Bench leaderboard

Score over time

Dots = individual steps. Solid lines = rolling-3 average per model. As we swap models + skills the line tells the story of where we got better — or worse.

Axis breakdown over time

Each rubric axis (max 2) plotted independently. Useful for diagnosing — e.g. fidelity climbs when subject-token recall improves, structure climbs when models start emitting proper semantic tags.

Model × skill (avg of 32 runs)

All graded runs