[HARDENING] The mumble discriminator fails its own must-detect gate; repairs stopped at the gate (PENDING-179)

The 2026-09-03 repair plan rested on one claim: that human_turns() is an
existing, controlled discriminator, and wiring it into four transcript
consumer sites was plumbing rather than classifier-building.

Tested against ground truth taken from the fool's own prompt text — not from
the function under test — the claim is false. 24 known mumbles, 41 known
non-mumbles, out of 65 transcripts:

  must-detect    22/24  two mumbles read as human-attended
  must-not-flag  35/41  and one "failure" is correct — b7e7eb39 is the
                        unattended session of 2026-08-31, which genuinely
                        has no human turn

human_turns() == 0 never meant "mumble". It means "nobody spoke", which is
equally true of an unattended real session. A mumble embeds the previous
session's text and so inherits its slash-command markers; it is excluded
only when the session it quoted happened to contain one. The two leaks are
exactly the two marker-free mumbles. Coincidence, not design.

No repair was made at any site, and no replacement discriminator was built.
Three controls passed and a fourth broke: all three test the question the
function was built for, none could see the question it was being reused for.
A successor written now inherits whatever made the first set look sufficient.

Also here, per the 2026-09-03 handover:
  - step 0 preservation ran: 76 transcripts, read-back PASS, 11 of them
    already pruned at source and surviving only in the archive
  - gate 2b: the composition claim in PENDING-178 stands (11 mumbles,
    ~a fifth, on 08-31). Two terms do not close and are reported as an open
    disagreement, not a correction — the reconstruction is a demonstrated
    lower bound and -178 enumerated live
  - gate 2c: five sites in four files; -178's [HARDENING] scope holds
  - MEMORY.md record-only arithmetic correction: N-now 44 -> 65 measured,
    composition 41 real + 24 mumble added, direction reversed (it is rising,
    not shedding), stale :331 pointer corrected to :513

Filed as an item rather than a PENDING-178 addendum on PENDING-145's
mechanism: a ruling claims a number, so an addendum would be suppressed the
moment -178 is ruled. The undecided filing moratorium is disclosed inside
the item.

governance-drift-check.py:513 excluded on receipt and not examined.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X3L79vgAnt1x2kxvf23Qt7
This commit is contained in:
David F Glidden
2026-09-04 10:22:23 +02:00
co-authored by Claude Opus 5
parent e5db321d6f
commit c150bdff17
4 changed files with 167 additions and 1 deletions
+1 -1
View File
@@ -35,7 +35,7 @@ permalink: claude-memory/memory
**Loud trigger — pointer suffices**
- [Chamber work: ground in constitution + charter + runbook FIRST](feedback-chamber-work-ground-in-constitution-charter-runbook.md) — **any chamber work *or talk about it*** begins there (incl. `reanchor:`). Repo CLAUDE.mds are pointers, not state; corpus claims come from a self-tested tool, never a hand grep.
- [Governance files are dotfiles symlinks](reference-governance-files-are-dotfiles-symlinks.md) — Edit/Write refuse to write through a symlink, so **edit the real `~/dotfiles/…` path** when appending — **`PENDING`/`REVIEWED` only.** ⚠ That refusal is a tool artifact, **not** a permission check, and this note is the documented route past the only friction guarding the constitution (PENDING-107). **`~/CLAUDE.md`/`~/REVIEWED.md`/L2 = `[ESCALATE]`, steward's hand — a jurist sign-off does not authorize one.**
- [Verification ladder](reference-verification-ladder.md) — the named instruments; reach for the gate the claim's shape demands instead of re-deriving one. ⚠ **FROZEN — REVIEWED-123 (2026-08-17): no additions, rewordings, removals or reorderings, from ANY source, whatever its authorization**, until the trial is graded (N-now **44/84 as of 2026-08-31** — *DOWN 7 from 51 on 08-25, and the direction is the point: the window is shedding faster than it gains, so grading may RECEDE rather than approach; the counter is a 30-day ROLLING WINDOW per PENDING-147, so it rises and falls as transcripts age in and out. Re-measure, never relay: `len(glob('~/.claude/projects/-Users-davidglidden/*.jsonl'))`, the trial's own method at `governance-drift-check.py:331`*). Entries earned meanwhile queue in **PENDING-141's owed-entries list** (4 rows). Reading it is unaffected — that is the point of the freeze.
- [Verification ladder](reference-verification-ladder.md) — the named instruments; reach for the gate the claim's shape demands instead of re-deriving one. ⚠ **FROZEN — REVIEWED-123 (2026-08-17): no additions, rewordings, removals or reorderings, from ANY source, whatever its authorization**, until the trial is graded (N-now **65/84, measured 2026-09-03** — *and the direction has REVERSED: 52 on 08-31, 65 today, **rising fast and rising for the wrong reason**. Composition, measured the same day by the fool's own prompt signature: **41 real + 24 mumble (36.9%)**. The earlier reading here — 44/84, "DOWN 7 from 51", "shedding faster than it gains" — was **wrong in the number and backwards in the direction**; corrected 2026-09-03 as a record-only arithmetic fix (PENDING-179 gate 2b). The counter is still a 30-day ROLLING WINDOW per PENDING-147, so it can fall; it is not falling now. ⚠ **Report the COMPOSITION with the count** — a bare 65 reads as 65 sessions and roughly a third is machine chatter (PENDING-178). Re-measure, never relay: `len(glob('~/.claude/projects/-Users-davidglidden/*.jsonl'))`, the trial's own method at `governance-drift-check.py:513`*). Entries earned meanwhile queue in **PENDING-141's owed-entries list** (4 rows). Reading it is unaffected — that is the point of the freeze.
- [Skill-harvest register](skill-harvest-register.md) — canonical home for proposed skills (governed analog of PENDING.md for tooling); `/wrap-up` §1.6 proposes, steward authorizes. **Rebuilt 2026-08-07** from the archive: **154 live proposals**, grouped by kind, each with an exact `archive:L###` pointer. ⚠ The 2026-08-01 compaction was *lossless but illegible* (55 scraped header rows, 95% of cells cut mid-word) — the count it advertised, 177, was never the number. ⚠ **THE STROKE-2 BATCH IS AUTHORIZED BUT SHOULD NOT BE EXECUTED YET — PENDING-141.** 41 rows stamped `S2` are authorized (2026-07-19, execution not a ruling) and 22 are `S2?` (unsettled). **BUT appending them triples the ladder from 20 entries WHILE a pre-registered trial is measuring whether the ladder is reached** (baseline 14%, graded at 84 transcripts). Ladder size is an uncontrolled variable in that design, so executing the authorized batch would confound the only check behind REVIEWED-95's causal claim. **⚖ RULED — REVIEWED-123 (2026-08-17): AUTHORIZED (a) HOLD, on six conditions, and the freeze is GENERAL rather than S2-specific.** Grading does NOT authorize the append — the batch returns for a *ruling*. Report **N-now at every wake** until lifted; 30-day review is a live `DEFERRED-DECISION`. ⚠ Filing new proposals now requires a **declared firing moment** (`/wrap-up` §1.6); one that cannot name it is documentation and must say so.
- [Copy-paste-clean governance drafts](feedback-governance-drafting-copy-paste-clean.md) — draft PENDING/REVIEWED blocks as plain fenced markdown; display formatting leaks into the placed record.
- [Verify-before-compose hook](feedback-verify-before-compose-hook.md) — chamber constitutional writes are BLOCKED without `<!-- GROUNDED-IN: … -->` + verbatim Grounding. Don't fight the block.