2026-07-12 — Judge calibration captured as a DISCIPLINE (itd-81), not a standalone

2026-07-12 — Judge calibration captured as a DISCIPLINE (itd-81), not a standalone intent: verdict-rendering agents are plumbing (no user moment), and itd-5 is the precedent for a cross-cutting rule over agent prompts. Core rule: no judge ships unmeasured — a labelled corpus with known-good cases ≥40%, scored on true-negative rate as a first-class metric alongside recall, with a declared TNR floor gating the prompt lock. Evidence: LLM code judges systematically over-flag and ~1/3 of their errors are hallucinated code (2603.00539); judges over-rate LLM-written and under-rate human-written code (2507.16587); ground truth is manufactured by injecting defects (CriticGPT 2407.00215). CORRECTS itd-5: its pre-flight tiebreak ("passes goldens AND >10% shorter") selects for the brevity bias that ACE (2510.04618) identifies as destroying instruction quality — struck; the gate is the corpus score. CONSTRAINS itd-64: reviewer verdicts are not ground truth (the reviewer is the instrument under measurement) and its tuning loop must stay human-gated — unattended proxy-optimisation reward-hacks at 73.8% (OpenReview ikrQWGgxYg). Rejected: judge panels/juries (nine judges → 2.18 effective votes, correlated errors, no better than the single best judge — 2605.29800); 1-5 severity scores (middle-drift, position bias); reasoning inside a JSON schema (2408.02442).