todo2code

Ticket 005 audit

Decision

Reject the live cross-language reranker as a production feature. Retain the offline contracts, schemas, tests, captured gold fixtures and research reproducer. Do not export or enable the reranker through the package, linker, CLI, MCP or A2A.

Communication audit

The final ticket produced 51 codex records and 4 tom-sapletta-com records after section-aware conversion. There are no blocking polarity conflicts. The final issue ownership is:

• 7 AGENT_CLAIM_WITHOUT_EVIDENCE findings require codex to attach commit or test evidence (the current implementation is intentionally uncommitted);
• 1 AGENT_HUMAN_DECISION_CLAIM_UNCONFIRMED finding requires tom-sapletta-com to record or reject the approval in the human-owned file;
• 8 AGENT_WORK_OUTSIDE_REQUEST warnings require tom-sapletta-com to record or reject the detailed scope that currently exists only in the conversation.

The agent may correct its seven evidence claims, but must not edit the human-owned participant file to silence the other nine findings.

Historical read-only material from wellmanifest/new-project commit 2b9e3c9 showed why a filename-only migration is unsafe:

• plain rename to user-*/ai-*: zero records and owner-specific migration warnings;
• typed Opus request/message sections: 9 human + 58 agent records, zero issues;
• typed GPT56Luna request/message sections: 9 human + 72 agent records, three unmatched request fragments and no false conflict between different files.

Offline reranker result

Gold v2 uses captured, structured decisions through the same runtime validators:

• expected cross-language relations: 6/6;
• forbidden cross-language relations: 0/6 violations;
• accepted: 6;
• abstained hard-negative cases: 1;
• deterministic linker remains 0/6 and unchanged.

Live tracked-repository result

• repository: subactor/platform;
• clean commit: 3e96573d587cb664741849ceba205bf303b9f418;
• current graph fingerprint: 250df4ff83f456fb371278d4a0c2cf17dd025582b6865a0cb2f01bd469fd1dd0;
• retrieval: the pinned multilingual E5 ranking captured by ticket 004;
• bounded payload: six reciprocal selected declarations, initially top-3 (18 candidates), then top-1 (6 candidates);
• model: qwen/qwen3.7-plus;
• declared evaluation revision: qwen3.7-plus@2026-07-31;
• privacy boundary: clean HEAD required; every projected declaration and module path had to be tracked; generated graph and result paths stayed outside the worktree.

Three live attempts failed closed:

1. top-3 returned a JSON value without a decisions array;
2. top-1 returned the top-level key judgments instead of decisions;
3. top-1, after an explicit key instruction, returned at least one confidence outside the required numeric 0..1 contract.

No accepted result artifact exists because invalid provider output is not promoted into t2c.semantic-rerank/v1. No relation was created, no coverage metric changed, and the two false embedding candidates from ticket 004 were not silently accepted.

Validation

• npm run verify: 251 tests, 250 pass, 0 fail, 1 local JDK skip;
• isolated CLI watch retry: 3/3 pass after one full-suite timing failure;
• gold v2 and v1: PASS;
• examples: 227 records, 97 relations, five SDKs: PASS;
• npm audit --omit=dev: 0 vulnerabilities;
• CLI, MCP, A2A and Docker smoke: PASS.

This site is open source. Improve this page.