Published on: July 20, 2026
Ready to accelerate your breakthrough? Send yourself an Un-Robocall™ • Get transcript when logged in
Send Strategic Nudge (30 seconds)Green in-lane · amber a little out · red drift. Every panel is a real commit, byte-identical on recompute. Tap any panel to open its shareable receipt.
Today the question landed in its simplest form: how big is the map, and how much of the territory does it actually cover? It came with a correction attached — stop calling it an index. An index helps you find a file. This thing is a compression map you can walk: the whole repository projected down into a 12-by-12 grid of text cells, each cell composed from two authored self-definitions, and every step any process takes recorded as a coordinate that opens back into one of those cells. This post documents where that map stands as of this afternoon, what winning looks like in three numbers, and what it will take for someone who is not us to find it useful. The code is public: github.com/wiber/thetacog-mcp.
The win condition for this post, declared before the first course: you leave and recompute. Every number below carries the command that produced it. If you leave merely nodding, the post failed — nodding is what a pitch produces; recomputation is what an instrument produces. Grade us on that.
Don't take a sentence of what follows on faith. Run npx thetacog-mcp attest-demo on your own machine, and here is exactly what comes back: a signed placement verdict — in-domain, out, or unplaced — built from a 144-cell grid on your silicon, with no model in the loop. That is why we can say "why we believe" without asking you to believe anything: the proof runs before the belief does. Everything below is commentary on artifacts that command can reach.
The claim you are invited to attack, one sentence: where a piece of work lands on this map is a decidable, recomputable fact; whether the work is good remains undecidable, and the receipt says so on every line. Swing at either half.
Twelve axes — Strategy, Tactics, Operations, and their nine children — each own an authored self-definition: a paragraph of text that says what that axis is. A cell of the map is the meld of two of them: the actor's text fused with the patient's. Cell B,A1 is Tactics acting on Strategy.Law, and its opening sentence is literally built from both parents' definitions — we measured it today, snippet by snippet: cell A,B carries two of Strategy's self-definition passages and four of Tactics', word for word, before the harvest loop ever densified it with repository passages. That is why this is not an index. An index points at content; these cells are content, composed first and grown second.
The walking part is just as literal. Every cycle the system runs, it records breadcrumbs — ply0: B,A1(0.057) — the coordinate plus how hard the walk gripped there. We checked every breadcrumb ever recorded: 1,873 walks, 27,766 hop-cells, and every single one dereferences into a non-empty cell. Zero dangling. That property is now a test (lineage-coords-resolve), so it cannot silently rot.
The map: 144 cells, 202.4 thousand characters of authored and harvested text, minimum cell 671 characters, median 1,477. The territory: 6,206 semantic files across docs, scripts, source, and tests, partitioned into 218 sections. Coverage is measured as grip — for each section, a bounded probe of its own text is compressed against the map (gzip-NCD, the same sensor everything else here uses), descending the hierarchy: twelve row axes first, then only the winning row's twelve cells. A section the map compresses well is territory the map speaks for.
Today's number: 105 of 218 sections gripped — 48.2%. And the useful half of the instrument is the remainder: the ten worst-gripped sections come out ranked, with the nearest cell named. The worst today is the application source itself (682 files, fit 0.971), followed by the analytics and troubleshooting docs. That list is not a to-do we wrote; it is the harvest frontier, computed — the demand engine's next targets fall out of the same measurement that reports coverage.
What we deliberately do not do: compress the whole repository at once. We measured that too — a single gzip pass over the full 57 MB takes 4.6 seconds, and compression distance needs pairwise passes, so whole-repo comparison would take minutes to hours. The map exists precisely so that never has to happen: the grep asks the repo for a lane's material in the lane's own vocabulary, and compression only ever touches candidate passages. The entire gzip cost of a mining cycle is about 25 milliseconds.
Recompute both numbers: node scripts/pmu/reef-coverage.mjs in the repo prints the map size, the grip percentage, and the frontier; run it twice and the bytes are identical — a guard enforces that the coverage number is not weather.
This morning the autonomous loop's narration stage — the one place a local model still ran every cycle — was 53 seconds of a 94-second tick. As of today the deterministic narrator is the default: the same cycle's story is read off the walk's own artifacts in under a tenth of a second, byte-identical across runs, and the tick dropped to 32 seconds, zero of them model time. The flip could not land silently even if we had wanted it to: the spec declared the literal marker in advance, the predicate is anchored to a real assignment, and the flip only landed with warm-model A/B evidence attached — measured against a model kept resident, with the memory cost stated both ways, because a cold-start comparison would have justified the flip on a lie.
