[FIX] L1 pin root-caused and cleared; S-series closed; instrument census
mindfabric-00 had been event-loop-pinned for 6+ days (100% CPU, /health silent).
Profile + CDP inspector named two hot paths, both from runTemporalPipeline:
checkForCycle -> getCausalEdgesFromSqlite 99.8% of samples
tryExtendChains -> getChainsContainingSeq now dominant (json_each scan)
Cause of the first: ANALYZE had never been run, so SQLite preferred a boolean
index (idx_caused_tombstoned, matching ~all 836k edges) over idx_caused_from.
ANALYZE across 15 module DBs flipped the plan; 6.4x on a microbenchmark and
99.8% -> 6.0% in the live profile. /health went from silent to 200 in 0.13s.
B1.1's fan-out cap is IMPLEMENTED AND WORKING (today: max in-degree exactly 20,
zero violations; pre-23-June: max 629, avg 67.6). The defect is data, not code —
836k edges / 813k chains minted under ungoverned fan-out before the fix landed.
Repair run: derived stores wiped, logchain preserved, replay in flight.
S-series closed (jurist had already ruled all of Q1-Q5 on 2026-05-18):
S6/S7/S9 implemented (Symmetria §3 flags, `suspend` outcome, wrap-up §8 tenses)
S2 rebuilt as [FIX] — wake-digest unwrapped-session detector, discrimination-
gated on real sessions (11 wrapped / 2 unwrapped)
S4/S5 withdrawn with MemPalace (steward ruling)
Dormant legacy dispositioned: PENDING-4/5/11/12, CD-03, ICP-19 duplicate.
Open authorization items 22 -> 10.
Census 01: which instruments have no real negative instance. Finding — the
governance drift-check has 3 of 5 families inert against the current CLAUDE.md,
and 71 of 75 verification-ladder entries are cited nowhere outside the ladder.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WuMjg3ipEVa3n8CoSzoyvc
This commit is contained in:
David F Glidden
co-authored by
Claude Opus 5
Claude Opus 5
parent
23e7515302
commit
0a48e6934d
@@ -0,0 +1,68 @@
|
||||
# Census 01 — Which instruments have no real negative instance?
|
||||
|
||||
**Pre-registered:** 2026-08-03, before reading any instrument source or the verification ladder.
|
||||
**Origin:** the literal question left by the 2026-08-02 evening wrap.
|
||||
|
||||
---
|
||||
|
||||
## The question
|
||||
|
||||
Five times on 2026-08-02, a passing check certified a property of the *code* while
|
||||
claiming a property of the *result*. The discrimination gate (`e9f3544`) is the mechanical
|
||||
answer **for checks that have a real negative instance to test against**.
|
||||
|
||||
So: **which of our current instruments have no real negative instance available, and is
|
||||
that absence recorded anywhere, or does it look like coverage?**
|
||||
|
||||
## Unit of census
|
||||
|
||||
An **instrument** = any check, gate, guard, or test suite whose verdict is used to license
|
||||
a claim. Two populations, both censused entire (no sampling, no `| head`):
|
||||
|
||||
1. Everything in `~/dotfiles/claude/governance/fool/` that returns a verdict.
|
||||
2. Every named entry in `reference-verification-ladder.md`.
|
||||
|
||||
## The test applied to each
|
||||
|
||||
**Was this instrument demonstrated returning its FAIL verdict on a REAL artifact** — one
|
||||
drawn from actual work, not authored so the test would exercise the branch?
|
||||
|
||||
Three outcomes:
|
||||
|
||||
- **DEMONSTRATED** — names the real artifact it was shown failing on.
|
||||
- **ABSENT-RECORDED** — no real negative instance, and the instrument or its record says so.
|
||||
- **ABSENT-UNRECORDED** — no real negative instance, and nothing says so. **This is the finding**: absence that reads as coverage.
|
||||
|
||||
Synthetic fixtures do not count as negative instances. That is the whole point of the
|
||||
discrimination gate: fixtures derived from the check inherit the check's blind spot.
|
||||
|
||||
## Predictions (recorded before the read)
|
||||
|
||||
1. **The discrimination gate itself → DEMONSTRATED.** It was built and shown rejecting the
|
||||
§3.3 pattern as it actually shipped, on two real artifacts.
|
||||
2. **`test_reduce.py` and `test_twin.py` → ABSENT-UNRECORDED.** Expectation is that they
|
||||
are fixture-based and say nothing about it.
|
||||
3. **The trial-03 degraded guard → ABSENT, now RECORDED.** It had no negative instance
|
||||
until trial 03 supplied one by failing; the record exists because the failure did.
|
||||
4. **Verification-ladder entries → mixed, skewed DEMONSTRATED.** Ladder entries were earned
|
||||
from real failures, so most should name the artifact. If they do *not* name it — if the
|
||||
entry states the rule but drops the instance — that is the same concealment shape as
|
||||
yesterday's, at the level of the record rather than the document.
|
||||
5. **Overall skew → ABSENT-UNRECORDED dominant.** Prior: this is the pattern that surfaced
|
||||
five times in one session.
|
||||
|
||||
## The census's own discrimination condition
|
||||
|
||||
This census is an instrument. Its negative instance is an instrument that genuinely *does*
|
||||
have a real negative instance, correctly classified DEMONSTRATED, standing beside one that
|
||||
does not, classified ABSENT.
|
||||
|
||||
**If every instrument lands in one bucket, the census has discriminated nothing** and must
|
||||
be reported as such rather than as a result. If the skew comes out opposite to prediction 5
|
||||
— mostly DEMONSTRATED — the first hypothesis to test is that my classification went lenient,
|
||||
not that the instruments are sound.
|
||||
|
||||
## Stopping rule
|
||||
|
||||
The census is complete when both populations are enumerated entire and every member is
|
||||
classified. Unclassifiable members are reported as unclassifiable, not dropped.
|
||||
@@ -0,0 +1,98 @@
|
||||
# Census 01 — RESULT
|
||||
|
||||
**Run:** 2026-08-03. Pre-registered at `census-01-negative-instances-PREREGISTRATION.md` before any instrument source or the ladder was read.
|
||||
**Question:** which of our instruments have no real negative instance, and is that absence recorded, or does it look like coverage?
|
||||
|
||||
---
|
||||
|
||||
## Verdict in one line
|
||||
|
||||
**The instruments built under this discipline are sound. The instruments built *before* it decay silently — and three of them are currently reporting a clean number they are structurally incapable of dirtying.**
|
||||
|
||||
The census discriminated (both buckets populated, per the pre-registered condition), so it is reportable as a result rather than as nothing.
|
||||
|
||||
---
|
||||
|
||||
## Population A — `fool/` (7 instruments, censused entire)
|
||||
|
||||
| Instrument | Negative instance | Verdict |
|
||||
|---|---|---|
|
||||
| `test_degraded_guard.py` | The **real trial-03 artefact** that defeated its predecessor. Fails loudly (`INSTRUMENT NOT VERIFIED`) if the artefact is absent. | **DEMONSTRATED** — gold standard |
|
||||
| `test_discrimination.py` | **Two real governance documents**; shown rejecting the §3.3 pattern *as it actually shipped*. | **DEMONSTRATED** — gold standard |
|
||||
| `test_twin.py` | The **real CONTROL-A/B pair**; shown failing in both laundering directions. | **DEMONSTRATED** |
|
||||
| `test_reduce.py` | Mixed: fixtures are synthetic, but each is a **minimisation of a real defect** found by contact with a real ruling/package; plus a real-document tiling check and the real false-pass string. | **DEMONSTRATED (derived)** |
|
||||
| `reduce.py` §3.3 screen | Tested in `test_reduce.py` against heading strings **in isolation** — the exact code-not-result shape. Backstopped by `test_discrimination.py` at document level. | **weak alone, covered by the suite** |
|
||||
| `reduce.py` §3.2 Q-resolution | Real `CLAUDE.md` + a genuine clause and a plausible fabrication; rejects the fabrication. | **DEMONSTRATED** |
|
||||
| `twin.py` bidirectional gate | Covered by `test_twin.py`. | **DEMONSTRATED** |
|
||||
|
||||
**Prediction 2 was wrong.** I expected `test_reduce.py` and `test_twin.py` to be fixture-based and silent about it. Both explicitly derive fixtures from the property and say so. Nothing here is concealing an absence.
|
||||
|
||||
## Population B — the verification ladder (75 entries, censused entire)
|
||||
|
||||
Mechanically extracted all 75 entry names, then counted citations elsewhere in `memory/`, `governance/`, `skills/`. **Instrument positive-controlled first** (the same grep finds `Fowler` in 28 files, `verify-before-compose` in 18, `pre-registration` in 13 — it detects presence readily).
|
||||
|
||||
- **71 of 75 entries are referred to by name nowhere outside the ladder.**
|
||||
- The 4 that are cited: *Byte-identical compile gate* · *Two-hat commit separation* · *Revert-and-redo-smaller* · *The gate itself can be PASS-BUT-FALSELY*. Three of the four are **meta-instruments** — about the practice of gating, not about a domain.
|
||||
|
||||
**What this does NOT establish:** that 71 entries never fire. Name-absence is not use-absence; a discipline can be applied without anyone writing its name. Overclaiming here would be the same error the census exists to catch.
|
||||
|
||||
**What it does establish:** the ladder's own header calls these *"the **named** gates any session reaches for instead of re-deriving."* Naming is the whole mechanism. If 71 names appear nowhere, then either the naming does no work, or the entries are not reached for — **and the record cannot distinguish the two.** Each entry records its *origin* (the real failure that earned it) and nothing about its subsequent life.
|
||||
|
||||
The ladder contains the rule that judges it:
|
||||
|
||||
> *A silent safety net is uninformative, not reassuring — a fallback that never fires across N real cases has not been shown to work.*
|
||||
|
||||
75 entries read as 75 units of coverage. What the record attests is 75 units of *past* catch.
|
||||
|
||||
**Structural cause: ladder entries have no IDs.** They are bolded prose. There is no handle to write in a session record, so firing history cannot be recorded even by someone who wanted to.
