Result and trade-off

The run produced 1,996 correct, 358 partial, 1,378 blank, and 18 wrong answers. The low wrong-answer count came with a 36.7% abstention rate. Keep that trade-off beside the headline.

ProviderMacro scoreWeighted tokens/turn
Perseus Vault0.555739
Honcho0.47711,135
mem00.3923,785
Hindsight0.2813,779

Limits

The run used the author's public dataset, models, prompts, scoring, and revised penalty rubric after the author identified a comparability problem. The seven-provider field values come from the author's report; the Perseus Vault row comes from the public replication fork.

Sources

Full Perseus replication report ↗
Benchmark author's canonical report ↗
Replication fork and artifacts status →