docs(marsh): close the loop on the tuned held-out + abstention runs — the null that removes an argument

The tuned-vs-generic held-out experiment and the abstention-rate probe (scripts/pmu/abstention-rate-probe.mjs,
data/pmu/study/2026-07-14-*.json) had already been RUN today but their findings never made it into
data/pmu/public-claims.json or the marsh-numbers-HELD.md doc — so the pinned-claims ledger and the actual
outreach material were about to walk into Thursday's call carrying a stale hope ("tuning should improve the
30%") instead of the measured result (it does not: tuned and generic both land 0.90/0.30, identical to three
figures). Recorded that as a NULL result next to drift_does_not_predict_defects, and recorded the abstention
rate (0/30) with the mechanism caveat that matters: UNPLACEABLE only fires when neither witness places a cell
at all, not on disagreement or a near-threshold score, so a 0% rate here does not yet prove the near-the-line
deadband behaves as claimed (that measurement, ε via cross-zlib-version max|Δσ|, is separately parked for
2026-07-15). Guard tests/pmu/public-claims-are-pinned.test.mjs still green.

Also delegated an unrelated pre-existing failure found while running the suite (tests/pmu/strike-price-chain.test.mjs,
worked-example job f5434057a has aged out of the live ledger) to the laboratory room via bifurcate (bf-1426) —
out of lane for today's Marsh de-risking task.

Originating-Terminal: 🎩 operator
Relevant-Rooms: 🧪 laboratory, ☕ network, 📐 architect
Story: The GCal 'RUN THE EXPERIMENTS NOW' fire found both blocking experiments already executed (self-prompts
sp-078/sp-081 record them fired ~20 minutes prior), but the loop was left open — data existed, the ledger and
outreach doc did not reflect it. Closing that loop is the actual de-risking work for a call that may land
Wednesday instead of Thursday.
Persona-Intent: Mike Carls (MD, Data & Analytics, FINPRO US) — reads a vendor who ran the exact test his own
hypothesis implied, got a null, and wrote it down before he could find it himself; walks away trusting the
0.90/4.48σ numbers MORE because the hopeful one was retracted in public, not because every number went up.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

commit 4c9a5dfa9 — 144×144 lattice triptych (on-chip)

◎ the ingest chose A1,A — Strategy·Law × Strategy — acting on B2,A1 — Tactics·Deal × Strategy·Law (grip 0.797) · the walk starts at the ACTOR's row · 9 file(s)

⏱ ingest 584ms (corpus→lattice) · definer walk 280ms (row→transpose→row, ballistic XOR per hop) · total 1177ms
ingest here = commit-scoped SENSING only (this commit's message + changed files → lit anchors) — deep seed authoring lives in the reef-self-loop, off the commit path

rows = ACTOR (lens) · cols = PATIENT (object) · diagonal = self-reference; ◎ at actor-row × patient-col — every panel, same orientation. Heat above the diagonal = actors ranking ShortLex-earlier than their patients (the walk follows definers uphill in rank).
The 12 canonical lanes (rows top→bottom, cols left→right): A·Strategy · B·Tactics · C·Operations · A1·Law · A2·Goal · A3·Fund · B1·Speed · B2·Deal · B3·Signal · C1·Grid · C2·Loop · C3·Flow

PAIR LATTICE (the map we normally see)

PRE-WALK SENSE · INTENT grid — the snapshot, no point of view

the no-point-of-view snapshot — what the sensor lit before any walk; the leaf walk below is what bridges the two grids

	ABC A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N A1–C3
ABC A1–C3 A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N

PRE-WALK SENSE · REALITY grid — the snapshot, no point of view

the no-point-of-view snapshot, reality side — same sensor law

	ABC A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N A1–C3
ABC A1–C3 A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N

PRE-WALK Δ · straight comparison of the two ingested grids (green both · cyan intent-only · amber reality-only) · 6% overlap · ◎ = where the competence pixel sits in the raw overlap

good = green overlap without being a copy — near-100% is self-similarity, not alignment

	ABC A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N A1–C3
ABC A1–C3 A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N

◎ the pixel this commit belongs to

A1,A — Strategy·Law × Strategy (actor) acting on B2,A1 — Tactics·Deal × Strategy·Law (patient) · grip 0.797

