AGENTS

Paraphase STRC validation

Public-data validation lane for Paraphase’s STRC/STRCP1 caller.

Hard rules

1. Read ../../AGENTS.md first. It wins on STRC vault policy.
2. Public human data only. Never copy family, clinical, collaborator, or controlled-access data here.
3. Pin every comparison to an exact commit and identical input bytes.
4. Compare v3.1.2 (paper baseline), v3.5.0 (last Python), v4.0.0 (Rust rewrite), and current main.
5. A v3.1.2 delta alone is not a Rust regression. Gate on v3.5.0 versus Rust.
6. Port parity is not biological truth. Use phased assembly evidence before making an accuracy claim.
7. Keep downloaded BAMs, references, build products, and run outputs out of git.
8. Keep sources/ and worktrees/ pristine unless preparing a deliberately scoped upstream patch.
9. Do not open an issue, PR, comment, or email without Egor’s explicit approval for that action.
10. Never make medical or patient-specific claims from this benchmark.

Success gate

• Same compact public fixture yields reproducible calls in two runs per version.
• v3.5.0, v4.0.0, and main agree semantically, or a minimized discrepancy is explained by orthogonal evidence.
• Current Rust test suite passes with the STRC fixture/test.
• Provenance records source URL, source MD5, tool versions, reference MD5, and fixture SHA-256.