The same discipline caught its own false positive hours earlier: a milestone predicate was briefly satisfied by a code comment describing the marker rather than the marker itself. The auditor caught it, retracted the emailed receipt, re-anchored the predicate, and shipped a guard asserting that predicates agree with the tree's actual state — never with documentation about it. A system that celebrates milestones must be able to un-celebrate them, or the celebration means nothing.
The end state is three numbers, each already instrumented. Total grip: every substantive section of the repository compressed into the map's cells — 48.2% today, steered upward by pointing the harvest at the computed frontier, with the loop already running unattended at a 50% duty cycle. Zero-model execution as the default: done as of this morning, and guarded — the verdict path is structurally unable to read a model's output. Provable portability: the verdict, the placement, and the receipt run from the public package on a stranger's machine with no repository access — the bundle is smoke-tested from a bare temp directory on every build.
What is honestly not proven yet, stated so it cannot quietly disappear: everything measured is on our own repository. The map's generality — bootstrapping a foreign repo's own 144 cells and reaching a comparable grip on territory we did not author — is the gap between an internal instrument and a product, and no number on today's tape speaks to it. Second: 48.2% is a floor reading from the instrument's first day; the grading probe is bounded and deterministic, but whether grip at 0.92 is the right ceiling for "the map speaks for this" will be recalibrated as the frontier sections get harvested and re-measured. The number is honest; its threshold is young.
Nobody needs the tesseract mathematics to use this. The transition happens through three operational shifts, each of which replaces a feeling with a number. For the engineer, documentation stops being a wiki that rots: the repository is the map, and the map densifies itself — the cells grow from real repository passages every cycle, and the coverage instrument tells you which territory is still unspoken for. For the operator of AI systems, prompting stops being finger-crossing: work is placed on the map locally, and drift past the policy line halts with an out-of-domain verdict and a signed receipt instead of a vibe. For the underwriter, incidents stop being ghosts: time-in-lane, off-domain rate, and drift velocity are countable, recomputable series — the difference between a vendor's claim and a book of business you can price.
The keys are already public: github.com/wiber/thetacog-mcp — npx thetacog-mcp attest-demo is the front door, and the full derivation lives in the book, Tesseract Physics — Fire Together, Ground Together, which earns the same claims through argument rather than commands. The map is at 48.2% and climbing on its own. What we can reach from here is a repository that finishes describing itself — and then does it to a repository that is not ours.
Ingredients only — what is on the record, for you to conclude from. The coverage instrument and its guard: scripts/pmu/reef-coverage.mjs and tests/pmu-simulator/reef-coverage.test.mjs in the public repo — determinism, schema, and a shared fit ceiling with the demand engine, asserted. The join-totality guard: tests/pmu-simulator/lineage-coords-resolve.test.mjs — 1,873 walks, every coordinate dereferences. The narrator flip and its evidence packet: NARRATOR_DEFAULT=B in the loop with warm-model A/B measurements attached, the false-receipt retraction preserved in the milestone dedup state. The timing measurements: 57 MB territory, 4.6-second single-pass gzip, 25-millisecond per-cycle compression share, all reproduced in the moving-parts ledger (docs/architecture/moving-parts-ledger-2026-07-20.md, part 13). The book carries the derivation: the divergent-series argument for why the walk terminates where a model cannot.
The to-do, ours and yours. Ours: steer the harvest at the computed frontier and re-measure grip weekly; wire the tick panel to the live tape; bootstrap a foreign repository's map — the generality gap named above, and the only claim we have not earned. Yours, if any of this held your attention: npx thetacog-mcp attest-demo, then swing at the attackable claim in section A. Related reading: The Post-Commit XOR Gate for how the receipt attaches to every commit, and Decidable on Silicon for the claim this whole instrument exists to defend.
![tolerance panel for commit f9cea3b — chore(commit-page): panel-repair 1f1766012 — re-grounded panels [panel-attached]](/blog/post-panels/2026-07-20-the-walkable-compression-map/f9cea3b8fac6db6d123d126177fcc74d95afa5e3.png)