Skip to content

R10 monitoring identity failure and incomplete restoration acceptance

The candidate was created and reached startup readiness, but managed load exited1 because Anvil Serving ignored Ignis --model-id when deriving served identity. Monitoring correctly withheld registration. The full 48-row native report remains planned with zero submitted, attempted, completed or scored jobs. No candidate generation, quality or inference-latency result is claimed.

The candidate was removed. The first incumbent restoration omitted explicit GPU order and was rejected by monitoring; the exact-order retry restored the original service and active monitoring. Routed coding/JSON/approximately128k needle/20of20 tool checks passed. Restoration acceptance remains incomplete: the clean-input Pi attempt performed its real read and completed with the declared provider/model, but returned extra prose and a duplicated digest. Both Pi attempts remain retained, with no further retries.

This packet was assembled after the attempt. Raw candidate HTTP bodies and the inner registration exception were not durably retained; source-backed independent diagnosis and startup logs are distinguished from raw API evidence. Private originals and frozen plans remain authoritative. Their hashes cannot be revalidated from sanitized derivatives; no complete native projection or repaired measurement is claimed. No promotion. Independent packet/size/publication gates remain separate.