boshu efe427ebb2 docs(pawl): DECISION — make --strict satisfiable via a measured agy certification (age-pawl-intent-zhndq.8) ...

--strict is the advertised strongest gate and CAN NEVER PASS today: it needs TWO distinct
strict-eligible COLD families and only codex qualifies (agy is A7-benched; there is no cold claude
adapter and never will be — LAW 0). An operator reaching for maximum rigor gets a guaranteed exit-5.

Decided by a genuine CROSS-FAMILY QUORUM (no self-grading; the author is neither judge):
  - codex (gpt family)   -> B: certify agy behind a measured bar
  - agy   (gemini family) -> C: B now, warm-strict deferred
Both reject warm-strict as the primary answer (it breaks strict's cold portability, and the measured
warm data says the bar may be unpassable: the tri panel reached FULL agreement 0/22 times vs dual
10/19). Both accept certifying agy — but only after a measured bar, never an unmeasured flag flip.

Grounded in live evidence from this session: with codex rate-limited, agy served as the sole
cross-family cold reviewer for ~8 beads and produced BOTH a real, correct defect catch (reconcile
falsely treated REBOUND verdicts as unbound) AND a false refute (wrong premise about source_hash,
disproved by 4 deterministic gates). Useful but noisier. On a strict door a false REFUTE fails SAFE
(HOLD) while a false CONFIRM would compromise the gate — so admit the noisy family IF its
false-CONFIRM rate measures ZERO.

The bar (all must pass before PAWL_STRICT_ELIGIBLE_FAMILIES is flipped): >=30 cold agy reviews over
real held-out beads + seeded good/defective diffs, adjudicated against deterministic gates; ZERO
false CONFIRMs; REFUTE precision >=80% (bounds the DoS risk both judges named); >=80% detection of
seeded defects; schema-conforming evidence. Until then --strict stays honestly UNAVAILABLE (exit 5,
never a silent single-family pass) — the status quo, so no code change holds the line.

.9 (F8b) is rescoped to RUN that certification; if agy fails the bar, .9 closes with the measured
evidence and strict stays unavailable — a real outcome, not a failure.

Claude-Session: https://claude.ai/code/session_01RDWPYn4ZS2SyBkcBvAM18u

2026-07-11 19:00:49 -04:00

..

2026-07-11-pawl-strict-satisfiable.md

docs(pawl): DECISION — make --strict satisfiable via a measured agy certification (age-pawl-intent-zhndq.8)

2026-07-11 19:00:49 -04:00