| One patched ordinary row kept one extra token relative to both stock and the binary64 reference. |
Preserved the exact row receipt; the strict local no-worse gate failed and startup stopped before weights. |
Review whether the predeclared conservative policy should remain the admission contract or whether upstream should address the split-search boundary. Do not weaken the completed run retroactively. |
Independent outcome review completed: the frozen gate failed correctly. The raw receipt records token, count, mass, input hash, and failed check. |
Open |
| The previous campaign lacked an independent ordinary-row oracle. |
Added the frozen binary64 numerical reference, seed 42 control, input-byte checks, prefix/subset checks, and bounded diagnostics. |
Keep the reference design and native receipts as the reusable diagnostic contract. |
Eight cases and 272 rows completed; no ambiguous oracle boundary, missing oracle token, prefix hole, invalid output, or input mutation was reported. |
Closed for this diagnostic |
| Post-exit effective cgroup values were unavailable. |
Retained separately from the primary kernel assertion; oom_killed is false. |
Preserve exit cause and post-exit observability as distinct fields in managed status. |
Candidate status reports exit 1, OOM-killed false, and unavailable post-exit observations. |
Open product limitation |
| The restored incumbent failed its single strict JSON smoke by emitting reasoning-like prose before the target object. |
Preserve the complete synthetic response and keep functional restoration acceptance incomplete. Operational identity/readiness restoration remains separately verified. |
Tracked evidence gap: add a matched pretrial control in any future experiment and keep exact-output acceptance separate from readiness. |
No matched pretrial smoke exists; candidate weights never loaded, so candidate causality is refuted and regression is unproven. |
Open |