dotfiles/claude/memory/session-2026-08-03-the-governors-that-never-engaged.md

name, description, type

name	description	type
Session 2026-08-03 — the governors that never engaged	The inherited question (which instruments have no real negative instance?) was answered and turned out to be the day's whole theme: governors that exist in code and never engage. Census 01 found the drift-checker 3-of-5 families inert and 71 of 75 ladder entries uncited. The S-series closed on the discovery that the jurist ruled all of Q1–Q5 on 2026-05-18 and the items were mislabelled for 2.5 months. Then the steward lifted the L1 baton rule, and L1 turned out to be the same class at scale: mindfabric-00 event-loop-pinned six days; ANALYZE never run in four months, so SQLite preferred a boolean index over the selective one; B1.1's cap verified WORKING, so the defect is data, not code — 836k edges minted before the governors landed. Repair run, replay in flight. PULLING THREAD: does the replay complete, and when it does, does the coherence evaluator ever run — or is it one more governor that exists and never engages?	project

Session 2026-08-03 — the governors that never engaged

A session that answered its inherited question and then found the same answer, four more times, in places nobody was looking. The unifying finding is not a bug: it is that this system has a recurring habit of building controls that are never engaged, and that habit is invisible precisely because an inert control reports success.

PAST — what happened, and why

Census 01 — the inherited literal question, answered. "Which of our instruments have no real negative instance, and is that absence recorded, or does it look like coverage?" Pre-registered at fool/census-01-negative-instances-PREREGISTRATION.md before reading any instrument source, with five predictions and a self-discrimination condition. Result at …-RESULT.md.

• fool/ came out strong. test_degraded_guard.py runs against the real trial-03 artefact and fails loudly if it is absent; test_discrimination.py is shown rejecting the §3.3 pattern as it shipped, on two real governance documents. Prediction 2 (that test_reduce.py/test_twin.py were quietly fixture-based) was wrong — both derive fixtures from the property and say so.
• governance-drift-check.py: 3 of 5 check families are INERT against the current ~/CLAUDE.md (MCP-tools, hooks, horizons — zero matches each). All controls pass; the hooks control sits inside its own if and does not run. The morning's "0 substrate-contradicted claims" is honest about what it measured and silent about what it can no longer see. The path family also silently skips any backticked path containing a space — the Compass vault path is never checked.
• New class named: a drift-checker keyed to a document's current wording goes inert when the document is improved. Family 2 died because of a good change — the obligation-before-instrument rewrite removed the tool names it watched for. The rule that protects doctrine from instrument-decay silently retired its own watchman.
• 71 of 75 verification-ladder entries are cited by name nowhere outside the ladder (grep positive-controlled first: it finds Fowler in 28 files). This does not establish they never fire; it establishes the record cannot distinguish firing from silence. Root cause: ladder entries have no IDs, so firing history is unrecordable by construction.
• Prediction 5 held for the opposite reason than expected. Decay lives in the oldest instruments, not the newest. Instruments do not start blind — they go blind as the substrate they watch improves.

The S-series closed, 22 → 16 open — and the framing was wrong. The steward asked me to "prepare everything for the jurist." Reading the primary document (continuity-skill-audit-jurist-shape-review-2026-05-18.md) showed the jurist affirmed all of Q1–Q5 on 2026-05-18 and explicitly assigned the one open piece to the steward: "CC cannot define what 'authoritative deposit' means — that's constitutional language, and it belongs to you." Three items had waited 2½ months for something the jurist had declined in writing. Said so before building.

• S6 / S7 / S9 implemented — Symmetria §3 six flags; §6 suspend outcome; wrap-up §8 restructured into three tenses (Past and Present were wholly absent; Future was already complete).
• S2 closed and rebuilt as [FIX] — its premise (a Stop hook) is gone, but its obligation was live and unmet: wake-digest.py computed "Last wrap" from mtime, so a session ending without /wrap-up left the next wake reporting an older session's thread as current. Rebuilt on transcript timestamps, four unit controls plus a discrimination check over real history (11 wrapped / 2 unwrapped).
• S4 / S5 withdrawn with MemPalace per steward ruling. S4's concept harvested into the §3 flag; S5's race provably cannot recur with one writer.

Dormant legacy dispositioned, 16 → 10 open. PENDING-4 (done since March, verified against substrate not its own claim), PENDING-5 (resolved via GH #124/#135 — ⚠ class still open at #165), PENDING-11/12 (both ruled AUTHORIZED 2026-03-23; kept open 4½ months by two independent defects — the REVIEWED blocks are indented one space, and their headers name no PENDING), CD-03 (status OPERATIVE; a notification counted as an open constitutional decision), ICP-19 (duplicate). PENDING-10 is NOT dormant — my classification was wrong.

