Fix a model and walk time: run-over-run gate scores from the committed ledger, with prompt-surface changes and comparability breaks drawn where they happen — thresholds applied by the page, never stored.
One point per run in the lens. A violet marker is a prompt-surface change; a dashed grey marker is a comparability break (a corpus, grading or protocol digest moved), and the series stops there rather than drawing across it — runs either side are not comparable (spec 14 §4).
Each mode points at one fix: descriptions, constrained decoding, scope shape, retry budget, or a bigger model. Shows the most recent runs in the window (8 unless the window filter says otherwise). Unlike the confusion and flip panels, which stay inside one comparability group, this one keeps every run in the window and marks a break with ⟂.