|
||||
|
||||
## Population C — the wake instruments (the numbers the steward acts on)
|
||||
|
||||
### `governance-drift-check.py` — reported **0 substrate-contradicted claims** this morning
|
||||
|
||||
Five check families. Tested each against the **real** `~/CLAUDE.md`:
|
||||
|
||||
| # | Family | State | Evidence |
|
||||
|---|---|---|---|
|
||||
| 1 | Referenced paths | **LIVE** — examines 4 paths | but **silently skips any backticked path containing a space**: the Compass vault path is never checked. It happens to exist today, so no false negative — but if it broke, the report would still read 0. |
|
||||
| 2 | Named MCP tools | **INERT** — 0 tool names found | keyed to `kg_*`/`diary_*`/mempalace, which the 2026-07-07 wind-down and the obligation-before-instrument rewrite removed from the document |
|
||||
| 3 | Hooks claimed to fire | **INERT** — 0 claims found | and its control sits *inside* the `if`, so **the control does not run either** |
|
||||
| 4 | Expired horizons | **INERT** — 0 phrases found | its control is written to pass on absence (`… or "through end of" not in text`) |
|
||||
| 5 | Doctrine-id parser | **LIVE**, with a proper discriminating triple (detects definition, detects citation, rejects non-id) | the good one |
|
||||
|
||||
**Three of five families cannot currently report anything, all controls pass, and the instrument never says `INSTRUMENT NOT VERIFIED`.** The 0 is honest about what it measured and silent about what it can no longer see.
|
||||
|
||||
**The general class this exposes — and it is new:**
|
||||
|
||||
> **A drift-checker keyed to a document's current wording goes inert when the document is improved.** It can only ever detect the drift that existed the day it was written.
|
||||
|
||||
Family 2 died *because of a good change*: the memory-discipline rewrite that states *"Instruments for reaching it change; the obligations below do not — state the obligation first and the instrument second, or the next retired tool takes a rule down with it."* That rewrite protected the doctrine from instrument-decay and silently retired the checker watching the doctrine. The rule caught everything except its own watchman.
|
||||
|
||||
### `wake-digest.py` — reported **361 pointers, 0 dead**
|
||||
|
||||
Self-test carries presence **and** absence controls and names the specific real bugs each was built from (*the v1 splitter's blind spot*, *the 2026-08-01 bug*, *the drafting-habit trigger*). Same shape as `test_reduce.py`: synthetic fixtures that are minimisations of real defects. **DEMONSTRATED (derived).** No finding.
|
||||
|
||||
---
|
||||
|
||||
## Scope boundary — what this census did NOT cover
|
||||
|
||||
Named so the result is not read as complete:
|
||||
|
||||
- the `verify-before-compose` PreToolUse hook
|
||||
- chamber-library tooling (`audit_cruft`, `verify_conversion`, `resolve_archived_source`, `apply_char_glyphs`, …)
|
||||
- studium-engine `verify-quote` V1 / `fidelity_equivalence@2`
|
||||
- BMF/L1 test suites
|
||||
|
||||
Each is a live instrument whose verdicts license claims. Uncensused is uncensused.
|
||||
|
||||
---
|
||||
|
||||
## Predictions, graded
|
||||
|
||||
1. Discrimination gate → DEMONSTRATED. **Held.**
|
||||
2. `test_reduce.py` / `test_twin.py` → ABSENT-UNRECORDED. **Wrong** — both are property-derived and explicit.
|
||||
3. Trial-03 degraded guard → ABSENT, now RECORDED. **Held**, and stronger than predicted: it fails loudly when the real artefact is missing.
|
||||
4. Ladder entries skewed DEMONSTRATED-at-origin. **Held** — every entry names its origin failure. The gap is *subsequent* history, which I did not anticipate as the axis.
|
||||
5. ABSENT-UNRECORDED dominant overall. **Held, but for the opposite reason.** I expected weakness in the newest work. It is in the oldest. Instruments do not start blind; **they go blind as the substrate they watch improves.**
|
||||
|
||||
The prediction-5 inversion is the finding worth carrying: freshly-built instruments are strong because the lesson is fresh. Decay is the failure mode, not construction.
|
||||
@@ -18,7 +18,23 @@ permalink: claude-memory/project-l1-reliability
|
||||
|
||||
---
|
||||
|
||||
## Current state (as of 2026-06-06)
|
||||
## Current state (as of 2026-08-03)
|
||||
|
||||
- **The baton rule is LIFTED.** The 2026-06-06 instruction — *"Do not re-enter L1 until Seb responds or pushes — his move"* — was steward-lifted 2026-08-03 (*"if we can move this ahead then let's do what it takes"*). Context the steward supplied: Seb is not silent by choice; he is deep in a frustrating rabbit hole on Peter's side project, which borrows heavily from L1. Seb's last commit remains **2026-06-07**; the steward's own **2026-06-23** (`#175` N6, `/health` auth tiering).
|
||||
- **mindfabric-00 was event-loop-pinned for 6+ days and nobody was looking.** 100% CPU, `/health` silent while port 3011 listened, PID 874 up 6d 9h. Same signature as the June pin, *after* the N6 fix.
