prd-to-evallisted

# PRD → Eval Disqualifiers become binary gates; tradeable qualities become anchored scores. Mixing the two — averaging a gate, or scoring fabrication 3/5 — is the root failure of most evals, and the one thing this skill refuses to do. ## Verification gates (defined first; output is blocked until all pass) - **G1 — Correct sorting:** every requirement lands on the right side of the line. Disqualifiers (fabrication, missing required fields, safety violations) are binary gates — partial credit on them is meaningless. Tradeable qualities (concision, tone, naturalness) are rubric dimensions. A scored disqualifier or a gated taste-call fails. - **G2 — Gates binary-testable:** every gate states a yes/no check procedure and is marked `mechanical` (string/count/regex-checkable) or `judge` (needs an LLM/human applying the stated check). A gate whose check can't be written as a definite procedure is reformulated or demoted to the rubric. - **G3 — Rubric anchored:** every dimension has concrete 1-anchor and 5-anchor descriptions (what a 1 actually looks like, what a 5 actually looks like). "Poor…excellent" scales fail. Every gate and dimension traces to a spec line — no invented requirements. ## Steps 1. **Harvest requirements** from the spec, keeping the source phrase for traceability. Split compound requirements ("accurate and concise" is two). 2. **Sort by the one question:** *if an output fails this and aces everything else, do we ship it?* No → gate. Depends/tradeable → rubric.