todo2code

Participant: Codex (AI agent)

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

  1. Expand multilingual gold coverage and classify positive and negative pairs.
  2. Map the synchronous linker, public API, pipeline configuration and cache boundaries.
  3. Compare local embedding, provider translation/projection and injected precomputed-topic strategies.
  4. Add a red contract test for the selected architecture.
  5. Implement one bounded candidate only if it remains auditable and optional.
  6. Run gold and controlled repository A/B.
  7. Complete full validation and readiness documentation.

Guardrails

Actual changes

Blockers