What problem does it solve?
When a skill corpus degrades on a new model, rewriting what looks wrong produces plausible fixes with no way to know whether any of them mattered. This Skill replaces guesswork with a measurement-first protocol: mine the run archive for a baseline, establish a noise floor with A/A runs, register a bar before any change, then cut in attributable passes until the bar clears.
Core Features & Use Cases
- Archive baseline mining: Extracts per-run traces, token rates, and phase markers from historical runs to build an outcome taxonomy and localize where runs die.
- Noise floor and registered bar: Runs two identical corpus builds through one harness to measure run-to-run spread, then registers a written bar before any edit exists.
- Adversarial corpus audit: Dispatches independent proposer and defender agents per unit so cuts require a real search of tests, docs, and git history, with protected categories like security guards and machine-readable strings.
- Halt taxonomy and cut passes: Classifies ten halt mechanisms (phantom handoffs, fictional seams, turn-ending output rules) and applies one problem class per pass with disjoint file ownership.
- Use Case: A team ports their 31-skill plugin to a new model and sees runs stalling at one phase boundary. This Skill mines 458 archived runs to localize the defect, proves the noise floor with A/A runs, then lands measured cut passes until eight consecutive clean runs clear the registered bar.
Quick Start
Ask the agent to retune the skills corpus in ./skills for the new target model, starting by checking that a benchmark harness with a build selector and run archive exists.
