut off rather than genuine semantic breadth. Concrete adjustment: rerun this exact sha with --max-depth raised one ply and compare litReality; if it doesn't converge, the truncation threshold for single-file diffs needs raising independent of overall time budget. Experiments: Experiment: add scripts/chat/*.sh to a builder/navigator owned-surface glob, then recompute this sha's offPct — measurable outcome: does the verdict flip from 'out-of-lane for architect' to 'in-lane for builder' at the same 39% mass, confirming the drift is attribution-only. · Experiment: rerun pmu-onchip --ballistic on sha fe0f43404 with walkTruncated forced false (raised --max-depth) and diff litReality against the current 113 — measurable outcome: quantify how much of offPct=39 is truncation artifact vs genuine spread. · Experiment: strip 'permission-backoff' and 'consent' from the reef dictionary for one dry run and recompute offPct on this sha — measurable outcome: if offPct drops significantly, those two tokens are the dominant magnet and are candidates for reef-dictionary rebalancing rather than room-glob editing.

    Measured outcome (deterministic, not self-graded): room mean offPct 22→35 over 4 commits since · dictionary ablation saved 930B

    A mistagged-but-correct one-line fix nudges this lane's apparent loss ratio up without any real defect underneath it, and it's restored by a config edit (owned-surface glob) rather than a code fix — the premium risk here is measurement noise, not engineering risk.

    Open the signed attestation →

80. 2026-08-26 · 📐 architect · offPct 25 · structural

    comms(newsletter): estimate the bounce instead of arguing about order — 66% to 81% on the offer

    

    Why it left the lane: The diff's three touched paths — docs/comms/newsletter/drafts/layout-W-diary-receipt-paired.html, scripts/comms/bounce-estimate.mjs, tests/comms/newsletter-litmus.test.mjs — match zero entries in 📐architect's declared 'Owned file/dir surface' (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). I checked every room's owned-surface list directly: docs/comms/newsletter/ and scripts/comms/ appear in NO room's formal ownership block — ☕Network claims only docs/comms/bookclub-*, and 🎩Operator's own focus_areas log even records newsletter as '(2c, ramp-anchored) · comms / newsletter / demoted', i.e. explicitly disclaimed. prewalkOverlap=3 against litIntent=140 (≈2%) shows the reef seed for this commit barely touched the architect-anchored intent corpus before the walk even ran — the content pulled hard toward comms vocabulary ('bounce', 'hazard', 'survival', 'reader-facing noun density', 'layout-W/Q') that has no home in the strategy/product lexicon the intent lane was seeded from. No region-coordinate narrative is cached for this sha, so I cannot cite the specific [row,col] cells the reality mass landed in — the diagnosis rests on the owned-surface glob mismatch and the seed-overlap collapse, not on decoded coordinates.

    The semantic pull: Strings 'docs/comms/newsletter/drafts/', 'scripts/comms/bounce-estimate.mjs', 'newsletter-litmus.test.mjs', and body vocabulary ('reach the offer', 'ghost-reader', 'layout-Q/layout-W', 'Resend', 'hazard/survival') anchor to a comms/newsletter coordinate that architect's owned surface never claims; nothing in the diff touches docs/strategy/, docs/03-product/, or CLAUDE.md.

    Intent–reality gap: The commit message and diff describe pure comms/newsletter engineering (a deterministic bounce estimator + a re-sequenced draft) with Relevant-Rooms listing 🎤voice and 🎩operator alongside 📐architect — but the receipt scored 100% of the changed files against architect's owned-surface lane, where none of them live, because no room's declared surface actually claims docs/comms/newsletter/ or scripts/comms/.

    • Add <li><code>docs/comms/newsletter/</code>, <code>scripts/comms/</code> — deterministic comms tooling + newsletter drafts</li> to 🎤voice's (terminal-voice.html) Owned file/dir surface — Relevant-Rooms already lists voice first for this class of commit
    • Add the same entry, with a note it stays read-only, to 🎩operator's Forbidden surface (kitty-operator.html) so 'demoted' becomes an explicit disclaimer instead of a dangling focus_areas note
    • Locate the script that reads the Originating-Terminal trailer to set the receipt's intent lane and add a fallback: when the trailer's room owns none of the changed files, resolve intent lane from Relevant-Rooms glob match instead
    • Audit this room's other 37 out-of-lane events in the last 14 days for the same signature (comms/docs paths with zero architect-glob overlap) to size how much of the 14→28 and 17→26 offPct regressions in the prior two lessons trace to this same unowned-surface gap
    • Update the not-yet-implemented scripts/room-drift-architect-check.sh spec so its 'flags any staged file outside the owned surface' logic checks against the UNION of all rooms' owned surfaces first, distinguishing 'landed in another room's lane' from 'landed in no room's lane'

    What we learned: CLAUDE.md explicitly sanctions Originating-Terminal diverging from content ownership ('that's fine and worth recording'), but the drift receipt's lane math has no such carve-out — it scores every commit against the originating room's glob alone. Combined with a genuinely unowned coordinate (docs/comms/newsletter/, scripts/comms/ claimed by zero room, confirmed by direct grep of every .workflow/rooms/*.html), a policy the project designed to be flexible about WHERE work happens becomes, mechanically, a guaranteed offPct spike generator whenever that flexibility is used from architect's terminal — a different failure shape than the prior three facts-only events, which carried no diagnosed mechanism at all.

    The missing guard (the deliverable): tests/rooms/lane-resolution-prefers-content-owner.test.mjs — asserts the receipt pipeline's intent-lane resolver falls back to Relevant-Rooms glob-match when the Originating-Terminal room's owned-surface has zero overlap with the changed files, and asserts a THIRD state ('unowned — no room claims this path') is surfaced distinctly from 'architect drift' so an unclaimed coordinate like docs/comms/newsletter/ cannot silently inflate one room's offPct.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litReality (79) reaching only 56% of litIntent's (140) spread on comms-domain content (dense, short, jargon-heavy vocabulary vs. the longer strategy corpus) suggests the reality-side walk is under-saturated for this content class rather than correctly right-sized. Concrete adjustment: raise --max-depth (or the ply budget) specifically when the reef seed's domain signature looks comms/voice-shaped, so the reality walk reaches a comparable lit-cell count before offPct is computed, rather than reporting bleed that is partly an under-walked reality side. Experiments: Re-run this sha's receipt with intent lane resolved via file-glob match (voice/operator) instead of Originating-Terminal and record whether offPct drops below the 15 threshold — direct test of the spec_rewrite's effect · Partition the room's other 37 recent out-of-lane events by whether the changed files match a non-architect Relevant-Room glob; measure what fraction of the 14→28 / 17→26 prior offPct swings would resolve under corrected lane attribution · Add a prewalkOverlap floor check: flag ingestSuspect=true whenever prewalkOverlap falls under ~5% of litIntent (this event: 3/140≈2%), then compare against a known in-lane architect commit's prewalkOverlap ratio to calibrate the floor

    Measured outcome (deterministic, not self-graded): room mean offPct 10→20 over 4 commits since · dictionary ablation saved 975B

    An unowned coordinate producing recurring false-positive offPct spikes degrades this room's loss-ratio signal with noise the operator cannot act on — cost is a diluted receipt (real architect drift gets lost among attribution artifacts) rather than a true premium event, and it's restored the moment docs/comms/newsletter/ and scripts/comms/ get a claiming room.

    Open the signed attestation →

81. 2026-08-26 · 🔒 vault · offPct 52 · mistagged

    feat(elicit): the cycle remembers what it already did, or it is a treadmill

    

    Why it left the lane: The diff's entire vocabulary — 'cycle_id', 'PRE.json'/'POST.json', 'priorWork', 'initiative.title', 'deferred', 'watermark', 'window', 'prompt_spec', 'gate' — belongs to the elicitation/cycle-engine subsystem (scripts/elicit/steer-cycle.mjs, tests/elicit/steer-cycle.test.mjs). None of that vocabulary or those paths intersect vault's owned surface (docs/01-business/patents/, docs/legal/, data-room*/, docs/incidents/). The reef placed the reality mass where the words actually live — laboratory/operator territory — which the commit's own Relevant-Rooms trailer already names ('🧪 laboratory, 🎩 operator'), while the receipt's room field still reads 'vault' from Originating-Terminal.

    The semantic pull: Terms like 'cycle remembers', 'priorWork', 'PRE.json/POST.json', 'fenced off', 'deferred' cluster tightly with the laboratory/operator reef corpus (pipeline self-improvement, agent dispatch) and carry essentially zero mass toward vault's patent/legal/data-room corpus — there is no patent, legal clause, or data-room artifact anywhere in this diff.

    Intent–reality gap: The commit message and trailers claim room=vault (Originating-Terminal), but every line changed is cycle-engine memory logic with zero surface overlap to vault's owned globs — the receipt is grading vault against work vault never claimed.

    • grep scripts/ and hooks/ for the Relevant-Rooms trailer generator and print its per-room glob-match scores for this sha to confirm vault scored 0
    • add scripts/elicit/**, tests/elicit/** to .workflow/rooms/laboratory.html's owned-surface glob (and operator.html as secondary)
    • add an explicit 'unowned surface' output state to the trailer generator, distinct from defaulting to Originating-Terminal
    • re-run the receipt pipeline against this sha and the last 12 vault out-of-lane events to see how many share this exact unowned-elicit-surface mechanism
    • confirm e6c2aa4ce's earlier fix (read Relevant-Rooms not Originating-Terminal) is still wired into the same attribution step this event exposes a gap in

    What we learned: The e6c2aa4ce fix repointed the receipt to read Relevant-Rooms instead of Originating-Terminal, but that only helps when Relevant-Rooms itself is computed from pure glob overlap. Here scripts/elicit/** has NO owning room glob at all, so the generator's fallback (whatever it is) still surfaces the origin terminal — vault — as if it were a legitimate top-3 match, at position one, with a real score of zero. An unowned code surface is indistinguishable, downstream, from a correctly-attributed one, and will keep re-triggering this exact class for every future elicit/-touching commit regardless of which terminal happens to be open.

    The missing guard (the deliverable): A test that fails if any room appears in a commit's Relevant-Rooms trailer while zero of its changed files match that room's declared owned-surface glob — this catches both the origin-terminal-default bug and any future unowned-surface gap before it ships a misattributed receipt.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. walkTruncated=true with litReality(143) > litIntent(123) suggests the walk is being cut off by the time/depth budget while still lighting substantial extra reality mass — right-sized-to-shallow. Concrete adjustment: re-run this sha with --max-depth increased by 2 and compare offPct; if it drops materially, the budget itself is contributing to the reading, not just the misattribution. Experiments: Re-run pmu-onchip --ballistic --start <this sha's coord> with --max-depth+2 and diff offPct against the current 52 to isolate truncation's share · Add scripts/elicit/**, tests/elicit/** to laboratory.html's owned-surface glob, then recompute prewalkOverlap for this sha — expect a measurable rise if the ownership gap is the driver · Instrument the Relevant-Rooms generator to log its raw per-room glob-match counts for this commit and confirm vault's logged score is exactly 0, isolating the fallback bug from a scoring bug

    Measured outcome (deterministic, not self-graded): room mean offPct 17→16 over 13 commits since · dictionary ablation saved 761B

    An unowned code surface silently misattributing to whatever terminal was open inflates vault's apparent loss-ratio on work it never touched, diluting the signal underwriters would actually price off vault's real patent/legal drift — restored only once scripts/elicit/** has a declared owner and the attribution step stops defaulting to Originating-Terminal.

