Failures, corrections and fix-forward dispositions

Observation	Disposition and remaining gap
Shorter Ignis thinking budget fails coding invariants	Larger 4,096 budget recovered the fixture; held-out budget sweep remains proposed.
Initial image preflight reaches 256-token cap	Retained failed finish; separate 1,024-token rerun passed. Fact-scoped prompts and contradiction checks remain proposed.
Image markers pass despite invented interpretation	Independent image review retained in the note;18/18 is not full caption faithfulness.
Vision works directly while model discovery reports text	Runtime metadata follow-up proposed; no routed-vision claim.
Host reaches 16GiB cap without OOM	Keep peak/max-event telemetry and load-time memory sizing; do not infer leak or comfortable headroom.
QAT vision startup fails	Earliest actionable error: bf 16 linear: unsupported shape; exited 1, OOMKilled=false. Text-only paired survivor does not qualify vision.
Original coding-candidate context outer job fails packaging	225/225 raw observations remain separate from the failed job; compact-summary fix was tested on later jobs, not used to rewrite this failure.
Native lexical quality check accepts wrong arithmetic	Excluded from independent correctness; behavioral coding grader is authoritative.
Explicit token counting protocol needed	Campaign fixed Anvil's protocol-declared adapter and tested bounded stability; no runtime-name route branch.
Initial reference capture incomplete	Historical identity remains partial; final restoration is declared-policy restoration only.
Windows-native repository harness unsupported	No repository score; no POSIX worker was provisioned in this campaign.
Full product suite fails Windows collection	Existing Unix-only dependency; targeted campaign gates and Ruff passed. This is not a full-suite pass.
Shared-memory retention unavailable during campaign	Sealed local evidence retained; not treated as successful shared synchronization.

Public chart annotation correction

Publication audit found the sealed private chart's blanket “29 completion tokens” annotation inaccurate for the Ignis cell: its native metric records 28 completion tokens, while both NInfer cells record 29. The public chart uses the source values and labels 28–29. All three target 16 words; tokenization differs. The protected original bundle is unchanged. This annotation correction does not change native scores or timing metrics and further limits cross-engine decode-rate interpretation.

Privacy and reproducibility

Redactions intentionally withhold private identity and text. Public scores are audit records of the independent private grading, not a public re-execution proof. No model, route or service is changed by publication.