Failures and durable disposition¶
| Observation | Cause or boundary | Disposition |
|---|---|---|
| Original ZIP gate0/3 | Raw numeric98101 violates string schema | Corrected with pinned compatible704 runtime; fresh3/3. Closed for this gate. |
| Latest upstream incompatible | Current HEAD expects artifactv3; supplied artifactv2 | Use exact704 compatible source. No conversion or checkpoint substitution. |
| Unsupported thinking control400 | 704 rejects chat_template_kwargs | Omit field; server no-thinking default. Low uses separate top-level control. Closed configuration issue. |
| Missing invalid tool payload | Runner discarded failed arguments | Bounded diagnostics implemented and independently reviewed;173focused tests. Closed. |
| Tool20 burst17/20 | C1 queue rejects3 requests429 | C1 declared protocol5/5; overload outcome retained. Capacity beyondC1 unqualified. |
| Boundary quoting | Model adds literal quotes in declared strings | Generic schema instructions improve18/21; original failures preserved. ExactCRLF still fails. Open quality boundary. |
| 128word counting | Early stop or no compliant visible output in off/low | Strict exclusion retained; separate32word short-output comparison, no long-output speed claim. Open. |
| Runtime build reproducibility | SHA-pinned executable in mutable build cache, apt resolution mutable | Managed reload verifies SHA, but clean immutable image remains required before deployment. Open. |
| Fullsuite Pi smoke | Existing Windows Node startup CSPRNG assertion from prior diagnosis | Separate ticket; do not attribute to candidate. Latest rerun:7933passed,416skipped,1failed at same Pi smoke. |
The independent reviewer found no CRLF normalization in704 parser; global normalization would corrupt the string contract. Exact-byte fields need a separately designed lossless application representation, not relaxed scoring here.
A GGUF control preflight was inadvertently sent before managed startup completed:4HTTP503 Loading model responses are retained separately. After managed load completion its preflight and repeated quality pass. Those correctness commands briefly overlapped; their latency is not comparison evidence. Capacity ran later in isolation with explicit warmups.