    Open the signed attestation →

82. 2026-08-26 · 🎤 voice · offPct 19

    feat(iamfim): report the SHAPE of a real-Chrome session, not that a page was requested

    

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 28→17 over 3 commits since

    Open the signed attestation →

83. 2026-08-26 · ☕ network · offPct 29

    rules(outreach): answer the peer directly — never sidewind to a junior channel

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct —→3 over 4 commits since

    Open the signed attestation →

84. 2026-08-26 · 🧪 laboratory · offPct 28 · mistagged

    feat(chat): a read door for the message spine that cannot report a cutoff as silence

    

    Why it left the lane: The reality mass lands almost entirely in two new files — scripts/chat/thread.mjs and tests/chat/thread-door-cutoff.test.mjs — plus one CLAUDE.md paragraph about 'THE MESSAGE SPINE'. None of these paths overlap 🧪 laboratory's declared owned surface (swarm-comms/, .thetacog/orchestrator/, thetacoach-crm-mcp/, punch-list*, the .claude/*.json configs). The vocabulary doing the pulling — 'thread', 'cutoff', 'iMessage', 'WhatsApp', 'syncs', 'messages.sqlite' — is chat/messaging domain content, which is the ☕ network room's declared territory (Messages persona), not laboratory's. The walk lights up cleanly (litReality=140, litIntent=127) but it's lighting up a lane laboratory never registered.

    The semantic pull: 'scripts/chat/thread.mjs', 'data/messages/messages.sqlite', 'iMessage', 'syncs table', 'thread-door-cutoff' are the terms that pull reality into the network/messages semantic block, while laboratory's own reef anchors (mcp health, punch-list, swarm-comms, crm-mcp) sit in an entirely different block — there is no shared axis for the walk to land in-lane on.

    Intent–reality gap: The commit's own Story trailer asserts 'the laboratory owns the chat-corpus surface, so the ask lands here' — but the room's actual owned-file-surface glob (used to compute the intent lane) grants laboratory no claim over scripts/chat/** or tests/chat/**, so the receipt measured 28% of the reality mass landing somewhere the intent lane was never fenced to cover.

    • Add `scripts/chat/**`, `tests/chat/**` to 🧪 laboratory's owned-surface glob in `.workflow/rooms/laboratory.html`
    • Run `/recontext` after the glob edit to confirm ☕ network doesn't ALSO claim scripts/chat/** (avoid a new CONTESTED surface)
    • Add 'chat', 'thread', 'cutoff', 'messages-spine' to laboratory's `vector_keywords` in `data/rooms.json` so `sync-vectors.mjs` projects chat-corpus work correctly going forward
    • Audit `scripts/chat/sync-whatsapp.sh`, `messages-db.mjs`, `followup-brief.mjs`, `sync-messages.mjs` — all pre-existing chat-corpus scripts with the same unregistered-surface gap
    • Add a coverage test asserting every top-level dir under `scripts/*/` maps to exactly one room's owned-surface glob (catches the next unowned directory before it drifts)

    What we learned: A commit's prose can assert room ownership ('the laboratory owns the chat-corpus surface') that the room's actual owned-surface glob config does not grant — the drift receipt scores against the declared config, not the commit's own claim, so an unregistered ownership assertion drifts regardless of how well-scoped the code is.

    The missing guard (the deliverable): A pre-commit or CI check (extending the /recontext skill's UNOWNED/CONTESTED logic) that fails when a changed top-level directory under scripts/ or tests/ is not covered by any room's owned-surface glob — surfacing the gap before the commit lands rather than after the receipt measures it.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Saturated, not shallow — every walked cell on both sides is lit (100% fill), which means the walk metric currently can't distinguish 'explored broadly but weakly' from 'explored narrowly but strongly.' Concrete adjustment: add a per-cell lit-strength floor so a cell only counts as lit above a threshold, rather than counting every visited cell — this would separate genuine off-lane mass from walk-saturation noise. Experiments: Recompute prewalkOverlap for this sha using the META-BULK symmetric-bulk method (prompt+cell vs candidate-rule+cell) and confirm 5 isn't itself a thin-mass matching artifact · Re-run the receipt with `--start` seeded from a hypothetical scripts/chat/** reef anchor instead of laboratory's current seed set, and diff offPct to isolate seed-choice vs walk-spread contribution · Ablate: recompute this commit's offPct with laboratory's owned-surface glob hypothetically expanded to include scripts/chat/**, and check whether it converges toward the room's recent ~10-13 baseline from the prior three lessons

    Measured outcome (deterministic, not self-graded): room mean offPct —→25 over 6 commits since · dictionary ablation saved 769B

    An unregistered-surface drift like this reads as a scope breach even though the code is sound, which inflates laboratory's apparent loss ratio for free — restoring the glob to match the commit's own ownership claim is the only fix that removes the false-positive premium hit without touching a line of working code.

    Open the signed attestation →

85. 2026-08-26 · 📐 architect · offPct 25

    rules(comms): the opener ARMS the reader — the tab was the wrong register, and the rule was mine

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 18→21 over 8 commits since

    Open the signed attestation →

86. 2026-08-26 · 🎤 voice · offPct 40

    feat(email): one thread per log type per day — dailyThread() extracts the rule two senders already hand-rolled

    

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 26→12 over 3 commits since

    Open the signed attestation →

87. 2026-08-26 · 🔨 builder · offPct 31 · mistagged

    fix(notify): the visit alert had the time all along and never printed it

    

    Why it left the lane: The commit's declared intent lane is narrow — the subject and README-style header in the diff frame this as ONE dropped field ('the visit alert had the time all along and never printed it'). But the new file src/lib/notify/visit-context.ts (220 lines) actually implements four independent concerns: (1) time rendering (clockIn/visitorClock), (2) a full geo-resolution sweep (readEdgeGeo scanning every x-vercel-ip-* header, mapsLink, preciseLocation, usStateName/countryName lookups), (3) a cross-file timezone-constant consolidation (OPERATOR_TZ replacing three separate 'America/Chicago' literals across activity/track and daily-thread.ts), and (4) an aggregate-digest correction (AGGREGATE_POINTER, reverting a 4th reimplementation of /api/cron/visitor-digest). Reality mass concentrated around the geo-resolution and cross-module timezone-consolidation work, which is genuinely a different lane than the single-bug notify fix the message claims — that's the 31% leaking outside the declared lane.

    The semantic pull: Strings that pull toward the geo/config lane rather than the notify/bugfix lane: 'readEdgeGeo', 'x-vercel-ip-*' (swept generically, not the 3 named headers), 'mapsLink', 'preciseLocation', 'usStateName'/'countryName' (imported from @/lib/geo/flag), and 'OPERATOR_TZ' exported and re-imported into src/app/api/activity/track/route.ts — a second, unrelated file. Each of those is infrastructure/config-boundary work (which headers exist, which timezone constant is canonical) layered on top of the one-line fix ('tod = clockIn(now, OPERATOR_TZ).text') the commit message actually promises. No cached encircled-region coordinates were available for this sha, so this diagnosis is grounded in the diff's domain composition rather than lattice coordinates directly — that gap should be closed (see missing_guard).

    Intent–reality gap: The message claims a single dropped-field bugfix ('not a missing capability, a dropped field'); the receipt measured a commit whose reality spans geo-header enumeration, a cross-file timezone-constant unification, an aggregate-digest correction, and a mapsLink NaN-bug fix — four concerns bundled under a one-concern subject.

    • Extract readEdgeGeo/mapsLink/preciseLocation from src/lib/notify/visit-context.ts into src/lib/geo/edge-geo.ts; re-import from visit-shape/route.ts
    • Add a doc note (docs/architecture or a header comment) drawing the src/lib/notify/* vs src/lib/geo/* boundary explicitly — ONE RULE LIVES IN ONE PLACE applied to domain ownership, not just data files
    • tests/notify/visit-context.test.mjs: add an assertion that visit-context.ts's exports are clock/render-only once split, so the geo sweep can't silently re-merge back in
    • Add a pre-commit check (or extend rules-for.mjs) that flags when a NEW file's exports import from two or more unrelated top-level src/lib/* domains (geo + notify here) without the commit subject naming both
    • Backfill this commit's Story trailer or a follow-up commit note stating the four concerns explicitly, since the shipped message already undersold what landed
    • Cache the encircled-region narrative for this sha's receipt so future interventions have literal coordinates rather than diff-domain inference

    What we learned: A commit message can be locally true in every sentence (each concern IS accurately described in the body) and still make a globally false single-concern claim in its subject line — the receipt caught scope creep the prose itself documented but the tagging didn't declare; 'the body already told the truth, the subject didn't' is a new failure shape not covered by the prior three facts-only interventions for this room.

    The missing guard (the deliverable): A pre-commit check that computes, for any NEW file touched by the commit, the set of distinct top-level src/lib/* domains its own new exports import from or are imported by outside the file; if that set has size >1 and the commit subject names only one domain noun, warn (non-blocking, per ITERATION DISCIPLINE's judgment-only enforcement) with the concerns listed — giving the room the exact multi-concern list this diagnosis had to reconstruct by hand.

    Detector probe: signal attribution — gzip reef 40% · ballistic walk 60%. Saturated, not right-sized: walkTruncated=true with 100% lit-rate on both sides means the walk hit its ply/time budget before it could show whether reality's extra 15 lit cells (140 vs 125) would have folded back toward intent given more depth, or diverged further. Concrete adjustment: bump --max-depth for commits whose diff creates a new file >150 lines (like this one) so the walk can fully resolve before truncating, rather than reporting offPct off a still-expanding frontier. Experiments: Re-run this sha's ingest with --max-depth+1 and compare offPct; if it drops meaningfully, walk truncation (not real drift) explains part of the 31% · Split commit-scope ingest per new-file when a commit creates a file >150 lines across >1 existing file, and compare per-file offPct to confirm the geo-resolution code is the actual out-of-lane contributor versus the clock-consolidation code · Ablate: seed the reef with a deeper prewalk sample size for this commit only and measure whether prewalkOverlap rises above 4 — if it stays near-zero even with more seed depth, the low overlap is a genuine intent/reality gap, not a sampling artifact

    Measured outcome (deterministic, not self-graded): room mean offPct —→19 over 6 commits since · dictionary ablation saved 563B

    An offPct of 31 against a 15 threshold, on top of 49 prior out-of-lane events in 14 days for this room, reads as a room whose commit subjects are systematically narrower than their diffs — raising this lane's loss ratio not because the code is wrong (it's well-tested, 138 new test lines) but because the declared-intent-vs-reality gap is the exact signal the receipt exists to price; restoring the lane means the domain-fan-out guard shipping and at least one subsequent commit in this room landing under threshold.

    Open the signed attestation →

88. 2026-08-26 · 🛰️ unattributed · offPct 21

    docs(deals): Kenneth today — the friends-and-family answer, and why the video is the wrong artifact

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Open the signed attestation →

89. 2026-08-26 · 📐 architect · offPct 21

    comms(newsletter): Z — the straight line, one job per beat, Grip as the pivot

    

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 13→14 over 7 commits since

    Open the signed attestation →

90. 2026-08-26 · 📐 architect · offPct 21

    comms(newsletter): C and D keep the original's grammar, and the widened copy survives typing

    

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 16→1 over 3 commits since

    Open the signed attestation →

91. 2026-08-26 · ☕ Messages Network · offPct 53

    feat(shoot): log-take.mjs — the take logger the shoot loop never had

    

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Open the signed attestation →

92. 2026-08-26 · 📐 architect · offPct 48 · structural

    ops(market): the market's actual assumptions, read off 3,055 unprompted messages — each with a dated, locatable receipt

    Why it left the lane: The commit's two new files sit at docs/ops/market-assumptions-from-the-record-2026-08-26.md and tests/ops/market-assumptions-carry-receipts.test.mjs. Neither path matches any entry in 📐 architect's declared owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md) — docs/ops/ and tests/ops/ appear nowhere in that list, so the reef seed for the reality side lands entirely outside the intent lane's territory before the walk even runs.

    The semantic pull: The content is dense with market/strategy vocabulary ('assumptions', 'market', 'insurability', 'liability', 'regulation') that reads as architect's domain, but the FILE PATH itself never touches docs/strategy/ or docs/01-business/strategy/ — the repo's own convention (the diff cites `docs/ops/the-abrasive-persuasion-matrix-2026-08-26.md` as prior art) routes delegated bf-* findings docs to docs/ops/ regardless of content topic, which is the directory the reef's seed actually walked from.

