You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hero verdict banner leading both reports: verdict word + one big memorable number (degradation ratio / threshold / steps analyzed) — color-blocked in HTML, color-boxed with a unicode sparkline of your rot curve in the terminal
Share card: 1200×630 social-ready card in the HTML report with Save as PNG and Download SVG buttons — pure inline SVG, still zero network calls
Per-model comparison: independent rot curve + verdict per model family (Opus vs Sonnet vs …), comparison table in both reports, small-multiple charts on a shared scale in HTML, additive models key in --json. Models under 150 steps fold into "Other".
Fixed
HTML chart y-axis was squashed when a tiny bucket's wide confidence interval set the scale (seen in real user reports) — low-confidence buckets no longer drive the axis
"CIs overlap" jargon now reads "within statistical noise"