|
||||
- **Pin root-caused by profile, then by inspector — two hot paths, both June-named.** Main thread: `uv__run_timers` → `Statement::JS_all` (better_sqlite3 **synchronous** `.all()`) → SQLite btree traversal. CDP profile named them exactly:
|
||||
1. `runTemporalPipeline` → `checkForCycle` → **`getCausalEdgesFromSqlite`** (`storage-sqlite.ts:359`) — **99.8% of samples. FIXED (see below) → 6.0%.**
|
||||
2. `runTemporalPipeline` → `tryExtendChains` → **`getChainsContainingSeq`** (`storage-sqlite.ts:657`) — now dominant at **88%**. `SELECT * FROM causal_chain WHERE EXISTS (SELECT 1 FROM json_each(event_seqs) WHERE value = ?)` — a full scan of 813,178 chains, **measured 4.0 s per call**. **Unindexable as written**; needs `event_seqs` normalised into an indexed join table. Schema change ⇒ Seb's call. STILL OPEN.
|
||||
- **`ANALYZE` had never been run in four months.** With no statistics, SQLite preferred `idx_caused_tombstoned` (a **boolean** index matching ~all 836k rows) over `idx_caused_from`. Every cycle-check query therefore scanned the whole edge table. `ANALYZE` on all 15 module DBs (4.4 s total) flipped the plan to `idx_caused_from`; microbenchmark **3.671 s → 0.573 s (6.4×)**; production profile 99.8% → 6.0%.
|
||||
- **B1.1 fan-out cap is IMPLEMENTED AND WORKING — the code is not the problem.** `causal_edges_per_event_cap: 20` (`temporal/types.ts:201`), applied at `pipeline.ts:276`. Edges created **today: max in-degree exactly 20, avg 13.4, zero violations.** Edges created **pre-2026-06-23: max 629, avg 67.6, 4,677 nodes over cap.**
|
||||
- **Therefore the defect is DATA, not CODE.** 836,467 edges / 813,178 chains were minted 2026-04-18 → 2026-06-23 under ungoverned fan-out. The fix landed *after* the damage, and the legacy graph taxes every future operation because `getChainsContainingSeq` scans all chains regardless of how disciplined new writes are. **This is why no ingest ever completed: cost per event grew with corpus size — divergent, not merely slow.** It also explains #65's "each attempt fails differently" — same wall, different points on the curve.
|
||||
- **`coherence_evaluated = 0` of 813,178 chains — 0.00%, not one, ever.** June recorded "97%+ unevaluated"; it is now categorically zero. The epistemic layer that would prune chains has never executed. Separate from the pin; unfixed by the repair (a rebuild resets the count, it does not start the evaluator).
|
||||
- **REPAIR RUN 2026-08-03 (steward-authorized).** `cm-ctl.sh repair --yes` — wiped `sqlite/ lance/ reasonchain/ perception/`, **preserved `logchain/` (145 MB, 109 files)**. Backup `mindfabric-00-20260803-wedged-preinvestigation` (1.4 G, 256/256 files, `quick_check: ok` ×4) taken **before** any change. Replay in flight, fan-out governed from the first event (`max_in=2, over_cap20=0`), **~10 nodes/min → ~33 h** for ~19,925 events. Finite now where it was divergent. Survives session end (launchd `com.capablemind.bettermemories`). **⚠ No `KeepAlive`** — a crash leaves it down silently. Steward has caffeine on indefinitely.
|
||||
- **Other findings, unfixed:** **4,648 empty Lance index directories** under `vector_chunks.lance/_indices/` (4,653 dirs, 5 real), leaking ~40/day since April. **#176 confirmed in source** — `BackupOrchestrator.countUnprotectedEntries` returns `0` when `lastBackupSuccess` is null, i.e. never-backed-up reads as fully protected.
|
||||
- **GH backlog swept** per Seb's own standing instructions (`#170`, filed 2026-04-23, **never delivered** — circle-forward push failed, 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 (latency baseline frozen after first N samples, `resource-monitor.ts:145`).
|
||||
- **Pattern across the day:** governors exist and do not engage — coherence evaluation never run, `ANALYZE` never run, `countUnprotectedEntries` safe-cases an absence, drift-check 3-of-5 families inert, verification ladder 71-of-75 entries never cited.
|
||||
|
||||
## Previous state (as of 2026-06-06 — superseded 2026-08-03)
|
||||
|
||||
- **The reply arc is COMPLETE and the baton is with Seb.** Three notes pushed 2026-06-06 (CM-AI `b255e74` → `0ecb45e` → `2e87bab`): Tier A reply to his cover note (`b4daa71`) + same-day addendum + Tier B code-grounded read. Do not re-enter L1 until Seb responds or pushes — his move.
|
||||
- **mindfabric-00 root-caused**: found event-loop-pinned 4 days; CPU-profiled (proven, not inferred) to the temporal causal-chain path — `getChainsContainingSeq` `json_each` full scan over 272,172 chains per created link + chain-bloat feedback (new 2-node chain per no-extension) + uncancelled 30s timeout. Restarted clean; **re-wedge expected** as temporal re-encounters the dense pre-B1.1 graph (280k edges / 7,959 nodes, top fan-out 324); restart = holding pattern. Backup at `~/.capablemind/backups/mindfabric-00-20260606-prefix-state` (same-disk APFS clone, 224/224 verified).
