Which model actually delivers?

Every number below comes from the same 110 published one-shot rounds — you can open any round and read the raw answer behind each cell. We keep the two things a model can be good at as two separate axes: writing a correct short answer, and writing code that actually runs. We do not blend them into a fantasy index.

Axis 1 — Text rounds, blind rubric scores

Scored against a written gold answer and rubric, without knowing which model wrote which answer (the reveal happens after judging). 1–10 per round; the bracket count shows how thick the base is.

Axis 2 — Code sims: does it run?

Share of code-round answers whose simulation is runtime-verified OK (renders, animates, no JavaScript error). Harness failures are excluded from the base — a broken cluster node is our fault, not the model's. The failure column shows what went wrong with the rest.

Trend — runnable share per round band

The permanent test keeps growing; this row shows each model's runnable-% over consecutive bands of ten code rounds (a dot means the model was not in the field for that band).

Frontier reference — outside the blind field

Answered the identical prompts, collected through context-free one-shot agents rather than the cluster harness — deliberately unscored on text and shown separately here so they cannot inflate the field.

Generated 2026-07-31 from the published rounds. No votes, no private variants, no composite score — how we measure.