    Intent–reality gap: The commit message and trailers claim this as architect-owned work (Originating-Terminal: 📐 architect, Relevant-Rooms includes architect first) — a legitimate delegated deliverable — but the receipt measured the reality lattice landing in docs/ops/ and tests/ops/, paths absent from every glob that defines architect's lane.

    • Add `docs/ops/*.md` and `tests/ops/*.test.mjs` to 📐 architect's owned-surface glob wherever it's canonically declared (.workflow/rooms/architect.html and/or the CLAUDE.md room-surface block)
    • Grep the last 90 days of commits for `docs/ops/*.md` creations, group by Originating-Terminal trailer, and check what fraction of authoring rooms already declare docs/ops/ in their own owned surface — this commit is likely one instance of a repo-wide gap, not an architect-specific one
    • Decide (and record the decision) whether docs/ops/ should be a per-room glob addition (as above) or declared a cross-room commons directory exempt from per-room lane scoring entirely — the piecemeal glob-add is the minimal fix, the commons declaration is the structural one
    • Re-run the receipt offline for sha 4dc4a7edc after the glob change to confirm offPct drops under 15 for this exact commit (a controlled before/after, not a projected number)
    • Audit the other 8 rooms' owned-surface lists for the same docs/ops/ / tests/ops/ blind spot before assuming this fix is architect-only

    What we learned: docs/ops/ is an established repo-wide convention for delegated bf-* findings-and-guard deliverables (this diff itself cites prior art living there) — yet it appears in ZERO room's owned-surface glob. Any room that correctly executes a delegated research task following that convention is mechanically guaranteed an out-of-lane receipt, independent of whether the work itself was scoped correctly. This is a hole in the mesh's ownership map, not a lapse in this room's judgment — and it's the kind of gap the prior three facts-only interventions for this room (offPct 14→28, 17→26, 14→1) had no way to name because they lacked the diff/commit-message context this pass has.

    The missing guard (the deliverable): A test over `.workflow/rooms/*.html` that unions every room's owned-surface glob and asserts it covers every directory that has repeatedly received delegated bf-* deliverables (detectable by grepping commit trailers for `bf-\d+` alongside new files under docs/ops/ or tests/ops/) — failing loudly the first time a directory accumulates N such commits with no room claiming it, instead of silently letting every future commit there mis-score.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Right-sized, not shallow or over-saturated — full lit=walks on both sides with no truncation flag means the walk ran to completion within its time budget on both lanes. No depth adjustment needed; the improvement is at the seed stage: re-seed the reality-side anchors from the commit's own changed file PATHS (docs/ops/, tests/ops/) in addition to content-derived vocabulary seeds, so directory-level signal registers explicitly instead of only through topical/content overlap. Experiments: Pull prewalkOverlap and reality-file-path prefixes for architect's last 42 out-of-lane events; if a majority show reality paths under docs/ops/ or tests/ops/, that confirms the directory-blind-spot as the dominant driver rather than a one-off for this commit · Add docs/ops/ to architect's glob and recompute the receipt offline for sha 4dc4a7edc as a controlled before/after ablation — measure whether offPct crosses back under the 15 threshold on the same commit, same corpus · Grep 90 days of commit history for `docs/ops/*.md` creations grouped by Originating-Terminal trailer, then check what fraction of those authoring rooms declare docs/ops/ in their own glob — quantifies how many of the 9 rooms share this exact gap versus it being architect-specific

    Measured outcome (deterministic, not self-graded): room mean offPct 13→39 over 5 commits since · dictionary ablation saved 646B

    An out-of-lane placement rooted in an unowned commons directory (not a scope error) is a low-severity loss for this lane's premium — it inflates the room's 42-event-in-14-days count with false positives; premium is restored the moment docs/ops/ is declared in-lane, which also retroactively explains a share of the room's other 41 events rather than indicting its judgment on each of them.

    Open the signed attestation →

93. 2026-08-26 · 📐 architect · offPct 17 · structural

    rules(comms): THE DIARY IS THE TAB — every book-club send opens on the invoice

    Why it left the lane: 3 of the 4 changed paths — docs/comms/newsletter/drafts/layout-A2-counted-words.html, docs/comms/newsletter/templates/a2.template.mjs, tests/comms/newsletter-litmus.test.mjs — sit entirely outside architect's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). Only CLAUDE.md is in-lane. The pull terms are literal: 'THE TAB', 'a2.template.mjs', 'newsletter-litmus.test.mjs', 'THE TAB — 2026-08-26' HTML block — these are comms/voice-room vocabulary and file-path tokens that the reef seeds toward the voice/newsletter coordinate, dragging the walk's lit mass off the architect lane even though the commit's Originating-Terminal trailer says architect.

    The semantic pull: 'docs/comms/newsletter/', 'templates/a2.template.mjs', 'tests/comms/newsletter-litmus.test.mjs', and the literal invoice-block markup ('THE TAB', 'pints outstanding', '18 pints', '0 billed') anchor to the comms/newsletter coordinate the reef already recognizes from prior newsletter-litmus commits — not to architect's strategy/product coordinates.

    Intent–reality gap: The commit message frames itself as a CLAUDE.md rule graduation ('Now a HARD RULE in CLAUDE.md, guarded') — squarely architect's job — but the receipt measured the bulk of the actual edited surface living inside the newsletter template, its rendered draft, and its litmus test, which are comms/voice territory per the commit's own Relevant-Rooms trailer (🎤 voice listed first).

    • Add tests/pmu-simulator/multi-room-commit-lane.test.mjs asserting the lane scorer checks Relevant-Rooms trailer membership before flagging offPct, not just Originating-Terminal
    • Update .workflow/rooms/architect.html's owned-surface declaration to explicitly EXCLUDE docs/comms/** and tests/comms/** (currently implicit by omission, should be explicit given repeat drift: 46 out-of-lane events in 14 days)
    • Update .workflow/rooms/voice.html to explicitly claim docs/comms/newsletter/** and tests/comms/** as owned surface, since this is where the mass keeps landing
    • Cross-reference the DELEGATE-TO-SUBTASKS rule from the new THE DIARY IS THE TAB rule block so a future editor sees the bifurcate obligation inline, not three sections away

    What we learned: A commit whose Relevant-Rooms trailer HONESTLY discloses multi-room work (voice, vault, operator) still gets scored 100% against the Originating-Terminal room's owned surface — the lane scorer has no mechanism to read its own commit-message convention, so correctly-disclosed cross-room work and silently-bundled cross-room work produce an identical offPct penalty. The trailer format already solves the attribution problem in prose; the receipt pipeline doesn't consume it.

    The missing guard (the deliverable): tests/pmu-simulator/multi-room-commit-lane.test.mjs — asserts that when a commit's Relevant-Rooms trailer lists room R, and the receipt's out-of-lane lit cells fall inside R's owned surface, offPct is NOT counted against Originating-Terminal alone; only mass landing outside every trailer-listed room's owned surface counts as true drift.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized, not shallow or saturated — both sides lit to completion with no truncation. The lever to pull is upstream: widen the reef seed to include the commit's own Relevant-Rooms trailer as additional anchor terms (voice, vault, operator) before the walk starts, rather than deepening max-depth. Experiments: Re-seed the reef with Relevant-Rooms trailer tokens as extra anchors for this commit class and re-run --ballistic; measure whether prewalkOverlap rises above 9 and offPct drops below 15 · Ablate the CLAUDE.md rule-block text as an isolated reef unit (strip the template/test diff) and re-score; if a pure rule-graduation commit scores offPct<10 alone, that confirms the template+test portion is the sole driver · Run gzip-NCD between the AMNESIA LEDGER vocabulary already seeded in CLAUDE.md and the new comms/newsletter template vocabulary to quantify how much of the 9-cell prewalkOverlap comes from shared ledger terms vs net-new template terms

    Measured outcome (deterministic, not self-graded): room mean offPct 18→22 over 10 commits since · dictionary ablation saved 1227B

    A trailer-honest multi-room commit scoring as a full-penalty drift event inflates architect's apparent loss ratio without a corresponding real defect, which erodes trust in the lane premium itself; restoring it requires the lane scorer to read Relevant-Rooms before pricing the deviation, not after.

    Open the signed attestation →

94. 2026-08-26 · 📐 architect · offPct 40 · mistagged

    test(chat): the probe guard asserted a substring, not that the fingerprint moves

    Why it left the lane: The diff is 100% inside tests/chat/probe-sees-top-conversation.test.mjs — it extracts the `node -e` reducer out of scripts/chat/gemini-probe.sh and drives it with DOM fixtures (HEAD_IS_OPEN, fingerprintOf, execFileSync). None of that mass touches architect's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). The pull toward tests/chat/ and scripts/chat/ is exactly what the commit says it's doing — extracting REDUCER, fingerprintOf, execFileSync — there is no strategy-doc content anywhere in the diff to anchor the reality lane back to architect's lane at all.

    The semantic pull: Every load-bearing string in the diff — `PROBE_PATH`, `REDUCER`, `execFileSync(process.execPath`, `fingerprintOf(dom)`, `HEAD_IS_OPEN` — is chat-probe/test-harness vocabulary, and the file itself lives under tests/chat/. The Relevant-Rooms trailer already admits this: '📐 architect, 🧭 navigator, 🧪 laboratory' — navigator owns scripts/chat/ and laboratory owns tests/, architect owns neither. No region narrative is cached for this sha, so the walk's exact out-of-lane coords aren't visible here, but the file-surface evidence alone accounts for the full 40% without needing them.

    Intent–reality gap: The commit message correctly scopes itself as a narrow test-guard fix inside tests/chat/ — it never claims to touch docs/strategy or CLAUDE.md. The receipt still scores it against architect's lane because Originating-Terminal says architect, even though the commit's own Relevant-Rooms trailer names navigator and laboratory as the actual surface owners — so the 40% offPct is the pipeline measuring the commit against a lane the commit itself never claimed to occupy.

    • Update the offPct/lane-scoring stage in scripts/pmu/pipeline.mjs (or wherever the commit's intent-lane is resolved) to parse the Relevant-Rooms trailer and pick the room whose owned-surface glob covers the most changed files, falling back to Originating-Terminal only when Relevant-Rooms is absent or ties
    • Add a one-line note to .workflow/rooms/architect.html's 'Owned file/dir surface' block: tests/** and scripts/chat/** are explicitly NOT architect surface (they belong to laboratory and navigator respectively) — makes the exclusion machine-legible instead of implicit-by-omission
    • Backfill-check: of the room's 40 out-of-lane events in 14 days, count how many are 100%-test-file diffs tagged architect-as-Originating-Terminal — if it's most of them, the lane-scoring fix above is the dominant lever, not a per-commit fix
    • Add a regression fixture (a synthetic commit whose changed files are 100% under tests/laboratory-owned globs but Originating-Terminal=architect) asserting the corrected scorer resolves the lane to laboratory, not architect

    What we learned: This room's receipt pipeline scores offPct against Originating-Terminal's owned-surface glob only, even though CLAUDE.md's own commit-trailer contract explicitly sanctions Originating-Terminal and Relevant-Rooms diverging — so a commit that follows the repo's documented convention to the letter gets flagged as drift by construction. Prior architect interventions were facts-only (no sensemaking cached) and the room's mean offPct trend after them went UP twice (14→28, 17→26) before one drop (14→1); this is the first pass with enough signal to say why: the scorer's lane input, not the operator's behavior, is likely the recurring cause.

