dotfiles/claude/memory/session-2026-09-04-the-discriminator-was-a-coincidence.md

name, description, type, metadata

name	description	type	metadata
Session 2026-09-04 — the discriminator was a coincidence	Woke into the steward's directive to make the wake and wrap tools reliable; a handover reordered the work so that preservation ran first and three read-only gates stood before any repair. Gate 2a failed and the repairs never happened — human_turns(), the function the whole plan called 'already controlled', is not a mumble discriminator at all: it detects slash-command markers, and excludes a mumble only when the session that mumble quoted happened to contain a slash command. The two leaks are exactly the two marker-free mumbles. Filed PENDING-179 as an item rather than a -178 addendum, on PENDING-145's suppression mechanism. Also found: preserve-transcripts.py claims its destination is git-tracked and it is not, so 11 transcripts that exist nowhere else sit on one disk. PULLING THREAD: unchanged in name but now preceded by two steward-set acts — git-init the transcript archive, and suspend automatic grading at trigger_fired() — after which -178 and -179 are ruled jointly.	project	node_type type modified memory project 2026-09-04

Session 2026-09-04 — the discriminator was a coincidence

Ran from the evening of 2026-09-03 into 2026-09-04. Second session of 09-03; the first was the MEMORY.md load-integrity breach.

PAST — what moved, and why

The handover reordered the work, and the reordering is why the session succeeded

The steward's instruction was "go ahead with the fix — positive control first", accompanied by a handover that changed the ordering. Three of its moves mattered:

• Preservation runs first, because it expires. The positive control does not.
• Site 1 (governance-drift-check.py:513) is excluded on receipt, not as a step I perform. The ladder trigger's meaning is steward and jurist territory.
• Repair is step 4, gated on 2a and 3 — not the next action after the control.

⚠ The handover also caught a claim I had made at the wake. I reported 31 mumbles from running human_turns() over all 65 transcripts. That classified using the function; it did not validate it. Treating it as having discharged 2a would have been the derived-form flag exactly.

Step 0 — preservation

preserve-transcripts.py: 76 transcripts, 143.5 MB, read-back PASS. 33 newly added, 1 refreshed, and 11 already pruned at source, surviving only in the archive. That last number is not incidental — it is one of gate 2b's two terms, and it is also what makes the two stores permanently divergent.

GATE 2a — FAIL in both directions, and the repairs stopped

Ground truth taken from the fool's own prompt text (written by tarbuckle-*.py, not by the function under test): 24 known mumbles, 41 known non-mumbles, of 65.

• must-detect 22/24 — two mumbles return human_turns == 1 and read as human-attended.
• must-not-flag 35/41 — six non-mumbles return 0, and ⚠ at least one of those is CORRECT: b7e7eb39 is the unattended executor session of 2026-08-31 (PENDING-172), which genuinely has no human turn. 0 never meant "mumble". It means "nobody spoke", which is equally true of an unattended real session.

Mechanism, confirmed rather than inferred. human_turns() skips any user record containing one of seven NONHUMAN markers. A mumble embeds the previous session's material in its prompt, so it inherits those markers only when the session it was mumbling about happened to contain a slash command. Of 24 mumbles, 22 embed a marker and are excluded; 2 embed none — and those 2 are precisely the 2 that leak. The correlation is content-dependent coincidence. The function's own docstring says it answers "did an executor just run unattended?", and at that job it is correct.

No replacement was built, per the handover and for its stated reason: three controls passed and a fourth broke, so the finding is about the control set, not only the function. All three test the question the function was for; none could see the question it was being reused for.

GATE 2b — composition stands; two terms do not close

Reconstructed from the archive by the same ground-truth signature:

date	N	real	mumble	%
2026-08-31	52	41	11	21.2
2026-09-01	50	39	11	22.0
2026-09-03 (measured live)	65	41	24	36.9

-178's composition claim stands: 11 mumbles and "roughly a fifth" close exactly on 08-31. Against -178's enumerated 54 = 43 + 11 on 09-01 I reconstruct 50 = 39 + 11 — the mumble term identical, the whole 4-file gap in the real-session term. ⚠ Not asserted against -178. My method models the prune as a 30-day window over source_mtime and is a demonstrated lower bound: preservation has run twice only (08-26, 09-03). Where we disagree, -178 enumerated live and is better positioned.