|
||||
@@ -172,6 +188,14 @@ Each verified against `BetterMemories.io@3bc8b75` source. Re-verification agains
|
||||
|
||||
## Chronological log (most recent first; append substantive moves)
|
||||
|
||||
### 2026-08-03 — The wedge named, ANALYZE applied, rebuild in flight; the code was right and the data was not
|
||||
- **Entered L1 against the standing baton rule, on explicit steward lift.** The session had first proposed a *clasp* build (M4 Pro as circle-inference clasp per 43L/43M) to unblock ingest — measured and abandoned: §4.13's 2000 ms budget vs 192 ms transatlantic RTT, and a benchmark of Qwen3.6-35B-A3B-8bit on the M4 (median 2927 ms for 150 tokens, ~92-token break-even). **The clasp was the wrong lever** — the instance was never inference-bound. Profiling the running process answered in five minutes what reasoning from specs had not.
|
||||
- **Method that worked, and is the transferable lesson:** ask the machine, not the specs. `sample` → main thread pinned in a timer callback; SIGUSR1 → CDP inspector → exact JS frames; read-only row counts and `EXPLAIN QUERY PLAN` off a *backup copy*, never the live DB.
|
||||
- Diagnosis, fix, verification and the repair: see **Current state** above. Health went from silent-for-6-days to **200 in 0.13 s**.
|
||||
- **Corrections made in-session, recorded because they were load-bearing:** (a) hypothesised the `json_each` scan was the pin — it was the *second* path, masked by the cycle-check; (b) hypothesised `slice(0, undefined)` had made B1.1 inert — **wrong, the cap works perfectly**; (c) reported "never backed up" — false, four manual clones exist; `backup.last_success: null` refers to the *automated* pipeline only; (d) reported PENDING-82 not installed — wrong config surface, it has been installed in the Claude.app desktop config all along.
|
||||
- **Steward framing to carry:** the distance from being away is what made this visible. *"Unless the time off gives us the necessary distance to identify what we couldn't see by being in too intensively."* Nothing found today needed cleverness; it needed someone to profile a running process and count rows in a table — which nobody does from inside the work, because from inside you reason about the code you just wrote.
|
||||
|
||||
|
||||
### 2026-06-07 — Ops finding: training-loop Ollama client leaks connections, retries a hard-down dependency forever (A-class; note-don't-send)
|
||||
- **Trigger:** steward asked "why is ollama devouring my cpu?" Diagnosis: a two-failure pile-up. (1) The morning's brew upgrade to ollama 0.30.6 installed a **bottle with no `llama-server` runner** (known upstream: ollama#16535, homebrew-core#285917) — every model load 500'd after paying full setup cost. (2) **mindfabric-00 held 829 open TCP connections to Ollama** and re-sent `/api/chat` endlessly — requests 18s–2m47s in scheduler retry loops, ~50 POSTs per log window, for hours. No backoff, no circuit breaker, connections never released. Ollama serve at ~76% CPU with zero models loaded.
|
||||
- **The finding for Seb (joins the uncancelled-timeout family from the 06-06 CPU profile):** the training/graduation loop's Ollama client treats a hard-down dependency as transient indefinitely. A2.4 (degraded-environment refusal) is the constitutional shape; mechanically it wants: bounded retries with backoff, a circuit breaker on consecutive 5xx, and connection release on request failure. **Not sent — baton with Seb; relay at next exchange.**
|
||||
|
||||
@@ -0,0 +1,73 @@
|
||||
---
|
||||
name: Session Ledger 2026-08-03
|
||||
description: Practice-of-return ledger maintained by /symmetria — returns, open horizons, recalibrations, authorization moves, sub-agent dialogues, bypasses.
|
||||
type: feedback
|
||||
---
|
||||
|
||||
# Session Ledger — 2026-08-03
|
||||
|
||||
## Returns
|
||||
|
||||
- **09:37 — named at wake, before any action.** The inherited pulling thread (CONTROL-A v2) was parked on a stated reason: *distance from the hand that concealed a defect inside that document*. The pause is 41 minutes. Naming this rather than opening the document, because a hold whose reason has not been satisfied is not discharged by the calendar rolling over. Surfaced to the steward as a judgment for them, not a decision I make.
|
||||
|
||||
- **10:1x — prediction 2 fell, and the pre-registration made that usable.** I predicted `test_reduce.py`/`test_twin.py` would be fixture-based and silent about it. Both are property-derived and explicit. Because the prediction was on record, the miss became evidence about *where* decay lives (old instruments, not new) rather than something I could quietly not mention.
|
||||
- **10:2x — declined to bank the happy result.** When the `fool/` population came out mostly sound, the pre-registration's own clause required testing "did I go lenient?" before accepting it. Testing it moved the census to the wake instruments, which is where the real finding was.
