Endpoint drift and subject eligibility

The Gemini sentinel fingerprint changed from 10/10 baseline matches to an oscillating sequence — 10 → 9 → 9 → 8 → 8 → 7* → [re-baseline 10] → 6* → 7* → 6* → 7* — on the unversioned endpoint, triggering block-boundary freezes, disclosure memos, re-baselining, and a later attestation gate; the Phase 4 F cross-vendor tier was demoted to descriptive-only rather than re-evaluating the rule on the data it fired on. GPT-4.1’s sentinel cells remained 10/10 throughout, and every deviant episode was a clean, valid single-token choice — a distributional change, not an infrastructure fault.

Separately, the original Claude Haiku candidate failed the registered Gate-0 behavioral entry gate (it could not complete a turn at max_tokens=16) and was replaced under a sealed amendment. Model identity and endpoint availability do not by themselves establish a stable, usable behavioral subject; deployed-model change over time is why behavioral sentinels exist in the architecture. Contaminated confirmatory spend: zero.