    The missing guard (the deliverable): A test asserting the drift-scoring stage resolves its intent lane from the best-file-surface-matching room among {Originating-Terminal, Relevant-Rooms}, not from Originating-Terminal unconditionally — feeding it a synthetic commit with Originating-Terminal=architect and 100% of changed files under tests/** (a laboratory-owned glob) and asserting the resolved lane is laboratory.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. litIntent/walksIntent = 139/139 and litReality/walksReality = 138/138 — both essentially 100% conversion, no unlit walks on either side. That's saturation, not a decaying cascade; the ballistic walk isn't demonstrating the ply-decay behavior the anti-rules ledger expects. Concrete adjustment: rerun this sha with a higher --max-depth and confirm the lit/walked ratio starts dropping below 100% at deeper plies — if it stays saturated even deeper, the walk region itself may be too small/dense for this commit's corpus and the aperture needs widening, not the depth. Experiments: Recompute offPct for this sha and the room's other 39 recent out-of-lane events using Relevant-Rooms-matched lane instead of Originating-Terminal-only, and measure how many drop below the 15 threshold — tests the mistagging hypothesis directly · Run a correlation pass over the last 30 days of receipts: prewalkOverlap value vs offPct value, to confirm low prewalkOverlap (<10) predicts high offPct (>30) as this row suggests — if confirmed, prewalkOverlap becomes a cheap pre-walk early-warning signal · For commits whose diff files are 100% under tests/** or scripts/**, add an early-exit heuristic that flags them mistagged-candidate before the full walk runs, then measure how much walk compute that early exit saves across the room's commit volume

    Measured outcome (deterministic, not self-graded): room mean offPct 13→8 over 6 commits since · dictionary ablation saved 831B

    A 40-events-in-14-days pattern of architect-tagged commits landing entirely outside architect's owned surface reads as a noisy, unreliable lane signal for this room's premium calculation — the fix that restores it is making the scorer trust the commit's own Relevant-Rooms trailer, which turns a chronic false-positive into a correctly-priced laboratory/navigator commit instead.

    Open the signed attestation →

95. 2026-08-26 · 📐 architect · offPct 18 · mistagged

    comms(newsletter): A2 — the diary stops confessing and starts counting; determinism agrees first

    

    Why it left the lane: Every one of the three changed paths — docs/comms/newsletter/drafts/layout-A2-counted-words.html, docs/comms/newsletter/templates/a2.template.mjs, tests/comms/newsletter-litmus.test.mjs — sits in the newsletter/comms reef territory (strings like 'template.mjs', 'newsletter-litmus', 'iamfim.com', 'Grip', 'attest-demo', the sticker/CID image markup). None of them touch the architect room's declared owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, docs/01-business/strategy/, CLAUDE.md). The commit's own trailers already say this: Originating-Terminal is 📐 architect but Relevant-Rooms lists 🎤 voice, 📐 architect, 🎩 operator — the content mass pulled toward the voice/operator comms coordinate, not architect's.

    The semantic pull: The diff is 100% newsletter-template code and marketing HTML (kick(), P(), lnk(), crest/rule/logo helpers, the A2 layout copy about determinism/Pacioli/T.J.Hooper) — a comms-pipeline artifact, not a strategy/architecture artifact. The only architect-owned-surface term anywhere near this event is the trailer word 'architect' itself, used to record physical location, not content.

    Intent–reality gap: The commit message claims and delivers a self-contained newsletter revision (A2 layout, template, litmus-window fix) — that intent is accurate. The receipt's lane, however, appears to be computed from Originating-Terminal (📐 architect) rather than from Relevant-Rooms, so a commit whose CLAUDE.md-sanctioned trailers already disclose 'this is voice/operator content, authored from the architect terminal' gets scored as an architect-lane drift instead of a correctly-routed comms commit.

    • grep hooks/post-commit and scripts/pmu/*.mjs for where a commit's drift-lane 'room' is assigned, and confirm it currently keys off Originating-Terminal rather than Relevant-Rooms
    • Add the CLAUDE.md clarifying sentence (spec_rewrite above) to the COMMIT MESSAGE STRUCTURE section
    • Add owned-surface globs docs/comms/newsletter/** and tests/comms/** to whichever room (voice or operator) the resolver should route comms commits to
    • Add a guard test with a fixture commit shaped like 6bddd7bff (Originating-Terminal=architect, Relevant-Rooms=voice/architect/operator, all files under docs/comms/) asserting the resolved lane is NOT architect-only
    • Backfill-check the room's 44 out-of-lane events in the last 14 days: partition by whether Originating-Terminal disagreed with Relevant-Rooms, to size how much of that count is this same mistagging mechanism

    What we learned: The two prior facts-only interventions made this room's offPct WORSE (14→28, 17→26) — plausibly because they tried to fix the symptom (nudge files/globs toward architect ownership) without ever checking which commit-trailer field the receipt actually keys its lane off. This event shows the field itself (Originating-Terminal) was never meant to be an ownership signal — CLAUDE.md says so in the same paragraph the pipeline apparently ignores — so no amount of glob-tuning on the content side fixes a resolver reading the wrong trailer.

    The missing guard (the deliverable): A test asserting the drift-receipt/lens room-resolver prefers `Relevant-Rooms` (matched to owned-surface globs) over `Originating-Terminal` when both are present in a commit's trailers, using a fixture commit identical in shape to 6bddd7bff.

    Detector probe: signal attribution — gzip reef 75% · ballistic walk 25%. Fully saturated within its budget (lit==walked on both lanes) but walkTruncated=true means it hit a depth/time ceiling before it could explore further — right-sized for what it covered, but truncation leaves open whether a deeper walk would have found more true overlap. Adjustment: bump --max-depth for this commit's re-ingest and confirm offPct is stable (supports mistagging) rather than converging lower (would suggest truncation was masking real overlap). Experiments: Re-run the pipeline on 6bddd7bff with an increased --max-depth / ply budget and compare offPct to the current 18 — if it holds near 18, truncation is ruled out as a driver and the mistagged-trailer theory is confirmed · Across the room's 44 recent out-of-lane events, split by whether Originating-Terminal and Relevant-Rooms agree, and compare mean offPct between the two groups as a measurable test of the mistagging hypothesis · Add docs/comms/newsletter/** as an explicit owned-surface glob for the voice or operator room's .workflow/rooms/*.html and re-seed the reef, then re-check prewalkOverlap on a similar future newsletter commit for improvement

    Measured outcome (deterministic, not self-graded): room mean offPct 17→21 over 4 commits since · dictionary ablation saved 1011B

    This event is noise, not loss — a mistagged lane, not a real architect-room boundary crossing — but left uncorrected it inflates the architect lane's apparent loss-ratio with false positives and buries the real drift signal underneath it; premium on this lane only normalizes once the resolver reads Relevant-Rooms and architect's offPct trend reflects genuine architect-content events only.

    Open the signed attestation →

96. 2026-08-26 · 🎭 performer · offPct 26

    chore(delegation): bf-4674 run_commit was a concurrent room's SHA — corrected to da1e310c4

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 8→49 over 3 commits since

    Open the signed attestation →

97. 2026-08-26 · 🎤 voice · offPct 46

    content(book): un-bury the lead — drift is the ordinary consequence of abstraction, and it opens the preface now

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 10→13 over 16 commits since

    Open the signed attestation →

98. 2026-08-26 · 🎭 performer · offPct 34

    docs(finance): the four finance answers, recomputed — and the rail is the only variable cost

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 7→49 over 3 commits since

    Open the signed attestation →

99. 2026-08-26 · 🔒 vault · offPct 19

    steer(post): cycle sc-2026-08-26-0ac3-c81 — 1/1 attempted initiatives produced a commit

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 20→26 over 4 commits since

    Open the signed attestation →

100. 2026-08-26 · 🎤 voice · offPct 30

     fix(email): route all nine repetitive log streams through dailyThread — one Gmail thread per type per day

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 27→18 over 4 commits since

     Open the signed attestation →

101. 2026-08-26 · 🔒 vault · offPct 21

     feat(elicit): bounded PRE→RUN→POST steer episode — declare it, do it, measure the gap

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 41→13 over 7 commits since

     Open the signed attestation →

102. 2026-08-26 · 🎤 voice · offPct 21

     fix(test): un-red the iamfim absolute-links guard — a fixed 2500-char window stopped reaching the line it guards

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 27→18 over 4 commits since

     Open the signed attestation →

103. 2026-08-26 · 🎤 voice · offPct 26

     comms(self): whatsapp CoS brief 2026-08-26-8 — JP asked the room the measurement question and nobody answered

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct —→17 over 18 commits since

     Open the signed attestation →

104. 2026-08-26 · 🔒 vault · offPct 33

     fix(elicit): the same false-send lie was still in the other half

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 19→20 over 8 commits since

     Open the signed attestation →

105. 2026-08-26 · 🎩 operator · offPct 26

     network(qualifier): two questions, a fence, and the seat that runs them — "we fix that" is the one line the caller may never say

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct —→16 over 3 commits since

     Open the signed attestation →

106. 2026-08-26 · ☕ network · offPct 20

     feat(comms): the overlap sheet — their LinkedIn against what we ship, ranked

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 19→21 over 4 commits since

     Open the signed attestation →

107. 2026-08-26 · 📐 architect · offPct 20 · mistagged

     comms(newsletter): X — The Second Entry, and 1494 becomes the fifth row

     

     Why it left the lane: The reality walk lit cells cluster around vocabulary the diff actually contains — "Pacioli", "double entry", "1494", "newsletter", "attest-demo", "iamfim.com", the file path docs/comms/newsletter/drafts/layout-X-second-entry.html, and tests/comms/newsletter-litmus.test.mjs — none of which fall inside 📐architect's declared owned-file-surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). No region narrative is cached for this sha, so the positional evidence is the glob mismatch itself: every changed path in the diff sits outside every architect glob, which is exactly what a 20%-outside-lane, 3-of-125-prewalkOverlap reading looks like when intent is seeded from a room lane the content never touches.

     The semantic pull: "newsletter", "Pacioli", "double entry", "attest-demo", "iamfim.com" and the path docs/comms/newsletter/drafts/layout-X-second-entry.html are comms/voice vocabulary with zero surface overlap against architect's strategy/config globs — that vocabulary is what pulled litReality (133 cells) into a region the litIntent walk (125 cells, seeded from the architect lane) barely touches, producing prewalkOverlap=3.

     Intent–reality gap: The commit message and trailers execute and describe pure newsletter/comms work (a new HTML draft plus its litmus-test pointer) but the event is filed under room 📐architect via Originating-Terminal; the receipt measured 20% of reality mass landing outside architect's declared surface because that surface has no comms/newsletter content in it at all — the claim was never that architect owns this content, only that the operator was physically sitting in that terminal.

     • Confirm docs/comms/newsletter/** and tests/comms/** are declared in 🎤voice's (or a comms-owning room's) `.workflow/rooms/voice.html` owned-surface block — add the globs if absent.
     • Add the one-sentence clarification to CLAUDE.md's COMMIT MESSAGE STRUCTURE section naming Relevant-Rooms (not Originating-Terminal) as the drift-lane source of truth.
     • Re-run a8ec03ac5's offPct computation seeded from 🎤voice's lane instead of 📐architect's and record whether it falls under the 15 threshold.
     • Re-run the same commit's ballistic walk with 2x --max-depth to rule out walkTruncated=true as a contributing cause before trusting the 20% figure as final.
     • Audit the other 38 architect off-lane events from the last 14 days for the same signature (Originating-Terminal=architect, Relevant-Rooms first entry ≠ architect) to size how much of this room's 39-event count is this exact mislabeling class versus real cross-lane work.

     What we learned: CLAUDE.md explicitly sanctions Originating-Terminal and Relevant-Rooms diverging ("that's fine and worth recording") but never states which field the drift-receipt pipeline treats as the intent lane for offPct — so a field designed to record operator location, not content ownership, is silently doing double duty as a drift-classification key. Every commit made from a terminal whose physical room differs from the content's real owner (a pattern CLAUDE.md itself permits) will register as off-lane drift against the wrong room, regardless of whether the work is correct.

     The missing guard (the deliverable): tests/pmu/room-lane-resolution.test.mjs — assert that for a commit whose changed files match zero globs of the Originating-Terminal room but match ≥1 glob of a Relevant-Rooms entry, the drift receipt's intent lane is seeded from the matching Relevant-Room, not from Originating-Terminal; fails red on this exact commit's file set today.

