cesnimda
12 lines
2.5 KiB
Markdown
12 lines
2.5 KiB
Markdown
# Local strategy structured-output correction — 2026-10-29
|
|
|
|
The final local synthetic worker smoke exposed an inherited contract conflict after the cumulative regression freeze. A strategy snapshot requested a JSON object through `GenerateSectionWithMetadataAsync`, which called `/cv/rewrite` in its default text mode. That route explicitly instructed the model to return a CV and "No JSON." The real queued operation reached the provider but failed with `invalid_provider_response`; it was not silently accepted. This is separate from CV factual extraction qualification.
|
|
|
|
The correction adds one explicit `StructuredJson` generation mode to the existing adapter. Strategy snapshots select it; the sidecar uses JSON generation with a trusted task/schema and delimited untrusted source, requires an object, and returns the existing envelope. The backend's existing strategy schema validation, retries, ownership, persistence and approval boundaries remain unchanged. Text rewrite and CV block modes retain their existing prompts. Older adapters fail closed rather than pretending to support typed output. Mode-specific cache keys prevent a text response from satisfying a structured request.
|
|
|
|
Evidence: `strategy-mode-red.log` retains the original rejection (422) of the new mode. `strategy-mode-focused/strategy-mode.trx` passes ten backend tests; the updated strategy fixture only accepts the typed overload. The source-layer candidate image passes all 83 offline Python tests, including JSON-path selection, absence of conflicting CV rules and non-object rejection. Live runtime proof and final full backend outcome are recorded in LOCAL_RUNTIME_ACCEPTANCE.md. Prior 1,447-backend/81-Python and browser/PDF results remain bound to the earlier freeze; later tests are reported separately rather than rewriting that evidence.
|
|
|
|
This does not qualify model factual accuracy. The original failed operation and temporary-account cleanup evidence remain under `local-ai/final-worker-journey-02/`. No real user documents or production infrastructure were used.
|
|
|
|
Final validation: complete backend rerun 1,448 passed, complete final-image offline Python 83 passed, focused static review no blocker. Actual local strategy queue succeeded on its first attempt and saved output was read back. Earlier failure and recovery are distinct observations. The application change is `caa0c6a`; final images and exact evidence are in LOCAL_RUNTIME_ACCEPTANCE.md. Broad browser/provider/PDF lanes were not rerun after the mode correction and remain explicitly bound to their earlier freeze.
|