Benchmarks

The public benchmark asks how many verified human outcomes an Agent organization delivers for the human attention it consumes. It evaluates the whole path from first prompt through delivery, authority and recovery—not one model response.

It measures effectiveness, human decisions and repair, elapsed time, tool and token use, duplication, crash recovery, chain of custody and missing evidence. Failed and incomplete attempts remain in the series. Missing telemetry is unknown, never silently zero.

Published baseline

The repository’s current human-readable report covers five declared quickstart-to-delivery attempts: three passed, one failed and one ended incomplete. The final two accepted passes delivered in 16m 59s and 16m 16s with zero Captain follow-up turns, repair interventions or manual operational actions. A held-out controlled interruption resumed accepted work without lost changes, duplicate effects or human repair.

Those numbers are scoped evidence, not a general production claim. Follow the result links to the exact subject, environment, revisions, limitations and sanitized artifacts.

Evaluation freezes evidence. Improvement begins only in a separate run after the bundle is fixed.