actor seed A1,A — The law provides the binding mandate and governance article that protects the long-horizon strategy substrate lattice. EU AI Act Article 14 requires human-on-the-loop oversight, with the supervisor able to interrupt the system at any moment. Compliance attaches to the artifact the artifact cannot escape — provenance is a regulatory primitive, not a feature. Constraints precede objectives. The cap-table, the license text, the consent envelope — each draws the inviolate line the action must respect, and the line is enforced before the optimization runs. A rule names what the system MAY NOT do. If the agent can do X, then 'forbid X' is not the constraint — it is wishful thinking dressed as policy. Constraints live in the gate, not in the prompt. Regulation is not advisory. The fine is the price tag stapled to the omitted control; insurance is the option pricing on the failure-mode the regulation already named. The lattice we build today is what next year's underwriter inherits. We set the substrate that downstream pricing rides on, and that choice is paid back across every quarter we keep operating. Direction is what you keep when you forget the calendar. Multi-year posture, irreversibly chosen, defines the boundary every quarter renegotiates inside. Strategy is the bearing the rotation is measured against. Capital lives where strategy is durable. Frame the ten-year shape of the business before optimizing the ten-week sprint; the substrate decisions compound while the surface decisions decay. The strategic question is which doors close on purpose. Anything that keeps every option open until next year has already chosen the option of inaction, paid in optionality cost.

patient seed B2,A1 — Deal terms and exchange rates must follow the binding mandate and governance rule of law. EU AI Act Article 14 requires human-on-the-loop oversight, with the supervisor able to interrupt the system at any moment. Compliance attaches to the artifact the artifact cannot escape — provenance is a regulatory primitive, not a feature. Constraints precede objectives. The cap-table, the license text, the consent envelope — each draws the inviolate line the action must respect, and the line is enforced before the optimization runs. The deal sets the exchange rate between two value substrates. Mis-priced exchange is a slow-bleed; correctly priced exchange compounds both sides, because each cycle of trade enriches the bench on both ends. Terms are the API between two organizations — every clause is a contract method that survives the handshake. Renegotiation is the breaking change; the deal is the stable interface either side can build against. Negotiation closes at the moment both parties' BATNAs cross. Push past that point and you've bought regret at retail; stop short and you've left margin on the table that the next deal will not return to you. A signed counterparty is more valuable than an unsigned promise. The signature is the substrate; the conversation that preceded it is the exhaust the signature converted into substrate.

INTENT · leaf-walk clouds · BLUE→VIOLET by ply (◎ = the competence pixel: actor∩patient)

good = the clouds concentrate where the commit SAYS it works

	ABC A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N A1–C3
ABC A1–C3 A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N

REALITY · leaf-walk clouds · ORANGE→RED by ply (◎ = the same pixel — the shared perspective)

good = the same clouds as intent — shipped where declared

	ABC A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N A1–C3
ABC A1–C3 A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N

DELTA · graded cloud divergence (green agree · magenta declared-not-done · amber done-not-declared) · 4% agree

good = mostly green; magenta = said-not-done (chase it in the code), amber = done-not-said (chase it in the docs)

	ABC A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N A1–C3
ABC A1–C3 A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N

TOLERANCE · cloud topology (dim green in-shape · hot amber bleed · red orthogonal concentration) · ⚠ TOO MANY

red only matters above the flip — a few amber is normal bleed

	ABC A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N A1–C3
ABC A1–C3 A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N

SHORTLEX-3 PROJECTION (ABC · A1..C3 · the 132 children — candidate, pre-ratchet)

PROJECTED INTENT · the commit’s claims, head→axis_i tail→axis_j (cyan)

the three-length view — good = claims land inside their own zone

	ABC A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N A1–C3
ABC A1–C3 A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N

PROJECTED REALITY · same projection law (amber)

same projection law — good = reality occupies the zones intent does

	ABC A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N A1–C3
ABC A1–C3 A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N

STRAIGHT COMPARISON · green both · cyan intent-only · amber reality-only · 5% overlap

good = the zones agree; cross-zone scatter is the drift to read

	ABC A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N A1–C3
ABC A1–C3 A1A–O A2A–O A3A–O B1A–O B2A–O B3A–O C1A–N C2A–N C3A–N

shortlex-3 zone occupancy (lit cells per diagonal square)
INTENT: 1 in the 3×3 ABC corner · 12 in the 9×9 A1..C3 square · 1884 in the 132×132 children square · 414 cross-zone
REALITY: 1 in the 3×3 ABC corner · 1 in the 9×9 A1..C3 square · 1328 in the 132×132 children square · 187 cross-zone
axes: identical 144 ShortLex names both sides (symmetric); weights asymmetric
zone 3 = candidate children, pre-ratchet (repo-derived dumps not yet past the perturbation-probe gate) · intent 109 claims @ θ 0.6406 · reality 20 claims @ θ 0.6094 · projection 806ms

SHAPE-MATCH σ = -0.11 — the real intent↔reality definer walk vs random-reality (actual 0.1013 vs random 0.104). >0 = the walk distinguishes real alignment from noise; aggregate over independent walks → the divergent series.
σ_drift · noise · below zero — this read is noise, not a measurement

Most out-of-lane row: A3 · Fund — that lane is drifting most (read the red band in that row).

Three-colour tolerance (out-of-lane = reality fires where intent is weak, in orthogonal blocks): green=in-lane agreement · amber=a few out-of-lane (tolerated) · red=FIRED — orthogonal out-of-lane 28% vs tolerance 25%.
922 green · 578 amber · 578 red · 168 XOR-friction cells.