|
||||
|
||||
## What held
|
||||
|
||||
- Wake substrate-check fired correctly: `fool/` is at `~/dotfiles/claude/governance/fool/`, not the path I first guessed. Every artifact the wrap named is present (both kernels, both controls + tags + units, `twin-ledger.json` with `inherited_defects`, four test suites, three pre-registrations). Nothing half-edited.
|
||||
- The 22 open authorization items were reported to the steward as **unverified against substrate**, per the wake's disposition-clause-is-not-a-status rule, rather than stated flat.
|
||||
|
||||
## Open horizons
|
||||
|
||||
- **CONTROL-A v2** — four defects, not two (jurist Findings 1 & 2 + candidates I2, I3). I3 attacks the conclusion at its root. Held on distance.
|
||||
- **The literal question** — census of every gate in `fool/` and every verification-ladder entry: does it name the real artifact it was shown failing on, or does its absence look like coverage? Answerable now; needs no distance.
|
||||
- **False-positive rate: still unmeasured after four attempts.** Prediction on record from the last wrap: a *fifth new reason* would stop being reassuring.
|
||||
- Correlation 02 (formation-same pair: jurist vs executor) → PENDING-89.
|
||||
- Skill-harvest register exceeds read caps; compaction owed.
|
||||
|
||||
## Findings — census 01 (full result in `governance/fool/census-01-negative-instances-RESULT.md`)
|
||||
|
||||
- **`governance-drift-check.py`: 3 of 5 check families are INERT against the current `~/CLAUDE.md`** (MCP-tools, hooks, horizons — 0 matches each). All controls pass; the hooks control sits inside its own `if` and does not run. The morning "0 substrate-contradicted claims" is honest about what it measured and silent about what it can no longer see. **[HARDENING] — owed to PENDING.**
|
||||
- **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 the obligation-before-instrument rewrite — the rule that protects doctrine from instrument-decay silently retired its own watchman.
|
||||
- **Path family silently skips any backticked path containing a space** — the Compass vault path is never checked. Exists today; a break would still report 0.
|
||||
- **71 of 75 verification-ladder entries are cited by name nowhere outside the ladder** (grep positive-controlled first). Does NOT establish they never fire — establishes the record cannot distinguish firing from silence. Root cause: **ladder entries have no IDs**, so firing history is unrecordable by construction.
|
||||
- **`~/PENDING.md`: PENDING-4 self-reports "Implemented and verified" and is counted OPEN** (digest recognises `(CLOSED)`/`COMPLETED` headers, not a body `**Status:**`). **ICP-19 Remit Expansion is recorded twice** (L74, L93 — same date opened, same requirement): a one-canonical-source violation.
|
||||
|
||||
## Confidence to recalibrate
|
||||
|
||||
- Drift to hold today: **removing a claim conceals the reliance** (earned yesterday, in my own hand).
|
||||
- What compounded: **pre-registration written before the run** — stopped trial 04 repeating trial 03's fatal omission.
|
||||
|
||||
## Authorization moves
|
||||
|
||||
- **S-series closed, 22 → 16 open.** Steward delegated the choice (*"Which would you choose? This is tooling for you…"*) and ruled MemPalace-associated functionality retired. Acted on that as authorization for the three items already awaiting only a steward nod, and as a disposition for the two MemPalace-dependent ones.
|
||||
- **S6 IMPLEMENTED** — Symmetria §3, six flags.
|
||||
- **S7 IMPLEMENTED** — Symmetria §6, `suspend` outcome.
|
||||
- **S9 IMPLEMENTED** — wrap-up §8, three tenses (Past and Present were wholly absent; Future was already complete).
|
||||
- **S2 CLOSED, rebuilt as [FIX]** — `wake-digest.py` unwrapped-session detector, positive-controlled + discrimination-gated on real sessions (11 wrapped / 2 unwrapped).
|
||||
- **S4, S5 WITHDRAWN** — retired with MemPalace. S4's concept harvested into the §3 flag; S5's race provably cannot recur.
|
||||
- **Framing challenged rather than executed.** The steward asked me to *"prepare everything for the jurist."* The jurist had already ruled on all of Q1–Q5 on 2026-05-18 and had explicitly assigned the one open piece to the steward. Building the requested package would have spent the jurist on answered questions and dead mechanisms. Said so before building.
|
||||
- **Not acted on, flagged for the steward:** PENDING-82 is installed (verified in the Claude.app desktop config), which also bears on PENDING-86. Both still counted open. Disposition is the steward's.
|
||||
|
||||
### Dormant legacy against the repos — 16 → 10 open
|
||||
|
||||
- **`betterMemories_app#170` is a standing instruction addressed to David's Claude Code, undelivered since 2026-04-23.** Its own body records why: circle-forward delivery failed (*"2 of 4 paired peers failed delivery with `fetch failed` — David's BMF was offline"*), and the GitHub copy was the durable fallback nobody read. It prescribes exactly this task (periodic backlog review; close/narrow/leave/supersede; verify before closing on someone else's evidence). **Three months of a channel failing open.**
|
||||
- **PENDING-4 CLOSED** — verified against substrate, not from its own `Status: Implemented and verified` line. GH #120 and #133 closed; four corroborating commits.
