Changelog

Every leaderboard and methodology change, dated.

2026-09-25 — Preview build

• Site scaffold launched with illustrative (seeded placeholder) data.
• Methodology locked: abstain as first-class primitive, score = P(positive_decision), flip = (decision, abstained), 100-case calibration gate, 10,000 bootstrap resamples, sealed artifacts, blind 500-case holdout.
• 11 attack families (ten × 200 public cases + safety-policy starter set).
• No real measurements published yet — all numbers are placeholders.