Both axes = the 144 ShortLex anchors, cell=row⊕col, on the chip. INTENT=docs+rules · REALITY=code · DELTA=friction.

tile dump — what the SimHash matched on the diagonal (inspect the sensor; tap to expand)

A,A The lattice we build today is what next year's underwriter inherits.

intent 72% The 0.30 is **paraphrase-invariance**: the sensor is **surface-form dependent**. An action *del

reality 69% const unplaceableRate = n ? +(unplaceable.length / n).toFixed(3) : null;

B,B Tactics is the choice of which leverage to pull at which beat.

intent 75% property of the **sensor** — it reads words, not roles, whether the universe around it is wide

reality 72% const AUTHORIZED = arg('--authorized', 'A3');

C,C Operations is the loop that runs whether anyone watches.

intent 83% **Here is the test I promised, and I ran it before writing this down.** My working hypothesis w

reality 77% // abstains, and on what. I will bring that." Never previously measured — this is that measurem

A1,A1 EU AI Act Article 14 requires human-on-the-loop oversight, with the supervisor able to interrupt the system at

intent 77% when *neither* independent measurement can place the input on any cell at all — not when the tw

reality 70% (no distinctive match)

A2,A2 The target is the cell coordinate the system MUST occupy by end-of-quarter.

intent 78% (no distinctive match)

reality 70% function callIssuer(jobId, text) {

A3,A3 Every dollar builds either the floor (substrate that lifts future bets) or the ceiling (cap on the next round)

intent 81% (no distinctive match)

reality 69% (no distinctive match)

B1,B1 Latency is the price of a missed beat.

intent 73% (no distinctive match)

reality 69% sigma_gzip: r1.physical_execution?.witness_simhash?.gzip_sigma ?? null,

B2,B2 The deal sets the exchange rate between two value substrates.

intent 81% would rather tell you that than let a hopeful number stand unchallenged.**

reality 72% (no distinctive match)

B3,B3 Signal-to-noise is the inverse cube of bandwidth — narrow the channel, the message arrives; broaden it, recipi

intent 73% (no distinctive match)

reality 73% const log = (...a) => process.stderr.write(a.join(' ') + '\n');

C1,C1 Power flows where the grid lets it.

intent 73% (no distinctive match)

reality 69% // Output: data/pmu/study/<date>-abstention.json

C2,C2 Each iteration is a hypothesis tested against the ground truth of the world.

intent 78% (no distinctive match)

reality 70% (no distinctive match)

C3,C3 Flow is the rate at which committed work crosses the finish line.

intent 78% Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

reality 69% rows.push({ id: it.id, kind: 'primary', label: it.label, chars: it.text.length,

commit context — 9 file(s)

docs(marsh): close the loop on the tuned held-out + abstention runs — the null that removes an argument

The tuned-vs-generic held-out experiment and the abstention-rate probe (scripts/pmu/abstention-rate-probe.mjs,
data/pmu/study/2026-07-14-*.json) had already been RUN today but their findings never made it into
data/pmu/public-claims.json or the marsh-numbers-HELD.md doc — so the pinned-claims ledger and the actual
outreach material were about to walk into Thursday's call carrying a stale hope ("tuning should improve the
30%") instead of the measured result (it does not: tuned and generic both land 0.90/0.30, identical to three
figures). Recorded that as a NULL result next to drift_does_not_predict_defects, and recorded the abstention
rate (0/30) with the mechanism caveat that matters: UNPLACEABLE only fires when neither witness places a cell
at all, not on disagreement or a near-threshold score, so a 0% rate here does not yet prove the near-the-line
deadband behaves as claimed (that measurement, ε via cross-zlib-version max|Δσ|, is separately parked for
2026-07-15). Guard tests/pmu/public-claims-are-pinned.test.mjs still green.

Also delegated an unrelated pre-existing failure f

data/pmu/public-claims.json
data/pmu/study/2026-07-14-abstention.json
data/pmu/study/2026-07-14-region-tuned.json
data/pmu/study/2026-07-14-region.json
data/room-bifurcations.json
data/self-prompts.json
docs/outreach/accounts/marsh-numbers-HELD.md
scripts/pmu/abstention-rate-probe.mjs
tests/pmu/heldout-tuned-spec-narrower.test.mjs

Produced on the chip by the running pipeline (scripts/pmu/pipeline.mjs → runPipeline → walk + xor), commit-scoped · on-chip per-walk 1112.81ns (this commit). Recompute: node scripts/pmu/commit-triptych.mjs --commit 4c9a5dfa9

Commit artifacts (open on a computer — Gmail mobile won't follow file:// links; the inline images above are the phone-viewable form): reef grid · this triptych · the HDL — all three are also attached to this email.