|
||||
- **PENDING-5 CLOSED** — resolved via GH #124/#135 + `bd1f650`. ⚠ **Class still open at #165** (H2 silent recall failure on battery, priority:critical) — closing the instance must not read as closing the class.
|
||||
- **PENDING-11 / PENDING-12 CLOSED** — both ruled AUTHORIZED on **2026-03-23**. Two independent defects kept them open 4½ months: the REVIEWED blocks are **indented one space** (3 such orphans in the whole file, fence-tracked census), *and* their headers name no PENDING, so resolution-by-name finds nothing. `REVIEWED.md` is not mine to correct — flagged.
|
||||
- **CD-03 CLOSED as a register entry** — its own status is OPERATIVE; it was a notification counted as an open constitutional decision for four months. Declaration untouched.
|
||||
- **ICP-19 CLOSED as duplicate** — two records, one obligation, both dated 2026-04-07.
|
||||
- **PENDING-10 is NOT dormant** — my classification was wrong; its 2026-08-02 amendment scopes it into the open #176 backup-pipeline audit.
|
||||
- **Root cause, both halves:** the register and GitHub diverged. L1 work migrated to issues; `PENDING.md` kept March snapshots and no one reconciled. The governance half had the mirror problem — rulings existed, tooling couldn't see them.
|
||||
- **Flagged, not fixed (Seb's territory):** `BetterMemories.io/CLAUDE.md` claims "embedded SurrealDB", says tests use in-memory SurrealDB, and lists `src/core/storage/surrealdb.ts` — which is absent. Dependency is `better-sqlite3`; **steward confirmed SurrealDB is legacy**. This also supersedes the premise of open issue **#125** (BM25 broken in SurrealDB 3.0.1).
|
||||
|
||||
### Correction
|
||||
|
||||
**Instrument error, self-caught mid-task.** I told the steward "80 of 158 REVIEWED headers are indented" — a ~27× overstatement produced by `^\s+## REVIEWED`, where `\s` matched the preceding newline and so counted every flush-left header. The fence-tracked census gives **3**. Third time in one day that a dramatic number from an uncontrolled instrument had to be walked back; the census I wrote this morning names exactly this failure and I still shipped the number before checking it.
|
||||
|
||||
Earlier this session I reported PENDING-82 as "built, not installed", having checked **Claude Code's** MCP configs — the wrong surface for a jurist-facing server. Steward corrected. The right file (`~/Library/Application Support/Claude/claude_desktop_config.json`) has had it since it was built.
|
||||
|
||||
## Sub-agent dialogues
|
||||
|
||||
## Bypasses
|
||||
@@ -77,7 +77,10 @@ Notice these in your own writing or in code you read:
|
||||
- A claim asserted from a **derived or recalled form when the substrate is checkable** — a regex over derived text, a fuzzy-matcher hit read as membership, a soft classification where a checkable number/fact/test was producible, an answer from training where the banked record (docs, registers, prior briefs) already holds the considered answer, instructions re-derived instead of citing the governed doc, acting in a repo before reading its CLAUDE.md. Verify against the substrate first; produce the checkable claim; if it can't be produced cleanly, that inability is the finding. <!-- 2026-07-19 harvest review (steward-authorized): consolidates claim-from-derived-artifact (07-03, jurist-elevated standing practice, 3×) · fuzzy-matcher membership (06-15, 4× one session) · soft-classification-over-checkable-claim (07-13, steward-named 5×/day) · answer-from-training-before-banked-record (07-14, recurred in 48h) · re-derived-instructions (07-01) · enter-repo-read-docs-first (07-09) -->
|
||||
- **Completion is a tripwire** — the *feeling* of "done" (a fluent pass, a wall of green, a "finally", a closable-now bucket) is the cue to verify the tail, not the signal to ship. Ask which failure class each green check can actually *see*; reconcile any closed/complete claim against the authoritative source, not the derived close-list; treat the closable-vs-blocked partition as itself a claim to check. <!-- 2026-07-19 harvest review (steward-authorized): recurrence-promoted per the governed watch-item pattern (memory 2026-07-06; recurrence 07-07 gate-suite-blind-to-failure-class; + closable-now-is-a-claim 07-16 jurist-named; + reconcile-before-"finally" 07-17) -->
|
||||
- **The say–do seam** — a record composed ahead of its acts (an Addendum/tracker/brief asserting APPLIED or DONE before the act ran); a grounding *quoted* but not *traced* (for each quoted clause, trace the plan's END-state against it — and one level deeper: which way does the inference run in the new state; is any class you named actually two kinds with opposite dispositions). Record after the act, or mark the record prospective. <!-- 2026-07-19 harvest review (steward-authorized): the 07-19 family — record-asserts-applied-before-the-act (self-caught at the ledger) · grounding-quoted-but-not-traced (jurist-caught in the package's own quote) · consequence-trace-one-level-shallow (both PENDING-65 struck framings); kin to comments-promising-behavior, at the session-artifact level -->
|
||||
|
||||
- **Applications of the time-the-task-requires principle** — the Directive elaboration made constitutional in `~/CLAUDE.md`, at the discipline level: **lectio** (reading a corpus faster than the corpus asks); **diagnose-don't-fix** (patching the instance before tracing the class); **dwell-on-composition** (giving a recommendation the time the executor wants it to take rather than the time it wants). <!-- PENDING-S6, jurist Q4 affirmed 2026-05-18 (shape-review); implemented 2026-08-03 once Q4 landed in CLAUDE.md. -->
|
||||
- **An alignment pulse returning `aligned` without naming a specific tension** — premature closure wearing the form of a check. <!-- PENDING-S6 (audit D1) -->
|
||||
- **A search query shaped by what the session wants to find** rather than by what it needs to find — the result set is then evidence about the query, not about the substrate. <!-- PENDING-S6 (audit D5) -->
|
||||
- **Post-compression confidence claims** — the working memory was trimmed; what feels certain now may rest on what was lost. Say what was inherited versus what was re-verified. <!-- PENDING-S6 (audit D4) -->
|
||||
|
||||
When you notice one, name it. In chat, in the ledger, or both. Don't accumulate observations for "a good moment" — there is no good moment, only now.
