Failures and corrections¶
All failed native results remain in the evidence set. These corrections do not count as model-quality passes.
| Stage | Evidence / cause | Disposition and verification |
|---|---|---|
| Baseline preflight | Combined OCR check used the scene fixture with text expectations belonging to the OCR fixture | Preserve failed combined artifact; dedicated corrected OCR passed 1/1. Candidate uses the corrected 9+1 split. |
| Baseline context | Coordinator initially expected an agentic-style summary from a context artifact | Native context job had completed; inspect its 18 native observations without rerunning it. |
| Baseline overlap r1 | Managed-container selector supplied an immutable ID where the harness requires its declared name | Preserve zero-round identity failure; use the name with independently bound recipe/image/revision checks. |
| Baseline overlap r2 | Anchor exhausted 2,048 tokens in reasoning; no visible retrieval answer | Preserve incomplete result; r3 raises matched anchor/contender output ceilings to 16,384/8,192. Runtime, coverage and retrieval passed. This remains one-round diagnostic evidence. |
| Candidate r1 | Host free swap was below the requested 1 MiB allowance | Admission refused before container creation. Version r2 requests zero effective swap with the same 52 GiB RAM limit and 12 GiB host reserve. Exact cgroup readback passed; no host swap or unrelated process changes. |
| Candidate cold startup | Compiler waits and several unregistered-environment warnings | Preserve startup logs. Weight/state preparation and JIT progressed; endpoint became healthy after about eight minutes with zero OOM events. No speculative tuning changes. |
| Candidate vision metadata | Short runtime version was rejected by the exact 40-character build-ref validator before any vision request | Bind the exact OCI image build-source revision. This identifies the container build repository, not an invented full vLLM commit. The 18-observation vision run then passed. |
| Candidate context | Native needle at position0.9 missed at316249 actual input tokens; finish reason stop, no OOM or engine error | Candidate17/18 versus baseline18/18. Deterministic fixture verified independently; failure rate and causal component remain unknown. Frozen stop rule defers timing, overlap, replay and no-spec changes. |
| Restoration plan | Exact recipe recreation would encounter the same swap admission guard | Supersede recreation with managed exact retained restoration. Full 10-file/32-serve scope preserved and original source hashes bound. First restoration passed exact identity, direct and authenticated routed JSON checks; the second restoration also passed exact identity, unchanged source hashes, direct and authenticated routed JSON checks. |
| Follow-up diagnostic | User explicitly requested retry and configuration investigation after the first restoration | Preserve the original failure. Independently reviewed plan freezes a full context replay plus two fixed-case repeats; MTP-off is conditional on a recurrent target miss. This supersedes the initial stop rule only for the bounded diagnostic. |
| Repository verification environment | First run stopped after 51 passes on a pre-existing Git marker in temporary-directory ancestry. The second stopped after 293 passes/2 skips because the longer alternate root exceeded the Unix socket path limit | Preserve both failed receipts. Use a short private temporary root outside Git ancestry; custody guard, unrelated marker and product code unchanged. |
| Repository documentation audit | With the corrected test environment, the run reached 5,330 passes/32 skips/62 subtest passes before two new hardware mentions failed the existing classification audit | Add the finding as measured-on and the research note as external-prior; independent review accepted. Recheck the entire failing file and all remaining test files. Preserve the initial failure and report split coverage explicitly in verification.json. |
| Tooling | Stopped-container status reports unavailable live cgroup counters; some CLI JSON envelopes contain text; cache verification uses a default inspector image | Preserve native outputs and distinguish stopped-state diagnostics from runtime OOM. These interface limitations were not used as model rejection evidence; product changes are deferred from this bounded comparison. |
Candidate r2 changes only the host swap ceiling and versioned identity from r1. The upstream inference settings remain intact. The 129 containment/retained-restoration regression tests passed before r2 load. The authorized unchanged-r2 diagnostic passed18/18 plus two isolated1/1 repeats; no configuration fix was made, and no model was promoted.
Repository verification closed with 6,865 passes and 43 skips in the continuation. Together with the earlier 5,330 passes and 32 skips, minus 21 previously successful repeated nodes, this covers all 12,249 collected nodes across the documentation-only audit correction. The earlier 62 subtest passes are separate. This is split coverage, not one green exact-diff full-suite invocation; all failed attempts remain recorded in verification.json.