Friction and fix-forward record

Strict JSON baseline failure — open quality limit

The single pretrial routed JSON smoke completed with a normal stop but included prose before the requested object. Strict parsing failed. The original public-safe response is retained in baseline-smoke.json. The request asked to disable thinking, but effective disablement was not established under the shared tier's existing overrides.

Disposition: retain this as a baseline quality failure and compare the separately planned restored-incumbent smoke. Do not retry, relax the exact-object oracle, or attribute a pre-existing failure to the candidate. The durable frozen plan requires the matched pair and separates transport/readiness/identity from output quality. The matched restored smoke also failed (66 completion tokens versus 53 before); operational restoration passed and no quality regression was established. The existing JSON-smoke issue remains the durable quality follow-up.

Prior strict kernel failure — policy distinction retained

The preceding trial's no-worse binary64 rule failed on a single extra boundary token. This trial froze a new native upstream count gate before lifecycle activity, preserved input/output integrity and binary64 safety, and retained the old criterion as a diagnostic. The same one-token difference reproduced; its no-worse diagnostic remains false.

Disposition: earlier failed findings remain immutable. The pinned source and independent review, plan, and native count arrays are the durable record. Offline recomputation of every stock/native and patch/native batch formula passed. This closes the admission-policy ambiguity for this trial only; it does not establish exactness, general numerical equivalence or model quality.

Tool coverage and claim routing — bounded preparation friction

The connected tool surface did not expose the exact runtime inspector, image build and isolated probe. The campaign owner used their registered Anvil Serving CLI commands. The ordinary claim path refused an enrolled root; the supported root-claim path was used with the original refusal retained privately.

Disposition: retain the invariant that these operations use their managed product owner. No raw model lifecycle workaround or production fix was introduced. The source-freeze receipt records the independently reviewed executable source; operational results remain separate.

Repeated reasoning close after the first boundary — open runtime diagnosis

The single candidate response emitted closing markers at positions 64 and 126 and failed both strict boundary and exact-answer gates. The prior token-0 loop did not recur, but this is not complete acceptance. Disposition: preserve the failure and stop without a retry. The tracked repeated-close ticket specifies the smallest offline MTP, thinking-budget and parser investigation. The independent live review verifies the raw token and trace claims; cause remains unresolved. The separate JSON-smoke issue remains open.