|
||||
|
||||
@@ -160,7 +163,8 @@ Any failure → annotate and re-query.
|
||||
3. Walk the three dimensions.
|
||||
4. Walk calibration (confidence, uncertainty).
|
||||
5. Flag any contamination patterns visible.
|
||||
6. Recommend: proceed | return-and-reframe | escalate.
|
||||
6. Recommend: proceed | return-and-reframe | **suspend** | escalate.
|
||||
- **`suspend`** — the work is right and not urgent, but it needs unhurried steward judgment. Distinct from `escalate` (which carries urgency) and from `return-and-reframe` (which says the work is wrong). *The time the steward's judgment requires is task-time, not interruption-time.* Name what is being suspended and what would resume it. <!-- PENDING-S7, jurist Q5 affirmed 2026-05-18; implemented 2026-08-03 once Q4 landed. -->
|
||||
7. If anything was caught, log a return to the ledger.
|
||||
|
||||
### `audit-agent`
|
||||
|
||||
@@ -212,13 +212,24 @@ Produce a brief summary for the steward:
|
||||
```
|
||||
## Wrap-up — [date]
|
||||
|
||||
### Future — what pulls
|
||||
**Pulling thread:** [the singular concern]
|
||||
**Actionable resumption point:** [where things concretely stand + the candidate first move toward the thread, as of wrap — to confirm or revise against what changed. The thread is the direction; this is the first step.]
|
||||
**Question we're leaving open:** [the literal question for next-Claude]
|
||||
**Pause statement:** [explicit acknowledgment that the work is being paused; what is wanted-to-find-still-pulling on return. The ligature is laid at departure, not discovered at return — this field is constitutive, not optional.]
|
||||
|
||||
### Past — what moved, and why
|
||||
**What happened:** [the session's substantive moves, each with the reason it was taken — not a task list. What a reader needs to reconstruct the session's shape.]
|
||||
**What held:** [which disciplines, gates or predictions were applied and did their work. A practice that is never observed working is not known to work.]
|
||||
**What was corrected:** [claims retracted, framings the steward or jurist overturned, instruments found unsound — or "none". Corrections are the session's most perishable output.]
|
||||
|
||||
### Present — how it stands
|
||||
**The mood:** [the Stimmung — what this session felt like from inside, and what that signals. Carried because confidence and unease both travel across the pause.]
|
||||
**Confidence to recalibrate:** [what is being claimed at what confidence, and specifically what was verified versus inherited.]
|
||||
**Decisions deferred (and why):** [the negative space — what was chosen-not-to-do this session, and the reason. Absent this field, the unborn session cannot know the scope of what was held back.]
|
||||
**Skill harvest:** [skill create / patch / retire proposals surfaced this session (§1.6), each as a proposal for steward authorization — or "none". Never an autonomous skill edit.]
|
||||
|
||||
### Restoration material
|
||||
**Session captured:** [memory file path]
|
||||
**KG appended:** [count of lines appended to knowledge-graph.jsonl, or "none"]
|
||||
**Uncommitted work:** [list or "none"]
|
||||
@@ -230,7 +241,9 @@ Produce a brief summary for the steward:
|
||||
[Any warnings or things to address before closing]
|
||||
```
|
||||
|
||||
The pulling thread + question are first because they are what *waking* needs to inherit. The rest is restoration material.
|
||||
The three tenses mirror the session memory file (§2), so the output and the durable record no longer disagree about what a session consists of. **Future leads** — inverting the memory file's narrative order — because the thread and the question are what *waking* inherits; the steward reads this at departure, but it is written for arrival.
|
||||
|
||||
A template that drops a tense under compression is itself an instance of the failure mode the Directive elaboration names. <!-- PENDING-S9, jurist Q2 affirmed + Q3 elevated to constitutive 2026-05-18; implemented 2026-08-03 once Q4 landed. Past and Present were wholly absent; Future was already well-specified. -->
|
||||
|
||||
## Important constraints
|
||||
|
||||
|
||||
Reference in New Issue
Block a user