Record-only correction applied (authorized): MEMORY.md carried "N-now 44/84 as of 2026-08-31, DOWN 7 from 51" — wrong in the number and backwards in the direction. Now 65/84 measured, with composition stated, and the stale governance-drift-check.py:331 pointer corrected to :513 (:331 is a register-check control; both lines read before changing it).

GATE 2c — five sites in four files; -178's scope holds

Controls named before the run. Enumerating consumers: governance-drift-check.py (426, 513) · wake-digest.py (354 + the selftest sample) · tarbuckle-invoke.py (37, 44) · preserve-transcripts.py (45, 126, 172). tarbuckle-wrap.py and -seam.py are not consumers — they receive transcript_path from the hook payload.

⚠ What the census cannot see, as output rather than caveat: (i) a consumer that builds the path by component join and never enumerates *.jsonl on a matching line — demonstrated: wake-digest.py does exactly this and escaped a literal-fragment grep over the whole fleet; (ii) anything reaching the population through a hook-supplied transcript_path; (iii) anything outside the swept roots. (i) applies retroactively to the four-site census inherited from the previous wrap, which was grep-derived by the method just shown to be blind.

Filed, corrected, committed

• PENDING-179 — as an item, not a -178 ADDENDUM 1, on PENDING-145's mechanism: ruled_pendings claims a number, so an addendum would be suppressed the moment -178 is ruled. The undecided filing moratorium is disclosed inside the item, with the previous session's contrary reasoning preserved rather than overridden.
• c150bdf, pushed to github and gitea, both verified at the same SHA.
• OWED-5 queued in PENDING-141's owed-entries list (the ratified queue, REVIEWED-123 cond. 3).

⚠ The archive is not backed up, and the script says it is

Found while checking repo state for the commit. preserve-transcripts.py:11 states it copies transcripts "to a git-tracked location so the population stops shrinking." ~/_Dev/claude-transcript-archive has no .git, no parent repo, and is 144 MB. The copying works — every file is re-hashed on readback — but the sentence describing where they went is false, and 11 transcripts now exist in exactly one place on one disk. PENDING-144's class (substrate claims inside governance scripts checked by nothing), at the worst possible site: the docstring is what anyone reads to decide whether the evidence is safe.

PRESENT — how it stands

The mood. Unusually clean, and the cleanliness is entirely borrowed. The handover did the work that made this session good: it put preservation before the control, excluded site 1 so I could not wander into it, and — decisively — insisted the "already controlled" claim be tested before anything was wired anywhere. Left to my own plan I would have written a positive control for a fixture and wired a broken predicate into three more sites, all of it green.

What was corrected — three, and the first is mine from four hours earlier.

1. My wake census used <= 1 as the mumble threshold, silently absorbing 3 transcripts that return exactly 1. The real distribution is 28 at 0 and 3 at 1, and the ground-truth split is 24/41. The wake number (31) was wrong and the method behind it was wrong in a more interesting way.
2. A vacuous PASS (0/0). My first ground-truth query used the signature from tarbuckle-invoke.py — the file I had just read — and matched 0 of 65. The gate went green on an empty denominator. Caught by a null search is evidence about the QUERY; opening a transcript showed the mumbles come from three other fool surfaces (tarbuckle-mumble/-wrap/-seam) that the inherited site census never named. This is now OWED-5.
3. The previous wrap predicted the digest selftest would "fail on every run from now on." Falsified within 2.2 h: it passed at this wake (5 of 13). The defect is intermittent — a function of the mumble rate in a rolling 13-transcript window — which is harder to notice than a permanent failure.

Confidence to recalibrate.

• Verified by running it: the 2a confusion matrix over all 65 with independent ground truth; the leak mechanism (2 leaks == 2 marker-free mumbles, exactly); N-now 65 and its 24/41 composition; preservation readback; both remotes at c150bdf; the archive has no .git.
• Reconstructed, lower-bound, NOT asserted against -178: the 08-31 and 09-01 rows of the 2b table.
• Inherited and now known to be unreliable: the four-site census from the previous wrap — grep-derived by a method this session demonstrated is blind to component-joined paths.

