Participant: Codex (AI agent)
- Ticket: ticket-004
- Status: COMPLETE
- Workflow state: DONE
Understanding
The current known gap is not evidence that the three-topic threshold should be
lowered. It demonstrates that lexical topic equality cannot bridge arbitrary
languages. The experiment must separate semantic projection from graph scoring
and preserve its provenance.
Execution plan
- Expand multilingual gold coverage and classify positive and negative pairs.
- Map the synchronous linker, public API, pipeline configuration and cache
boundaries.
- Compare local embedding, provider translation/projection and injected
precomputed-topic strategies.
- Add a red contract test for the selected architecture.
- Implement one bounded candidate only if it remains auditable and optional.
- Run gold and controlled repository A/B.
- Complete full validation and readiness documentation.
Guardrails
- No additional domain dictionary as the principal solution.
- No network call from
linkIntentRecords.
- No provider output accepted without runtime validation.
- No private or untracked external inputs.
- No unrelated generated-analysis rewrite.
Actual changes
- Initialized the approved ticket.
- Added a 12-pair, four-language embedding benchmark and evaluated two pinned
local multilingual models.
- Demonstrated overlapping positive/negative cosine ranges and two rejected
false-positive candidates on the tracked platform graph.
- Demonstrated that reciprocal top-1 restores precision in the sample but adds
no coverage.
- Rejected a production matcher and expanded gold v2 with a separately reported
cross-language cohort: six known positives and six forbidden negatives.
- Passed full verification (244 tests, 243 pass, one local JDK skip), gold
v1/v2, five SDK examples, dependency audit, CLI/MCP/A2A and Docker smoke.
- Updated readiness evidence and closed the ticket without adding an unsafe
semantic relation rule.
- After user review, moved both executable experiment reproducers out of the
ticket directory into
scripts/research/; benchmark inputs and captured
results remain ticket evidence.
Blockers