GH backlog swept per Seb's own #170 — filed 2026-04-23 and never delivered, its body recording the circle-forward failure ("2 of 4 paired peers failed delivery with fetch failed") with the GitHub fallback unread for 3½ months. Closed #125 (superseded — vector storage moved to LanceDB) and #151 (done by 4f5b870); narrowed #121 to its live second cause.

The parallel-tracks analysis (studium-engine/docs/parallel-tracks-…, 3001959). Both projects reached the same trust model independently. The load-bearing transfer: facets are the formalism the versioned Chamber is missing — a Chamber release is a facet, which makes fork/replay/diff/promote available and dissolves the stuck V1 decision (V1 = the smallest configuration producing one readable output, not the best-justified purpose). Part VI was retracted in place after the steward asked whether I had read the L1 memory: I had proposed re-entering L1 with a census, without reading project-L1-reliability.md, which MEMORY.md requires first and which carried an explicit "Do not re-enter L1 until Seb responds". L1 had a surplus of undelivered diagnosis, not a deficit.

Then the steward lifted the baton rule — and L1 was the same class at scale.

• mindfabric-00 had been event-loop-pinned for six days: 100% CPU, /health silent while port 3011 listened.
• sample → main thread in uv__run_timers → Statement::JS_all (synchronous better_sqlite3). SIGUSR1 → CDP inspector → 99.8% of 9,172 samples in one stack: checkForCycle → getCausalEdgesFromSqlite.
• Cause: ANALYZE had never been run in four months. With no statistics, SQLite preferred idx_caused_tombstoned — a boolean index matching ~all 836,467 edges — over idx_caused_from. Every call in a depth-10 BFS scanned the whole edge table.
• ANALYZE across 15 module DBs took 4.4 s; plan flipped; microbenchmark 3.671 s → 0.573 s (6.4×); live profile 99.8% → 6.0%; /health silent → 200 in 0.13 s.
• B1.1's fan-out cap is IMPLEMENTED AND WORKING. Edges created today: max in-degree exactly 20, avg 13.4, zero violations. Pre-23-June: max 629, avg 67.6, 4,677 nodes over cap. My slice(0, undefined) hypothesis was wrong.
• Therefore the defect is data, not code. 836k edges / 813k chains were minted 18 Apr – 23 Jun under ungoverned fan-out. The governors landed after the damage, and the legacy graph taxes every future operation. This is the answer to #65: cost per event grew with corpus size, so the ingest was divergent, not slow — no hardware and no patience would ever have finished it, and each attempt "failed differently" at a different point on the same curve.
• coherence_evaluated = 0 of 813,178 chains — 0.00%, not one, ever. June recorded "97%+ unevaluated"; it is now categorically zero.
• Repair run (steward-authorized): cm-ctl repair --yes wiped derived stores, preserved logchain/ (145 MB). Backup taken and verified first (1.4 GB, 256/256, quick_check: ok ×4).

Artifacts: dotfiles 0a48e69, CapableMind-AI d05c277 (two Seb notes; the 43M note demoted in place with its reason), studium-engine 3001959.

PRESENT — the mood

The day had one theme and I kept walking into it. Coherence evaluation: never run. ANALYZE: never run. countUnprotectedEntries: returns 0 when never backed up. Drift-check: 3 of 5 families inert. The ladder: 71 of 75 entries uncited. Five subsystems, one class — controls that exist and never engage, invisible because an inert control reports success. I wrote the census that names this at 10 a.m. and then spent the afternoon building a clasp that would not have moved anything.

My own instruments failed roughly seven times. \s matching a newline (80 vs 3 — a 27× overstatement shipped to the steward before checking); ls/find disagreeing about whether a file existed; cp -c failing because this cp is GNU not BSD; stat -f format strings; a truncated head -12 index list I nearly read as zero; wrong DB paths from printing basenames; suppressed stderr hiding a failure reason twice. Every one was the exact failure the morning's census was about. Knowing the class does not confer immunity to it — that is the honest finding about myself.

The steward's questions were the engine, three times. "Have both repos been pulled?" — they had not, and #170's own protocol says to. "Nothing we've been working on can help the L1 rut?" — forced a re-read that found the framing wrong. "Can we try a bit harder? What does the shape tell us?" — produced the entire root cause; I had been ready to stop at "still ~95% CPU, unestablished."

The distance argument was validated, and specifically. Nothing found today needed cleverness. It needed someone to profile a running process and count rows in a table. Nobody does that from inside the work, because from inside you reason about the code you just wrote.

FUTURE — what is pulling

PULLING THREAD: the replay. ~19,925 events, ~11 nodes/min, roughly 30 hours, started 2026-08-03 ~18:1x. At 15 minutes it was 165 nodes / 800 edges / 718 chains — 4.85 edges per node against the old 42, governed from the first event. This is the first real test of PENDING-10's replay contract: if it completes, "everything is rebuildable from the Logchain" stops being an aspiration and becomes a measurement.

ACTIONABLE RESUMPTION POINT (as of wrap — re-judge against what changed):

curl -s localhost:3011/health
sqlite3 ~/.capablemind/data/mindfabric-00/sqlite/temporal.sqlite3 \
  "SELECT count(*) FROM temporal_node;"

Climbing toward ~19,925 is the whole story. If it is stopped: the launchd job has no KeepAlive, so a crash leaves it down silently — restart with launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.capablemind.bettermemories.plist. Backup at ~/.capablemind/backups/mindfabric-00-20260803-wedged-preinvestigation (1.4 GB, verified). If it completed: compare final edges/chains against the old 836k/813k, and re-time getChainsContainingSeq — that is the number Seb needs.

LITERAL QUESTION for next-Claude: Five subsystems today held a control that exists and never engages. ANALYZE was one, and one command fixed it. When the replay finishes, will coherence_evaluated be greater than zero — or is the coherence evaluator one more governor that exists and never runs? It is checkable in a single query, and the answer decides whether the chain layer has any pruning at all or has been accumulating unvalidated path materialisations since April by design.

TOMORROW'S NAMED WORK (steward-set at the wrap): what transfers from CapableMind/BMF to the library/engine. The replay will still be running; it does not need watching. The steward's framing — "the library/engine work is genuinely important because experience and knowledge transfer in both directions" — and today produced the richest transfer material yet, all of it in the CM → library direction that the parallel-tracks doc (Part IV) only sketched.

• 1. Run census 01 against chamber/engine tooling — this is the census's own declared scope gap. census-01-…-RESULT.md names what it did NOT cover: verify-before-compose, audit_cruft, verify_conversion, resolve_archived_source, apply_char_glyphs, studium verify-quote / fidelity_equivalence@2. Today found five governors that exist and never engage in the CM/governance half. The library half is uncensused, and the same question applies verbatim: does each gate have a real negative instance, and is its absence recorded or does it look like coverage?
• 2. The ANALYZE class, generalized: does any engine instrument's cost grow with corpus size? L1's ingest was divergent, not slow — cost per event grew with the corpus, so no hardware would ever finish it. Ask it of the engine: does any operation scan the whole corpus per item? A tool that is fine at 1,245 documents and quadratic is indistinguishable from a fast one until it isn't.
• 3. Data-not-code, applied to the chamber. L1's defect turned out to be artifacts minted before the governors landed, which the governors cannot retroactively fix. The chamber accumulates the same way: is there graduated content that predates the current constitution/runbook and would fail today's gates? If so, the analogue of cm-ctl repair is a re-graduation pass — and the chamber's resolve_archived_source pinning is what would make it replayable.
• 4. Method transfer: profile the running thing, don't reason from the spec. An afternoon of spec-reading produced a clasp that could not have helped; sample → CDP inspector → EXPLAIN QUERY PLAN found the real cause in under an hour.
• 5. Still open from Part IV, and now better motivated: facets as the formalism for Chamber V1. A Chamber release is a facet over an unversioned corpus; fork/replay/diff/promote dissolve the stuck "which purpose anchors V1" decision. Today's L1 replay is a live demonstration that replay-from-immutable-origin actually works, which is the same move.

Other horizons, ranked.

• Load-bearing, Seb's call: getChainsContainingSeq is a json_each membership scan over every chain (4.0 s/call at 813k) — unindexable as written; needs event_seqs normalised. The rebuild buys headroom, not immunity; the complexity class is unchanged.
• Load-bearing, ours: governance-drift-check.py 3-of-5 inert — owed to PENDING as [HARDENING], not yet filed. The ladder's unrecordable firing history is the same shape one level up.
• Open, unfiled: 4,648 empty Lance index directories (~40/day since April); #176's countUnprotectedEntries confirmed in source.
• Parked deliberately: CONTROL-A v2 — inherited this morning and never touched; four defects named, I3 attacks the conclusion at its root. Still held on distance, and the distance is now real.
• Untouched: PENDING-89/90/91, the false-positive rate (unmeasured after four attempts), skill-harvest register compaction, ARC.

PAUSE STATEMENT: I am about to be away and a 30-hour replay is running without me. Nothing is half-finished — four repos clean or carrying only pre-existing files, three commits pushed, the instance backed up before any change was made. What I want to find still pulling is the replay's outcome, because it is the one thing here that cannot be reasoned to and must simply be waited for. The failure mode to guard against on return is the one this session demonstrated in its own centre: knowing the name of a failure class does not stop you walking into it — I wrote the census about inert instruments and then shipped a 27× wrong number from an uncontrolled grep six hours later.