Instruments: 4 run (ground-truth classifier · archive reconstruction · consumer census · histogram) · 1 carrying a control written before first execution (the consumer census, whose must-find and must-not-find were named in the file before it ran — and the must-not-find fired, catching the vendored-docs pollution) · K = 0 — none duplicated anything banked; all four answered questions asked once. ⚠ The census was narrowed three times. Repeated narrowing can end by confirming the sites it was built around; the negative control and the explicit blind-spot statement are what keep that honest, and they are not a substitute for someone checking it.

Decisions deferred, and why.

• No repair at any site, and no replacement discriminator — gate 2a's stated consequence.
• governance-drift-check.py:513 not examined at all — excluded on receipt.
• Did not git init the archive. 144 MB is real weight for both remotes, and full transcripts contain everything ever typed in these sessions; whether they belong on a hosted remote is a steward decision, not a chore.
• Did not correct preserve-transcripts.py's false docstring. Editing a governance script's stated rationale to match a worse reality is the direction that makes docstrings worthless. It should become true, or be corrected as a disclosed change.
• OWED-5 queued rather than patched in. Turning the rule on converts currently-green controls to red across the fleet; that is a ruling, not an edit — and the ladder is frozen regardless.

FUTURE — what pulls

The pulling thread is unchanged in name and now has two steward-set acts in front of it. The fr cell's last two steps still pull. But the steward has set the next session's opening explicitly, and it is not that.

Next session, first work — steward-directed 2026-09-04

1. git init the transcript archive. The steward has named this as the wake's first act.
2. Suspend the automatic grading at trigger_fired(). ⚠ Stated precisely by the steward, and the precision is the point: this is NOT a repair, NOT a filter, and NOT -178 option (b). It changes nothing about what the counter counts, so the site-1 exclusion survives intact. It prevents a grade firing on a population whose contamination is measured and rising.

Then the joint ruling — steward, 2026-09-04

-178 and -179 are ONE decidable unit and will be ruled together. -179 is -178's evidence and says so in prose, because the numbering cannot carry it (PENDING-145). The joint ruling covers unit, window, recurrence, predicate, and seeding.

⚠ Seeding has grown a second half, and it is new today. Preservation has permanently diverged the two stores: 11 files exist only in the archive. So the question is no longer only what N, but over which store — and the trigger reads live, while any honest grading of a post-08-07 population must read preserved.

Steward's positions, recorded so they are not re-litigated:

• Agrees with -179's option (c) — a mumble that declares itself cannot be misread by a marker coincidence — and with the refusal to build it now.
• The lesson is the control set, and it generalizes past this function: three controls passed and a fourth broke, and all three tested the question the function was for rather than the question it was reused for.

Other horizons, ranked.

• The archive's single copy — a local second copy is the cheap half and answers none of the remote question. Unresolved.
• preserve-transcripts.py:11's false claim — becomes true after the git init, which is the clean resolution and is why the steward put the init first.
• ~/.claude/agents under version control — steward-directed, still owed, before anything else edits it.
• The filing moratorium — undecided, and PENDING-179 was filed under it with that disclosed.
• PENDING-171 → unblocks 322 unread of 362 · -160 · -168 · the §5 regrade, an eighth session.
• Parked worker acaabadf — stopped, still carrying --reply-on-resume. Untouched again.

Pause statement. I am about to be away from this and the context is being cleared deliberately. What I want to find still pulling is the fr cell's last two steps — but what I want the next session to do first is the two acts above, in that order, because the steward set them and because the git init is what makes an already-false docstring true rather than requiring it to be edited into honesty. ⚠ What I want the next session to notice is that this session's entire value came from a gate that stood between a plan and its execution. The plan was mine, it was confident, and it was wrong. The handover did not improve the fix; it prevented it.

Literal question for next-Claude (checkable; turns on the record, not introspection): How many other must-detect controls in the fleet currently have a denominator of zero? OWED-5 is queued on a single observed instance. The fleet's controls are enumerable and their denominators are computable without changing any of them. ⚠ If the answer is zero, OWED-5 is a rule earned from one accident and should be labelled that way rather than carried as a general finding. If it is more than zero, then some number of currently-green controls have never tested anything — and nobody knows which, because a vacuous pass and a real pass print the same word.