Skip to content

docs: record real-provider Hybrid ask retest without release promotion - #87

Merged
Air000000 merged 1 commit into
mainfrom
codex/hybrid-acceptance-retest
Oct 17, 2026
Merged

Air000000 merged 1 commit into
mainfrom
codex/hybrid-acceptance-retest

Conversation

@Air000000

Copy link
Copy Markdown
Owner

Result

One explicitly authorized real-provider retest after PR #86, using the unchanged 7 fictional documents and 8 fixed cases. All eight flow checks passed: two answer cases, four insufficient-evidence refusals, two no-context cases (empty/deleted). SQL/Chroma and document storage were temporary; JWT, tenant/category scoping, Dense default and typed audit/metrics checks passed in the isolated ASGI app.

  • Exactly 14 Beijing embedding + 6 Singapore rerank + 6 Singapore classification + 2 Beijing generation attempts; all 28 HTTP 200, no retries or extra runs.
  • Reported usage: 14525 tokens (374/430/13278/443 by kind); request bodies: 63541 bytes. ¥2 was the conservative authorized estimate, not a bill.
  • Generation completed in 3480 / 1889 ms with finish_reason=stop. Timeout was not reproduced; a single successful run does not establish root cause or latency reliability.
  • Preserves the first failed report, including its unknown possibly billed generation usage.

Still not release accepted

The NovaVPN 1.0 explicit version constraint is again audited NOT_REQUIRED. The 2.0 mismatch case labels version_and_time SUPPORTED despite acknowledging the source only covers 1.0; other MISSING facets correctly preserve overall refusal. Both semantic warnings remain visible, and frozen v5.3 NO_PROMOTION / FAIL is unchanged. This eight-case flow check is not an independent quality evaluation.

Read-only docker info still fails (dockerDesktopLinuxEngine pipe unavailable). Running-server HTTP, container restart/persistence and rollback acceptance remain unexecuted. Ticket Agent is still Dense, switches stay default-off, no public deployment.

Changes / verification

Only README and a sanitized JSON evidence report. No prompt/rule/code/.env changes, no business documents/TechQA/gold labels/secrets in model input bodies, no ticket creation. No further paid calls.

Complete offline suite rerun: 632 passed in 48.16s with temporary database, dummy credentials and external-socket/private-data guards. Report JSON, executed source SHA256 hashes, per-kind limits, totals, and all eight expected/observed statuses verified; git diff --check passed.

@Air000000
Air000000 merged commit 28e2e54 into main Oct 17, 2026
1 check passed
@Air000000
Air000000 deleted the codex/hybrid-acceptance-retest branch October 17, 2026 14:53
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant