Keep-or-discard harness loop

The idea

A harness change is kept only if the declared capability metrics stay within tolerance, one efficiency metric improves by the minimum, and frozen held-out cases confirm it. Campaigns count every attempt.

The problem
Cheaper harness changes can quietly lose capability.
Why it matters
A cheaper failure is still a failure, so only measured gains are recommended.
Where it stands
In production as of

A step-by-step replay of this idea is not published yet. The library adds one when a recorded run shows it.

Where it lives in the product

  • Run waste analysis

    When a run ends, the platform analyses its steps and requests for repeats, missed fusions, context growth, large outputs and long logs. It estimates the tokens each mechanism would save.

  • Self-improving template harness

    After work finishes, two proposers edit the template's harness from the run ledger. Each candidate is probed on the same request. A composer and a blind selector pick one, which a person adopts (or auto-adopt does, if the checks pass).

Turn the numbers into shipped work.

Crewshal runs these choices for you: a persistent AI worker with memory and rules, on your Claude and Codex subscriptions.