     Detector probe: signal attribution — gzip reef 75% · ballistic walk 25%. walkTruncated=true with litReality (133) slightly outpacing litIntent (125) suggests the walk was cut off before it could fully explore either side, which inflates uncertainty on the 20% figure rather than confirming it. Concrete adjustment: for commits touching >1 distinct file domain (here: a comms HTML draft + a test file), double --max-depth or the time budget before trusting offPct, and re-tag the receipt with a LOW-CONFIDENCE marker whenever walkTruncated=true survives the retry. Experiments: Re-run a8ec03ac5's walk with 2x --max-depth and compare offPct against today's 20 — if it converges under 15, walkTruncated was the dominant cause, not lane mismatch. · Re-seed intent from Relevant-Rooms glob-match instead of Originating-Terminal for the last 39 architect off-lane events and measure whether the room's mean offPct trend (14→28, 17→26, 14→1 across prior interventions) drops once terminal/surface conflation is removed. · Track prewalkOverlap/litIntent as a standalone leading indicator and test whether a <5% ratio predicts offPct>15 with high precision across the historical corpus — if so, gate the expensive full walk behind a cheap prewalkOverlap check.

     Measured outcome (deterministic, not self-graded): room mean offPct 12→16 over 7 commits since · dictionary ablation saved 990B

     A mistagged-lane event like this one is a false claim against architect's loss ratio rather than a real one — it inflates the room's offPct trend without reflecting bad work, and premium on this lane is only restored once the pipeline stops charging Originating-Terminal for content it was never declared to own.

     Open the signed attestation →

108. 2026-08-26 · 🎩 operator · offPct 16

     feat(ops): the event script — seal a countable GET before you walk in, or do not go

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct —→15 over 14 commits since

     Open the signed attestation →

109. 2026-08-26 · 📐 architect · offPct 49

     comms(newsletter): A1 — sticker first, Hooper closes the pattern, and the flattering sentence is banned

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 14→18 over 4 commits since

     Open the signed attestation →

110. 2026-08-26 · 🔒 vault · offPct 39 · mistagged

     fix(elicit): the backlog was counting re-declarations, and my own alarm believed it

     Why it left the lane: The diff touches only scripts/elicit/steer-cycle.mjs and tests/elicit/steer-cycle.test.mjs — a dedup fix for the priorWork() deferred-list. Vault's owned surface is docs/01-business/patents/, docs/01-business/patents/v20/, docs/01-business/patents/prov-*/, docs/legal/, data-room*/, docs/incidents/ — zero glob overlap with either changed file. The commit still carries Originating-Terminal: 🔒 vault and, critically, Relevant-Rooms: 🔒 vault, 🧪 laboratory, 🎩 operator — vault listed FIRST in the field the drift pipeline is supposed to score lane-fit against, with laboratory/operator (the rooms that actually plausibly own scripts/elicit/) demoted behind it. prewalkOverlap=3 is very low, meaning the reef's own seed match between the intent lane's vocabulary (patents/legal/incidents) and the reality lane's vocabulary (backlog/cycle/deferred/dedup) found almost nothing in common before the walk even ran — the walk then had to spread (litReality 125 vs litIntent 114) into territory the seed never anchored, and 39% of that lit reality mass landed outside vault's lane because vault was never a genuine candidate lane for this content.

     The semantic pull: Nothing in the diff pulls toward patents/legal/data-room content semantically — 'deferred', 're-declared', 'ledger', 'drowning detector', 'arithmetic' are elicitation/scheduling vocabulary, not filing/claims vocabulary. The pull is not lexical, it's attributional: the Relevant-Rooms trailer generator placed 🔒 vault ahead of the two rooms (🧪 laboratory, 🎩 operator) that DO plausibly touch scripts/elicit/, which is the exact same symptom e94a8cd14 diagnosed ('Relevant-Rooms itself lists vault first despite vault's owned-surface glob matching zero changed files'). No region narrative was cached for this sha, so the only positional evidence available is this glob mismatch, not a specific out-of-lane coordinate pair.

     Intent–reality gap: The commit message makes no claim whatsoever about patents, legal, or the data room — it is entirely about a backlog-counting bug in the elicitation cycle tracker. The receipt nonetheless measured it against vault's lane because the Relevant-Rooms trailer names vault first with no genuine glob support, so the gap is not between what the work claimed and where it landed — it's between what the work claimed (nothing vault-shaped) and what the TRAILER claimed (vault-shaped) on its behalf.

     • Audit the Relevant-Rooms scorer for the exact backfill/fallback branch that includes a zero-match room (this is the second time it's been implicated — e94a8cd14 then this sha) and remove the fallback-to-Originating-Terminal path
     • Add a glob for scripts/elicit/** to whichever room's .workflow/rooms/*.html most plausibly owns backlog/cycle tooling (laboratory is the standing candidate — both prior events and this one place laboratory in Relevant-Rooms already)
     • Add a unit test asserting Relevant-Rooms never contains a room whose glob-match count against the commit's changed files is zero
     • Add a drift-pipeline test: a commit whose Relevant-Rooms (post-trim) has zero real vault-glob overlap must not be scored against vault's lane at all
     • Backfill-check the last 19 vault out-of-lane events in 14 days for the same zero-glob-match signature before assuming each is a distinct root cause
     • Cross-reference e94a8cd14's fix commit to confirm whether it touched the trailer GENERATOR or only the pipeline's field-read, since this event shows the generator side is still unfixed

     What we learned: e94a8cd14 already proved the pipeline can read the 'right' field (Relevant-Rooms) and still mistag, because the field being read can itself be wrong — but the fix that shipped for that lesson evidently patched the READ side only. This event is the same defect recurring on the SAME room 13→19 events later, which means the actual root — whatever generates Relevant-Rooms — was never touched. The durable fact: 'read the correct field' and 'the correct field is correctly populated' are two separate claims, and this repo has now shipped a fix for the first while leaving the second's regression untested, twice.

     The missing guard (the deliverable): A test (and a pipeline-level assertion) that Relevant-Rooms can never include a room with a zero-count glob match against the commit's own changed-file list — i.e. the trailer generator's output must be checkable against the same glob data the drift pipeline uses to score lane-fit, so the two can never silently disagree the way they did here.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent==walksIntent (114/114) and litReality==walksReality (125/125) — every walk step lit a cell on both sides, i.e. full saturation, yet walkTruncated=true. A fully-saturated-but-truncated walk on a mistagged lane is not informative about walk depth at all; re-run this same commit's walk against the CORRECT candidate lane (laboratory/operator) before concluding anything about depth — if it still saturates at the same ply count against a semantically closer lane, the walk is right-sized; if it saturates faster, this commit's current depth is oversized for the actual content. Experiments: Re-run this sha's ballistic walk seeded from 🧪 laboratory instead of 🔒 vault and compare offPct — if it drops toward the 0-15 in-lane band, that's direct confirmation the lane choice (not the walk) produced the 39% · Add prewalkOverlap to the room-assignment decision itself: refuse to score a commit against any candidate room whose prewalkOverlap falls in the bottom decile of that room's own historical distribution, and log it as a lane-candidate rejection rather than an out-of-lane event · Backtest the last 19 vault out-of-lane events: compute prewalkOverlap for each against vault vs. against its own Relevant-Rooms runner-up, and check what fraction would resolve to in-lane if scored against the runner-up — this sizes how much of vault's 19-event count is this same mistagging class

     Measured outcome (deterministic, not self-graded): room mean offPct —→21 over 5 commits since · dictionary ablation saved 915B

     A vault-lane commit that never touches patents/legal/data-room content and still scores 39% off-lane reads to an underwriter as vault's premium being priced against noise rather than real drift, which erodes the lane's signal value across all 19 recent events; restoring it requires the Relevant-Rooms generator to stop padding with a zero-glob-match room so vault's offPct trend reflects only commits that actually touch its owned surface.

     Open the signed attestation →

111. 2026-08-26 · 🔒 vault · offPct 51 · mistagged

     feat(ops): the at-rest close now kicks the steer cycle, detached

     

     Why it left the lane: scripts/pmu/drift-watch.sh (line 14 + 27-36) resolves the receipt's room ONLY from the commit's Originating-Terminal trailer, falling back to 'builder' — it never reads Relevant-Rooms at all. This commit's own trailers read 'Originating-Terminal: 🔒 vault' but 'Relevant-Rooms: 🎩 operator, 🔒 vault, 🧪 laboratory' — the commit correctly self-diagnosed that operator/laboratory own the content (scripts/ops/session-close-cycle.sh, scripts/elicit/steer-cycle.mjs, nohup/disown detach pattern, tests/ops/*), but drift-watch.sh used the terminal label instead, so commit-drift.mjs graded steer-cycle/session-close/detach content against vault's patent-and-legal intent spec. The mass had nowhere honest to land inside that spec, hence offPct=51.

     The semantic pull: Terms actually in the diff — 'steer-cycle.mjs', 'nohup', 'disown', 'detached', 'session-close', 'movement since the last cycle', 'aggressive cadence' — are all Operations/pipeline vocabulary with zero overlap to vault's owned surface (docs/01-business/patents/, docs/legal/, data-room*/, docs/incidents/). No region narrative was cached for this sha, so the pull is inferred from vocabulary-vs-owned-globs rather than a cached lattice coordinate, but the mechanism is confirmed directly in the receipt-attribution code, not just inferred.

     Intent–reality gap: The commit message trailers already state the correct attribution — Relevant-Rooms lists 🎩 operator first, 🔒 vault only third — but the receipt pipeline's room-resolution code ignores Relevant-Rooms entirely and keys off Originating-Terminal, so the receipt claims a vault lane the commit itself never claimed as primary.

     • scripts/pmu/drift-watch.sh: implement Relevant-Rooms-first room resolution as above (the headline code_rewrite)
     • tests/pmu-simulator/ or tests/ops/: add a fixture test asserting drift-watch.sh resolves ROOM=operator for a commit message carrying Originating-Terminal:vault + Relevant-Rooms:operator,vault,laboratory
     • CLAUDE.md COMMIT MESSAGE STRUCTURE: land the spec_rewrite clause in the same commit as the code fix, not as a separate prose-only pass
     • docs/architecture/anti-rules-ledger.md or the SELF-IMPROVING FIRES section: note that e6c2aa4ce's diagnosis went unshipped for one full cycle — this is the concrete instance the 'ship the prevention, not just diagnose' rule exists to catch
     • scripts/pmu/commit-drift.mjs (or wherever the receipt row's room field is finally written): confirm it consumes drift-watch.sh's ROOM var and doesn't independently re-derive from Originating-Terminal a second time

     What we learned: A sensemaking lesson that correctly diagnoses a bug (e6c2aa4ce: 'the receipt pipeline is reading Originating-Terminal instead of Relevant-Rooms') is not self-enforcing — drift-watch.sh line 14 still documents the broken behavior verbatim as the intended design, with no Relevant-Rooms parsing anywhere in the file. The intervention loop can diagnose the same root cause twice if the first diagnosis never produces a commit against the code it names — this is the SELF-IMPROVING FIRES rule failing on its own most recent output, and it's the first time in this room's lesson history where the fix target was named exactly, by file and line, and still didn't land.

     The missing guard (the deliverable): A test that feeds drift-watch.sh a synthetic commit message with divergent trailers (Originating-Terminal: vault / Relevant-Rooms: operator, vault, laboratory) and asserts ROOM resolves to 'operator', not 'vault' — the guard e6c2aa4ce should have shipped alongside its diagnosis.

