todo2code

Ticket 005: Audited cross-language reranking

• ID: ticket-005
• Owner: tom-sapletta-com
• Status: DONE
• Workflow state: DONE
• Created: 2026-07-31

Goal and scope

Evaluate a two-stage cross-language linking path: semantic retrieval may create only a bounded candidate list, while a separate structured reranker must cite repository-owned evidence and may abstain. Retain a production change only when it closes the six current cross-language gold gaps, preserves every forbidden pair and improves coverage on an additional tracked repository.

Executable implementation belongs in src/ and regression coverage in test/. Optional experiment reproducers belong in scripts/research/. This ticket directory is limited to governance, inputs, captured outputs, decisions and logs.

The approved continuation adds a prerequisite communication audit: verify that the governance-standard user-* and ai-* files are converted into distinct human/agent Intent DSL records, compare their intent, and identify the participant who must respond when scope, polarity or coverage diverges.

Acceptance criteria

• AC-01: Define a versioned candidate and reranker contract with explicit model/provider identity, score, cited record IDs and abstention reason.
• AC-02: Keep network/model calls outside the synchronous deterministic linkIntentRecords boundary and preserve the current offline default.
• AC-03: Candidate generation is bounded and cannot create a relation by itself.
• AC-04: The reranker accepts a candidate only with repository-owned evidence; unsupported, ambiguous and multi-module statements abstain.
• AC-05: Gold v2 cross-language recall rises from 0/6 to 6/6 while all six cross-language forbidden pairs and all existing hard negatives remain clean.
• AC-06: A tracked repository outside the ticket-004 primary pair shows improved implementation coverage without a manually rejected new relation.
• AC-07: Any dependency or provider is pinned, licensed, security-reviewed, cacheable and optional; no private or untracked source is transmitted.
• AC-08: Full verification, both gold versions, examples, dependency audit, CLI/MCP/A2A smoke and Docker validation pass.
• AC-09: If the quality boundary is not met, reject the candidate without a production semantic rule and preserve the measured failure.
• AC-10: No executable source is stored under project/ticket-005.
• AC-11: Governance-standard user-* and ai-* files are recognized without front matter, while ticket specifications and generated evidence are not misclassified as participant communication.
• AC-12: Communication analysis reports an explicit response owner for missing response, human-agent conflict and agent work outside the human request.

Non-goals

• Growing the hand-written Polish dictionary.
• Lowering the three-topic lexical floor.
• Treating embedding similarity as implementation evidence.
• Enabling provider-dependent behavior by default.
• Choosing one module for a genuinely multi-module requirement.

Participants

• user-tom-sapletta-com.md
• ai-codex.md

Evidence

• preprompt.md
• audit.md
• ai-codex-logs.txt
• changelog.md

Approval

• Decision: approved
• Evidence: user instruction to handle the next todo2code tickets and audit user-*/ai-* Intent DSL divergence
• Date: 2026-07-31

Conclusion

The communication prerequisite is retained. Governance user-* and ai-* sections become distinct human/agent Intent DSL records, and each detected divergence names the role and participant who must respond.

The semantic production candidate is rejected. Captured gold decisions satisfy 6/6 expected cross-language pairs with zero forbidden pairs, but three live OpenRouter attempts on the clean tracked subactor/platform snapshot failed the structured contract before any relation could be materialized. The provider first omitted decisions, then returned judgments, and finally returned an invalid non-numeric confidence. Consequently AC-06 and AC-07 were not demonstrated. The deterministic linker remains unchanged, and the experimental reranker is not exported from the package, CLI, MCP or A2A.

This site is open source. Improve this page.