todo2code

Preprompt and technical directives (ticket-005)

• Task title: Audited cross-language reranking
• Created: 2026-07-31
• Governance source: wellmanifest/new-project

Requirements

1. Use retrieval only to produce a bounded shortlist.
2. Require a separate structured decision with explicit abstention.
3. Ground every accepted decision in repository-owned records, paths, symbols or capability terms.
4. Preserve exact-target precedence and the deterministic offline linker.
5. Record provider/model/revision, input hashes, scores and cited evidence.
6. Cache model-derived output by content and model identity.
7. Evaluate tracked snapshots only; never transmit untracked or private data.
8. Reject the approach unless it clears gold and real-repository precision gates.
9. Store executable source outside project/ticket-*.

Referenced evidence

• project/ticket-004/iteration-01.md
• project/ticket-004/audit.md
• evaluation/gold/v2/dataset.json
• src/graph/linker.ts
• src/core/text.ts
• docs/READINESS.md

Approval boundary

Initialization records the user’s request to continue, but implementation waits for review of README.md and ai-codex.md as required by P-CORE-008.

This site is open source. Improve this page.