     Detector probe: signal attribution — gzip reef 75% · ballistic walk 25%. Saturated, not shallow — both sides hit a 100% lit-to-walked ratio, and walkTruncated=true reflects a time-budget cut after full saturation, not premature stopping. No depth adjustment needed here; re-running with the correct room seed (operator, not vault) is the actual lever, not --max-depth. Experiments: Re-run commit-drift.mjs for c1cf8a5c2 with --room operator instead of --room vault and diff offPct against 51 — if it drops near the 15 threshold, that's the ablation proof the defect was attribution, not code, before merging the drift-watch.sh fix. · Grep the last 90 days of commit-drift receipt rows for cases where Originating-Terminal-room != Relevant-Rooms[0]-room, recompute offPct with Relevant-Rooms[0] for each, and measure whether the 9-room mean offPct drops — this validates the fix's blast radius before it ships broadly. · Add a --explain flag to drift-watch.sh that prints which trailer supplied ROOM (relevant-rooms | originating-terminal | default) on every run, so every future receipt row is self-auditing on precisely this failure mode.

     Measured outcome (deterministic, not self-graded): room mean offPct 15→12 over 10 commits since · dictionary ablation saved 1021B

     A mistagged-room false positive inflates 🔒 vault's apparent loss ratio on work it never claimed and never touched — restoring premium requires no code changes to the underlying work, only fixing the attribution bug so the receipt is priced against the room that actually built it.

     Open the signed attestation →

112. 2026-08-26 · ☕ network · offPct 31

     network(targeting): the seat is who, the pre-contact tell is how, the hook qualifies nobody

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 15→14 over 4 commits since

     Open the signed attestation →

113. 2026-08-26 · 🔒 vault · offPct 24 · mistagged

     fix(chat): the 'access data from other apps' popup was a diagnostic nobody needed

     Why it left the lane: The diff's actual vocabulary — fdaScope, fs.accessSync, TCC, 'Full Disk Access', chat.db, Library/Messages, EPERM/EACCES, writeDenial/readDenial/clearDenial — is macOS permission-and-process-plumbing content living entirely in scripts/chat/sync-messages.mjs and tests/chat/no-tcc-canary-probes.test.mjs. Vault's owned surface (docs/01-business/patents/, docs/legal/, data-room*/, docs/incidents/) shares zero tokens and zero path prefixes with either changed file. The mass had nowhere in vault's lane to land, so it lit cells the receipt correctly scored as out-of-lane — the pull isn't semantic drift inside real work, it's that the intent corpus seeded from vault's glossary (patent/legal terms) has near-zero gzip overlap with a TCC-permission bugfix (prewalkOverlap=4 against litIntent=133 confirms the seed barely touched the reality corpus before the walk even started).

     The semantic pull: 'fdaAdvice', 'TCC', 'Full Disk Access', 'chat.db', 'writeDenial/readDenial/clearDenial' in the diff have no counterpart anywhere in vault's owned-surface globs (patents/, legal/, data-room*/, incidents/) — there is no shared vocabulary to pull the walk back into lane, only apart.

     Intent–reality gap: The commit's trailer claims Originating-Terminal 🔒 vault and lists vault first in Relevant-Rooms, but the receipt measured 100% of the changed-file mass (scripts/chat/sync-messages.mjs, tests/chat/no-tcc-canary-probes.test.mjs) landing outside vault's owned surface — the work is a correct, well-guarded macOS-permission fix; only the room attribution is wrong.

     • Add 'scripts/chat/**' and 'tests/chat/**' to the owned-surface glob list of whichever room should actually claim OS-integration/system-ops work (🎩 operator is already the closest match in this commit's own Relevant-Rooms list) in .workflow/rooms/operator.html
     • Locate and fix the trailer-generation script (per CLAUDE.md, it scores against .workflow/rooms/*.html 'Owned file/dir surface' globs) so a zero-glob-match commit cannot silently retain Originating-Terminal in Relevant-Rooms
     • Run the existing /recontext skill against the full repo tree to enumerate every currently-UNOWNED path prefix (scripts/**, tests/** subtrees especially) before the next room-glob edit, so this is fixed structurally rather than one path at a time
     • Re-tag this commit's trailer in a follow-up commit correcting Relevant-Rooms now that the true owner is known
     • Add tests/rooms/room-glob-coverage.test.mjs asserting every file changed in the last N commits matches ≥1 room's owned-surface glob, flagging UNOWNED paths as a build warning

     What we learned: e94a8cd14 diagnosed that the trailer generator falls back to Originating-Terminal as 'relevant' with no real glob match, and it measurably helped (offPct 21→12 over 11 commits) — but this event, with vault's out-of-lane count actually climbing to 18 in 14 days, shows the deeper structural fact: the fix addressed HOW the fallback fires, not that NO room in the 9-room ownership map (.workflow/rooms/*.html) owns scripts/chat/** or tests/chat/** at all. A correctly-behaving trailer generator still has nowhere honest to route this work, because the coverage gap is in the glob definitions themselves, not just in which field gets read.

     The missing guard (the deliverable): tests/rooms/room-glob-coverage.test.mjs — walk every room's owned-surface glob against a sample of recently-changed paths (or the full tree) and fail if any path under scripts/ or tests/ matches zero rooms; this is exactly what the /recontext skill already computes on demand, so the guard should call recontext's UNOWNED output and turn it into a red test rather than a report nobody re-reads.

     Detector probe: signal attribution — gzip reef 60% · ballistic walk 40%. litIntent=133 and litReality=138 are both near the 144-cell ceiling (~92-96% lit) — the walk is close to saturated on both sides, which means it's providing little differentiating signal about WHERE specifically the mismatch concentrates. Adjustment: when litIntent/litReality both exceed ~90% of the lattice, lower --max-depth or add an early-saturation check so the walk stops burning budget once it's no longer discriminating. Experiments: Run /recontext against .workflow/rooms/*.html to produce a current UNOWNED-path report and measure how many of the last 90 days' commits landed on paths with zero glob match · Re-run the trailer generator against the last 20 vault-tagged commits with a 'require ≥1 real glob match, else UNOWNED' rule and measure whether vault's room-mean offPct trend improves versus the current 21→12→24 oscillation · Sample 10 correctly-tagged vs 10 mistagged historical commits and compare litIntent/litReality saturation levels — if saturation correlates with mistagging, that becomes a cheap pre-walk flag

     Measured outcome (deterministic, not self-graded): room mean offPct —→21 over 5 commits since · dictionary ablation saved 708B

     An 18-event/14-day out-of-lane rate for vault, on top of two prior facts-only interventions that made the trend worse (11→24, 17→27), reads as a lane whose attribution — not its underlying work — is unreliable; premium is restored by fixing the glob coverage and trailer generator once, structurally, rather than by another per-commit sensemaking patch.

     Open the signed attestation →

114. 2026-08-26 · 📐 architect · offPct 21 · structural

     comms(newsletter): the live renderer fails 13 of the 20 guards — measured, logged, and NOT sent

     

     Why it left the lane: The diff is 100% scripts/comms/newsletter-delta.sh + data/comms/newsletter-delta.ndjson — strings like 'render-email.mjs', 'layout-A2', 'guards_pass', 'bounce-estimate.mjs', 'newsletter-litmus.test.mjs' are comms-pipeline vocabulary, not strategy vocabulary. None of it touches docs/strategy/, docs/03-product/, .workflow/rooms/*.html, docs/01-business/strategy/, or CLAUDE.md — architect's actual owned surface. No region narrative is cached for this sha, so the exact [row,col] anchors the walk lit aren't visible to me; what IS visible is that 100% of the changed-file mass sits outside every glob architect owns, which is sufficient to explain a 21% offPct on its own without needing the coord list.

     The semantic pull: 'newsletter', 'render-email.mjs', 'guards_pass/fail', 'finishing_pct', 'bounce-estimate.mjs' are all comms-pipeline reef terms that would seed strongly into whichever coordinate the reef assigns scripts/comms/* — a lane this commit's Originating-Terminal (architect) doesn't own, but that Relevant-Rooms (voice, architect, operator) already flagged as plausible.

     Intent–reality gap: The commit trailers self-report multi-room relevance (Relevant-Rooms: voice, architect, operator) and the body is entirely a comms-pipeline diagnostic, yet the receipt scores the commit against a single lane (architect) whose owned surface the diff never touches — the trailer already told the pipeline this was cross-lane and the receipt scored it as a lane violation anyway.

     • Locate the receipt's room/lane-resolution code path and confirm today it reads Originating-Terminal rather than Relevant-Rooms
     • Add scripts/comms/** and data/comms/** to voice's (or operator's) owned-surface glob in .workflow/rooms/*.html if not already claimed there
     • Amend CLAUDE.md COMMIT MESSAGE STRUCTURE per spec_rewrite above
     • Add a guard test asserting lane resolution prefers Relevant-Rooms glob-match over Originating-Terminal when they diverge
     • Sample the room's other 44 out-of-lane events from the last 14 days and check what fraction share this same terminal-vs-surface mismatch shape before assuming this fix covers the whole 21%→lower trend

     What we learned: CLAUDE.md itself already distinguishes 'where the operator sat' (Originating-Terminal) from 'what surface the commit touches' (Relevant-Rooms) precisely because they diverge — but the drift receipt's lane scoring apparently still keys off the former, so the pipeline is measuring against a field the project's own convention says is not the authoritative lane claim. Prior lessons for this room were facts-only with no mechanism; this is the first time the actual attribution bug (which trailer field the scorer reads) has been named.

     The missing guard (the deliverable): A test that, given a commit whose changed files intersect zero owned-surface globs of its Originating-Terminal room but intersect the owned-surface globs of a room named in Relevant-Rooms, asserts the receipt's lane equals the Relevant-Rooms match — not Originating-Terminal.

     Detector probe: signal attribution — gzip reef 45% · ballistic walk 55%. Saturated-and-truncated, not too shallow in the usual sense: both lanes hit 100% lit-to-walked ratio while still being cut off by the budget (walkTruncated=true). Concrete adjustment: re-run this exact sha with --max-depth increased by ~2 plies and compare offPct — if it drops, the 21% partly reflects an unfinished walk rather than a stable semantic verdict. Experiments: Add 'newsletter', 'render-email', 'comms', 'bounce-estimate' to voice/operator's reef seed keywords, replay this sha, and measure whether litReality mass shifts toward voice/operator coordinates · Replay this sha with --max-depth +2 (deterministic) and record whether walkTruncated flips to false and offPct changes · Recompute the receipt using a Relevant-Rooms-derived lane (voice+operator combined glob) instead of the Originating-Terminal-only lane for this sha and diff the offPct against the shipped 21%

     Measured outcome (deterministic, not self-graded): room mean offPct 18→20 over 4 commits since · dictionary ablation saved 822B

     An architect-lane commit that is 100% comms-pipeline code reads as unmeasured cross-domain drift and inflates this lane's apparent loss ratio; premium is restored once the lane resolver scores by Relevant-Rooms surface-match instead of Originating-Terminal, so genuinely in-lane strategy work stops sharing a bucket with mistagged comms diagnostics.

     Open the signed attestation →

115. 2026-08-26 · 🔒 vault · offPct 22

     fix(elicit): a blind scraper is not a quiet operator — the gate must say which

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 19→11 over 3 commits since

     Open the signed attestation →

116. 2026-08-26 · 🔒 vault · offPct 28

     feat(elicit): the deferred pool is a graveyard unless something counts it

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 19→26 over 4 commits since

     Open the signed attestation →

117. 2026-08-26 · 📐 architect · offPct 20

     comms(newsletter): A6 — name an argument, link its post; and a 404 that had shipped four times

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 20→20 over 6 commits since

     Open the signed attestation →

118. 2026-08-26 · 📐 architect · offPct 23 · structural

     comms(newsletter): A5 — restore Grip as the pivot; a pasted draft is not a spec

     

     Why it left the lane: Every changed path in this commit — docs/comms/newsletter/drafts/layout-A5-grip-then-determinism.html, docs/comms/newsletter/templates/a5.template.mjs, tests/comms/newsletter-litmus.test.mjs — sits outside every glob in architect's declared owned-surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). The diff's own vocabulary — 'newsletter', 'a5.template.mjs', 'litmus', 'Grip.', 'determinism', 'Pacioli', 'iamfim.com', 'kick()'/'P()' helpers — is comms/voice-domain content with zero lexical overlap against architect's strategy/product corpus, so the recursive walk necessarily lights cells outside architect's lane: prewalkOverlap=4 against litIntent=135/litReality=133 means the reef seed had almost no shared ground between the claimed lane and the actual reality corpus before the walk even ran.

