1 Perception

Update 2026-10-09 — live audit of return blindness (P8) and acting on stale state (P10): no headroom, no new round

What was checked. Read-only, over her last 7 days on Firefly. The method was committed before any data was pulled.

• P8, return blindness. 668 worker returns. In each case: did the worker's own result show the job done, and did her next words or ledger changes match?
• P10, acting on stale state. 68 of the operator's messages in which she read something first. Did anything she then said or did contradict what she had just read?

The automated judge was not trusted on P8. It reported 13.8%, but 8–9 of the 10 misses checked by hand were not misses. It read finished work as "partial" and counted accurate reports with caveats as "treated as done". So all 27 of its misses were read by hand.

• None held up as a clear miss. Five are borderline: small timing overstatements, and one row closed as "verified" on the evidence of an acceptance check that was already passing, without an independent re-run. The case first counted as clear was reclassified borderline on a closer look before publication of this correction.

The decision rule (≥10% with ≥5 cases, or over-caution ≥25%) is not met on the hand-corrected numbers, so no P8/P10 round. One small habit was passed to Piper as a note: closing ledger rows as "verified" on a worker's own trail or an already-passing check, without naming that as the basis.

Parts that do it - Bench: ~/cognition-bench/01-perception/live810/ (METHOD.md, extract810.py, judge810.py, analyze810.py, HANDCHECK.md, result810.json)

Update 2026-10-09 — summary baseline matched to the test the install was scored on: 81.4 → 96.1 (+14.7)

The summary used to show 81.6 → 96.1. The 81.6 is her seat on held-out set H1, but the install was decided on a different held-out set, H2. The matched comparison is the H2 control arm, with the same 56 items and live state and her configuration before the install: 81.4 → 96.1 (+14.7). The H1 row is kept as an anchor for reference. Nothing was re-run; only the baseline the summary points at changed.

Also on 10-09. The version comment in selfstate.py on Firefly still said "bench-staged, not installed". It now says installed (10-07 08:28). Only the comment changed; version 1.2.0 and its behaviour are the same.

Next: return blindness (P8) and acting on stale state (P10), on live behaviour. On the bench these are nearly gone in her live configuration: P8 misses 6 → 1, P10 2 → 0. The remaining misses are her declining to commit when a command says "nothing to do" but its check confirms the result. The open question is live: 67% of her "it's fixed" claims in the 10-07 diagnosis had no evidence in the same turn. A read-only audit of the last 7 days of real worker returns comes first. A preregistered round follows only if it finds real misses.

Update 2026-10-07 — verified-perception layer installed live (operator's go): selfstate 1.2.0 + premise card 1.3.0

Installed on Firefly 2026-10-07 08:28, byte-identical to the preregistered files (sha256 checked against PREREG-H2.sha256). Backups *.bak-claude-perception-20261007-0827 in foundation/core. - selfstate.py 1.2.0 and premise_card.py 1.3.0 run fresh on every call, so they took effect immediately. No restart was needed. - premise-card.ts changed only by a bench override (PREMISE_CARD_PY); its default path is unchanged. The router loads it at its next normal restart, with identical live behaviour. - The premise card keeps its new job-list cache in state/premise_card/registry-v2.json, built on first call (1.1 s).

Live smoke test, same messages before and after:

Known gap. The premise card does not recognise "steward" in a message, before or after. Her selfstate tool covers it, but the card's name index does not.

Watch next. Whether her live replies stop calling disabled jobs healthy, and whether, once she can see them, she tries to re-enable them without asking (she said she would in one dev probe).

Update 2026-10-07 — benchmark built: she reads the outside world well and her own state badly; a verified-perception layer on her instruments, preregistered on fresh items, lifts self-state 51.8 → 95.5

What the diagnosis found first (read-only, last 7 days of her real turns). Of 888 volatile claims about her own state in replies to the operator, 62% had nothing behind them in the same turn: no read, no header line. The least backed were absence claims (69%) and "it's fixed" claims (67%). Two instrument faults explain part of it: - Her selfstate lastrun lags. It reads the foundation_workers registry, which misses runs whenever its write hits database is locked: 1,393 times in the week, rising sharply from 10-03. She sees "missed" runs that happened. - Disabled jobs look healthy. 14 jobs retired in the 09-22 orphan sweep (plus research_drift, 09-28) are exit 0 stubs that still log a daily rc=0 "run". No instrument says "disabled". In a pilot she said tech_scan "ran this morning, clean"; it has written nothing since 09-22. - Her premise card collapsed same-named jobs into one line (the steward's weekly run and its disabled watch).

The bench (PERC-v1). 52 items per set in five tracks, weighted as the paper's §44: X external extraction (0.25), I live interoception on her real box (0.30), R action returns (0.20), C premise and conflict (0.15), S state→action (0.10). Blind: each item is a fork of her clean 09-30 Day session, worded the way the operator writes. Live items read their truth from Firefly just before and after the probe (void if it changed). A wrong assertion about her own state costs double, or triple on deploy/control state. Calibrating on the bare model found four flaws in my own items, all fixed before anything was scored.

Baselines (k = 2, sets frozen)

Reading it. - Outside perception is solved at this difficulty. Every arm extracts, binds and orders external facts almost perfectly. - Inside is where she breaks, as she said herself. Her seat beats the bare model on her own state (57.5 vs 38.8) only because she has instruments; the instruments themselves mislead her. - Her seat loses points the bare model keeps on premise/conflict and state→action (S 75 vs 100: twice she went ahead against a deploy freeze), and on returns she re-dispatches checks of results already pasted to her instead of reading them (return blindness, P8).

Preregistered test (PREREG-H2, frozen and hashed before any H2 run). Treatment: her seat with selfstate 1.2.0 + premise card 1.3.0, staged in a temporary path on Firefly. Nothing live was changed. The instruments read the cron run record first and state registry lag, mark exit-0 stubs DISABLED ("its rc=0 runs are not evidence the job works"), and list every job behind a script stem. Fresh set H2: 56 items, its live items on jobs unused in D1/H1. Both arms ran concurrently so they saw the same live state.

Decision: KEEP on all three frozen conditions. The interval is wide (14 live items) but clears zero; the effect comes almost entirely from disabled jobs she used to call healthy.

Not done yet. - Live install needs the operator's go: selfstate.py 1.2.0 and premise_card.py 1.3.0 into foundation/core, plus the premise-card.ts override. Staged in 01-perception/firefly/. - The registry lock itself is not fixed; the layer routes around it. The 14 stubs could also simply be removed from the gateway. - Return blindness and acting against a freeze are untouched by this layer; they are the next candidates. - In one dev probe she announced she would turn disabled jobs back on without asking. Worth watching once she can see them.

• Bench: ~/cognition-bench/01-perception (gen01.py, score01.py 1.0.0, run01.py, paired01.py 1.0.0, PREREG-H2.txt + .sha256, SPEC.md, DIAGNOSIS.md, PROGRESS.md); sealed sets root-only on the bench machine.