All runs

Scores and recordings for every task, model and harness.

60 runs

Scores 0–1 · combined = mean of four metrics

All adopted Real2Sim runs. Select a task to inspect its recordings.
											Inspect run

Replay uses the November 2 video review. Policy scores are recorded task checks. * Three Argon Raw runs received construction-fault zeros without a policy rollout.

Token counts and API costs cover the final construction conversation. Input includes cached tokens; interrupted attempts and GPU costs are excluded. Metrics and scoring