     The semantic pull: The pull terms are comms-pipeline nouns with no architect analog: 'newsletter', 'template', 'litmus', 'Grip', 'determinism', 'a5.template.mjs' — these are the exact words the CANONICAL COMMUNICATION PROTOCOL and paradox-voice corpus own, and Relevant-Rooms itself names 🎤 voice first, not architect. offPct=23 is the measured share of reality mass that landed in that voice-shaped territory instead of architect's strategy/product territory.

     Intent–reality gap: The commit message claims a legitimate comms(newsletter) change and its own trailers admit the terminal (📐 architect) and the content-relevant rooms (🎤 voice, 📐 architect, 🎩 operator) diverge — CLAUDE.md explicitly sanctions that divergence as 'fine and worth recording.' But the receipt measured this commit's reality against architect's owned-surface lane specifically, and no glob anywhere in architect's declared surface covers docs/comms/newsletter/ or tests/comms/ — so a commit CLAUDE.md calls fine by design reads as a 23%-over-threshold lane violation by the pipeline.

     • Add docs/comms/**, docs/comms/newsletter/**, tests/comms/** to voice's owned-surface glob in .workflow/rooms/voice.html
     • Add the Relevant-Rooms-over-Originating-Terminal scoring rule to CLAUDE.md COMMIT MESSAGE STRUCTURE
     • Audit all 9 rooms' owned-surface globs for other unclaimed live directories (docs/outreach/, docs/decks/, docs/ops/, data/messages/ are candidates)
     • Add a guard test asserting every docs/ and tests/ subdirectory with commit activity in the last 90 days is claimed by ≥1 room's owned-surface glob
     • Re-score architect's last 49 out-of-lane events against Relevant-Rooms[0] instead of Originating-Terminal and report how many drop below the 15% threshold

     What we learned: The room's elevated event count (49 out-of-lane events in 14 days) is plausibly not accumulating operator drift at all — it's a structural artifact of an entirely unclaimed directory (docs/comms/newsletter/, tests/comms/) that guarantees any comms commit routed through ANY terminal reads as out-of-lane for that terminal's room, regardless of content quality. This explains why the two prior facts-only interventions (ffaf8d2ac, ff5110f85) left offPct WORSE — restating the drift never touched the missing glob — while feeb3558c's 14→1 improvement over 14 commits suggests an earlier session found and fixed an analogous ownership gap for a different directory and it held.

     The missing guard (the deliverable): A test asserting every directory under docs/ and tests/ with recent commit activity is claimed by at least one room's 'Owned file/dir surface' listing in .workflow/rooms/*.html — failing when a live, actively-committed directory is unclaimed by all 9 rooms.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. walksIntent(135)≈litIntent(135) and walksReality(133)≈litReality(133) — nearly every walked cell lit, so the walk is close to saturated rather than shallow, and walkTruncated=true confirms it hit the depth/time budget rather than terminating naturally. Adjustment: raise --max-depth or the time budget specifically for large single-file comms diffs (this commit's HTML draft is 107 lines in one block) so the walk isn't truncated before it can find the correct, more-distant voice-room coordinates. Experiments: Re-run the reef seed for sha d94f1ecb8 with docs/comms/newsletter/** added to a room's corpus and measure whether prewalkOverlap rises above 4 · A/B the lane-scorer: recompute offPct for architect's last 49 events using Relevant-Rooms[0] instead of Originating-Terminal and measure how many drop below the 15% threshold · Extend each room's ingest corpus to include its full Relevant-Rooms trailer history (not just owned-surface globs) for 90 days, then re-check whether prewalkOverlap improves on comms-tagged commits

     Measured outcome (deterministic, not self-graded): room mean offPct 20→16 over 11 commits since · dictionary ablation saved 1289B

     This lane's apparent loss-ratio is inflated by a manufactured 23%-over-threshold event stemming from an unclaimed directory rather than real drift; premium is restored by declaring docs/comms/newsletter/**+tests/comms/** to voice and re-scoring lane-conformance against Relevant-Rooms so the receipt reflects genuine boundary crossings instead of a configuration gap.

     Open the signed attestation →

119. 2026-08-26 · 🔒 vault · offPct 16

     feat(elicit): the watchdog can see the steer cycle — and open is not the same as stale

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 26→12 over 10 commits since

     Open the signed attestation →

120. 2026-08-26 · 🔒 vault · offPct 21 · mistagged

     fix(chat): stop the popup cycling — the sidebar IS the recency index

     Why it left the lane: The entire diff lives in scripts/chat/*.{mjs,sh} and tests/chat/*.test.mjs (tcc-gate.mjs, gemini-sidebar-order.sh, gemini-web-all.sh, corpus-reconcile.mjs, whatsapp-sync.mjs, hourly-path-is-token-free.test.mjs) — zero files under vault's declared globs (docs/01-business/patents/, docs/legal/, data-room*/, docs/incidents/). The receipt's intent lane was seeded from the room label 'vault', not from the diff. The commit's own trailers explain why: Relevant-Rooms lists '🧭 navigator, 🔒 vault, 🎩 operator' with vault appearing purely because Originating-Terminal is also vault — the same fallback e94a8cd14 diagnosed. Terms like 'recordDenial', 'kTCCServiceSystemPolicyAppData', 'refused ... m ago', 'gate' read as permission/audit vocabulary that could plausibly gzip-NCD-match legal/incident corpus text, but the more direct driver is structural: vault was fed in as an anchor lane with zero true file overlap, so essentially all 131 lit reality cells were guaranteed to register outside it.

     Intent–reality gap: The commit message is internally coherent and fully describes chat/TCC infrastructure work (stopping a macOS permission-dialog loop). The receipt measured drift not because that work drifted, but because the room the receipt attributes it to — vault, whose lane is patents/legal/data-room/incidents — never had any claim on this diff in the first place.

     • Write scripts/ops/relevant-rooms.mjs (real glob-match scorer, changed-files in, top-3-or-fewer out)
     • Wire it into the commit trailer generation step (hook or session convention) so Relevant-Rooms stops being hand-filled
     • Update CLAUDE.md's Relevant-Rooms bullet per spec_rewrite above
     • Add tests/rooms/relevant-rooms-real-match.test.mjs asserting every room named in the last N commits' Relevant-Rooms trailer has ≥1 real glob-matched file
     • Audit vault's other 19 out-of-lane events in the last 14 days for the same zero-match signature to size how much of the room's offPct trend this one defect explains
     • Decide and document which room (if any) owns scripts/chat/** + tests/chat/** — currently no room's owned-surface glob claims it

     What we learned: The Relevant-Rooms trailer field documents a scoring rule but names no generator — Originating-Terminal has one ('detected by scripts/open-room-session.sh --print-label'), Relevant-Rooms does not. With no deterministic source, the field is filled by the committing session's own judgment, which structurally self-selects the room it's physically sitting in. e94a8cd14 fixed WHICH field the pipeline reads and measured a real gain (offPct 21→12), but that fix left the field itself hand-authored; this event is the same defect surfacing again at its actual source — a spec with no running code behind it.

     The missing guard (the deliverable): tests/rooms/relevant-rooms-real-match.test.mjs — for every commit's Relevant-Rooms trailer, assert each named room has at least one changed file matching that room's owned-surface glob in .workflow/rooms/*.html; fails loudly if a zero-match room (especially one equal to Originating-Terminal) appears.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Both sides are saturated (100% lit-to-walked), not shallow — every node the walk visited also lit. That's not evidence of under-walking; it suggests the walk isn't discriminating relevant from irrelevant once it starts from a bad seed. Concrete adjustment: run this same commit with --max-depth reduced by half and compare litReality — if it stays near 100% saturation, the anchor set (not depth) is the lever to fix, confirming the seed-not-walk diagnosis above. Experiments: Re-run the receipt pipeline for this sha with room label forced to whatever scripts/ops/relevant-rooms.mjs actually computes (likely fewer than 3 rooms, possibly none) and measure whether offPct drops below 15 · Backfill scripts/ops/relevant-rooms.mjs against the last 20 vault out-of-lane events and count how many share e1cce21c4's zero-owned-surface-match signature · Ablate the reef seed: exclude Originating-Terminal from intent-lane seeding entirely for one week of commits and compare mean offPct across all 9 rooms before/after

     Measured outcome (deterministic, not self-graded): room mean offPct —→21 over 5 commits since · dictionary ablation saved 500B

     A mistagged-room false positive inflates vault's measured loss-ratio on work it never touched, diluting the receipt's signal for genuine vault drift (patent/legal boundary crossings) — premium on this lane is restored by shipping the deterministic Relevant-Rooms scorer so vault's offPct trend reflects only commits that actually touched its owned surface.

     Open the signed attestation →

121. 2026-08-26 · 🔒 vault · offPct 20 · mistagged

     chore(elicit): the episode keeps its record and sheds the rest

     Why it left the lane: The commit message is written in the same forensic, self-diagnostic register as docs/incidents/ postmortems — 'the episode keeps its record and sheds the rest', 'only trip-encircled.png is the artefact the episode is actually read through', 'Found by checking whether anything I built was left half-finished, not by a test.' That vocabulary (record/episode/artefact/half-finished, root-cause narration) gzip-NCD-matches vault's docs/incidents/ corpus even though neither changed file (.gitignore, tests/elicit/steer-cycle.test.mjs) sits anywhere under vault's owned globs (docs/01-business/patents/, docs/legal/, data-room*/, docs/incidents/). The reef seeded on prose, the walk then saturated fully inside that seeded region.

     Intent–reality gap: The trailers claim Originating-Terminal 🔒 vault and Relevant-Rooms 🔒 vault, 🔨 builder, 🧪 laboratory — but zero changed files (.gitignore, tests/elicit/steer-cycle.test.mjs) match any vault owned-surface glob. The receipt correctly measured the work as builder/laboratory pipeline-hygiene content; the trailer over-claims vault.

     • Add glob-match enforcement to the Relevant-Rooms trailer generator (no fallback to Originating-Terminal on zero match)
     • Update CLAUDE.md's COMMIT MESSAGE STRUCTURE section with the explicit zero-match rule text above
     • Add tests/vault/relevant-rooms-glob-match.test.mjs asserting a synthetic commit touching only tests/elicit/ + .gitignore does NOT get vault in Relevant-Rooms
     • Down-weight incident/postmortem-register vocabulary (record, episode, artefact, half-finished, root-cause narration) in the reef's vault-seed dictionary so prose style alone can't pull placement
     • Re-run this sha (e62e96057) through the pipeline after the guard ships to confirm offPct drops below 15

     What we learned: The prior fix (e94a8cd14) addressed the Originating-Terminal fallback narratively but shipped no enforcement guard, so it regressed by trend alone (13→17 vault out-of-lane events in 14 days) rather than being structurally closed. This event exposes a SECOND, independent channel into the same failure: even a commit whose Originating-Terminal fallback were fixed would still mis-seed vault, because the commit message's own forensic/incident-report VOICE (borrowed from CLAUDE.md's incident-writing convention) gzip-NCD-matches vault's incidents corpus on vocabulary alone, with zero path overlap. Room mistagging in this repo has two separate root causes — a trailer-generation fallback bug and a prose-register bleed into the reef seed — and fixing one without the other leaves the class alive.

     The missing guard (the deliverable): A pre-commit check that computes actual owned-surface glob matches for the staged diff against .workflow/rooms/*.html and fails/warns if any room named in Relevant-Rooms (especially one equal to Originating-Terminal) has zero matched files — forcing the trailer to name the truth or explicitly declare 'no owned-surface match'.

     Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Right-sized, not shallow or under-saturated — both sides hit 100% lit with no truncation. The asymmetry worth adjusting is the SEED, not the depth: reality's walk covered 134 cells vs intent's 105, consistent with the incident-prose vocabulary giving the reality seed a wider footprint in vault's corpus than the actual diff paths justify. No depth-cap change indicated; the lever is upstream at seeding. Experiments: Strip incident/postmortem-register terms (record, episode, artefact, half-finished, root-cause narration) from vault's reef seed dictionary and re-run e62e96057 through the pipeline; measure whether offPct drops below 15 · Add idf-style down-weighting for vocabulary shared across ALL 9 rooms' commit messages (not room-specific) in reef seeding, then re-run the last 17 vault out-of-lane events and compare offPct trend before/after · When Relevant-Rooms lists a room with zero glob-matched files, seed that room's reef contribution from the diff paths only (never the commit message prose) and measure whether reef_signal_pct's share of drift on such commits falls

     Measured outcome (deterministic, not self-graded): room mean offPct 19→28 over 8 commits since · dictionary ablation saved 757B

     A room whose trailer claims content it never touched is an unpriceable attribution risk — vault's loss-ratio reads worse than its actual exposure until the glob-match guard ships, at which point offPct on pipeline-hygiene commits should fall to the room that actually owns them and vault's premium reflects only real patent/legal/incident work.

     Open the signed attestation →

122. 2026-08-26 · 📐 architect · offPct 29

     comms(newsletter): A4 — one unbroken sequence, and adjacency IS the argument

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 18→23 over 4 commits since

     Open the signed attestation →

123. 2026-08-26 · 🔒 vault · offPct 35 · mistagged

     fix(elicit): a send that failed was being recorded as sent

     Why it left the lane: The diff's actual operative vocabulary — `const token = `steer-${cycle_id}``, `'--token', token`, `sent = /spec thread OPEN/.test(out)`, `recordStage(cycle_id, 'pre-sent', ...)`, and the whole spec-thread.mjs `--token` override — is cycle/pipeline/email-dispatch machinery. None of it touches anything in vault's declared owned surface (docs/01-business/patents/, docs/legal/, data-room*/, docs/incidents/). The changed files (scripts/elicit/steer-cycle.mjs, scripts/pmu/spec-thread.mjs, tests/elicit/steer-cycle.test.mjs) sit entirely outside vault's glob set, so the reef placement for this commit has nowhere in vault's lane to land — offPct=35 is the walk finding real mass and none of it fitting the claimed room's territory. No region narrative is cached for this sha, so I anchor on the owned-surface mismatch itself rather than inventing coordinates.

     Intent–reality gap: The commit message and trailers claim this is vault's work (Originating-Terminal: 🔒 vault) and describe a legitimate, well-scoped bugfix to a send-status bug. The receipt measured the reality lane as pipeline/cycle code with zero overlap in vault's owned surface — the fix is correct, but the room claim on it is not.

     • Read the script that generates the 'Relevant-Rooms' trailer (grep for the '3 highest-scoring matches against .workflow/rooms/*.html' logic referenced in CLAUDE.md's commit-message-structure section) and check whether it unconditionally includes Originating-Terminal in the top-3 even when that room scores zero real glob matches — vault leading this commit's Relevant-Rooms list with 0 files in its owned surface is the tell.
     • Confirm which room's .workflow/rooms/*.html owned-surface glob actually should claim scripts/elicit/, scripts/pmu/, tests/elicit/ (operator and laboratory are the other two Relevant-Rooms named on this commit) and verify their glob lists currently include those paths.
     • Repoint the receipt pipeline's lane/offPct computation (the successor to e6c2aa4ce's fix) to key off the Relevant-Room with the highest real glob-match score among the three listed, not the first-listed or Originating-Terminal by default.
     • Re-run the last 13 vault out-of-lane events from the last 14 days through the corrected attribution to see how many reclassify to operator/laboratory — that count is the test of whether this is systemic or a one-off.
     • Add a guard asserting the Relevant-Rooms trailer generator never lists a room with zero owned-surface-glob matches unless every candidate room scored zero (i.e. no silent self-inclusion fallback).

     What we learned: e6c2aa4ce already diagnosed 'read Relevant-Rooms not Originating-Terminal' and measured a real improvement (offPct null→3), but this event — 13 out-of-lane vault events in 14 days, offPct=35 — shows the underlying attribution is still broken even when Relevant-Rooms is populated, because Relevant-Rooms itself lists vault first despite vault's owned-surface glob matching zero changed files. The prior fix addressed WHICH field the pipeline reads; this event exposes that the field being read can itself be wrong — the trailer generator appears to fall back to including Originating-Terminal as 'relevant' even with no real glob match, which re-legitimizes the mistag instead of exposing it.

     The missing guard (the deliverable): A test asserting: for any commit, if a candidate room's owned-surface glob matches zero of the commit's changed files, that room must not appear in the computed Relevant-Rooms trailer unless every other candidate room also scored zero — i.e. no silent self-inclusion-of-Originating-Terminal fallback when it has no real file-surface claim on the diff.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Saturated, not shallow — walksIntent=litIntent and walksReality=litReality means every walked cell lit, with walkTruncated=true suggesting the walk hit its depth/time budget before naturally tapering off. A fully-saturated walk with a truncation flag can't distinguish 'real broad drift' from 'ran out of budget mid-spread.' Adjustment: lower --max-depth on a sample of these truncated runs and check whether litIntent/litReality stop pinning at 100% — if they do, the current budget is the thing manufacturing apparent saturation, not the underlying semantic spread. Experiments: Recompute prewalkOverlap seeding from the non-origin Relevant-Room entries (laboratory, operator) instead of Originating-Terminal (vault) for this commit and its 12 siblings, and measure whether offPct drops — this directly tests the mistagged-attribution hypothesis against the detector's own numbers. · Halve --max-depth on a batch of recent walkTruncated=true commits and check whether litIntent/litReality stop saturating at ~92%/84% — if the saturation persists at a lower budget, it's real signal, not a budget artifact. · A/B the lane-threshold computation using 'best-matching Relevant-Room glob' vs 'Originating-Terminal glob' across this room's 13 recent out-of-lane events and report the resulting mean offPct for each — the comparison the checklist's item 4 needs before the guard can be written correctly.

     Measured outcome (deterministic, not self-graded): room mean offPct 21→12 over 11 commits since · dictionary ablation saved 887B

     A mistagged-room event inflates vault's measured loss-ratio with noise that has nothing to do with vault's actual risk surface (patents/legal/data-room), which either masks real vault drift under false positives or trains the room's premium on the wrong signal entirely — restored only by fixing the attribution step so vault's offPct reflects vault's own file surface, not whichever terminal the operator happened to be sitting in.

     Open the signed attestation →

124. 2026-08-26 · 📐 architect · offPct 32

     comms(bookclub): unsubscribe martin@xpanda.ai with a dated record, in blocklist not tombstone

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 13→14 over 5 commits since

     Open the signed attestation →

125. 2026-08-26 · ☕ network · offPct 27

     guard(comms): catch dead links to our own domain before a send ships one

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 15→14 over 4 commits since

     Open the signed attestation →

126. 2026-08-26 · 🎭 performer · offPct 49

     feat(fundraise): the gift-range chart — 11 checks, $300k lead, 26 qualified prospects

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct —→29 over 4 commits since

     Open the signed attestation →

127. 2026-08-25 · 🎤 voice · offPct 18

     feat(outbox): a draft with no clock is a note to self — schedule the send, one clock, $0 idle

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 27→12 over 3 commits since

     Open the signed attestation →

128. 2026-08-25 · 🔨 builder · offPct 18

     fix(email): the welcome-email service captured its clients at import, so no test could reach them

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 25→17 over 3 commits since

     Open the signed attestation →

129. 2026-08-25 · 📐 architect · offPct 21

     feat(deck): the captain frame — print deck back to parity at ten slides

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 27→15 over 3 commits since

     Open the signed attestation →

130. 2026-08-25 · 🔒 vault · offPct 21

     feat(milestones): the chief-of-staff read — what can close without waiting

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct —→32 over 9 commits since

     Open the signed attestation →

131. 2026-08-25 · 🎤 voice · offPct 20

     docs(mdx-cook): bank the ten things the revision rounds actually taught

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 25→16 over 8 commits since

     Open the signed attestation →

132. 2026-08-25 · ☕ network · offPct 23

     feat(comms): the forwardable ledger — sent/ is not the same claim as delivered

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 38→34 over 3 commits since

     Open the signed attestation →

133. 2026-08-25 · 📐 architect · offPct 37

     content(iamfim): the fork above the fold — two doors, two nightmares, one slot

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 17→24 over 7 commits since

     Open the signed attestation →

134. 2026-08-25 · 🔨 builder · offPct 22

     fix(ops): the job-health probe was printing a success line as evidence for a failure

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 27→15 over 3 commits since

     Open the signed attestation →

135. 2026-08-25 · 📐 architect · offPct 41

     fix(surfaces): restore two blocks I deleted by accident, and converge the buy gesture on /pricing

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 20→16 over 3 commits since

     Open the signed attestation →

136. 2026-08-25 · 🎭 Alacritty Performer · offPct 57

     feat(ops): one cross-process slot stack for qwen and heavy shell workers

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct —→26 over 3 commits since

     Open the signed attestation →

137. 2026-08-25 · 🎭 Alacritty Performer · offPct 30

     fix(ops): wire the hot paths onto the slot stack, and serialise ollama at the server

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct —→26 over 3 commits since

     Open the signed attestation →

138. 2026-08-25 · 📐 architect · offPct 61

     content(surfaces): the reason a developer actually sends it, and a receipt an executive can just look at

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 20→18 over 3 commits since

     Open the signed attestation →

139. 2026-08-25 · 📐 architect · offPct 25

     content(resources): the grip deck as a file, not only as a route

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 20→22 over 4 commits since

     Open the signed attestation →

140. 2026-08-25 · 🎤 voice · offPct 49

     fix(routes): link /big-map and /C from /pixel, /dispatches from /blog

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 27→23 over 3 commits since

     Open the signed attestation →

141. 2026-08-25 · 🔒 vault · offPct 37

     feat(ops): notify when the role you asked for starts acting like a different one

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct —→21 over 4 commits since

     Open the signed attestation →

142. 2026-08-25 · 📐 architect · offPct 36

     content(surfaces): the fork in the deck's own grammar, and /cta becomes the ratchet

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 19→27 over 3 commits since

     Open the signed attestation →

143. 2026-08-25 · 🎤 voice · offPct 24

     fix(rooms): the role-assignment construct missed the gerund form

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 26→26 over 5 commits since

     Open the signed attestation →

144. 2026-08-25 · 🔨 builder · offPct 49

     feat(spe-channel): three pages that link is not yet a channel — declare the order, then guard it

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 19→23 over 6 commits since

     Open the signed attestation →

145. 2026-08-25 · 🔒 vault · offPct 27

     fix(pulse): the probe was blind to the chat the operator was typing in

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 26→25 over 6 commits since

     Open the signed attestation →

146. 2026-08-25 · ☕ network · offPct 28

     test(comms): a paste-ready block ships the bytes the human pastes, not markdown

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 28→37 over 3 commits since

     Open the signed attestation →

147. 2026-08-25 · 📐 architect · offPct 21

     feat(stickers): five designs on one line, and the line lands on iamfim additively

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 28→18 over 4 commits since

     Open the signed attestation →

148. 2026-08-25 · 📐 VS Code Architect · offPct 35

     feat(liability): re-sequence as the 1+1 reveal, kill the pitch-slap

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct —→19 over 6 commits since

     Open the signed attestation →

149. 2026-08-25 · 📐 architect · offPct 30

     feat(playbook): the ladder was on the wall with nothing above it to say when it fires

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 22→7 over 3 commits since

     Open the signed attestation →

150. 2026-08-25 · ☕ network · offPct 19

     feat(comms): the Paul email in the corporate register, with the proof kept

     

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct —→10 over 4 commits since

     Open the signed attestation →