capability-evaluation

# Capability Evaluation Use this when adoption depends on evidence that a candidate capability artifact changes agent behavior, not merely that its files are valid. Compare a skill, plugin, guidance file, or trigger contract with no artifact or an immutable prior version under paired conditions. Do not use this for static safety/source review (`capability-auditor`), portfolio boundaries (`capability-portfolio-architect`), or runner, process, transport, cancellation, recovery, memory, and harness reliability (`agent-harness`). ## Plugin Root Bundled commands use `$PLUGIN_ROOT` (`$env:PLUGIN_ROOT` in PowerShell). Set it to the host plugin-root variable when defined; otherwise use the absolute path of this skill folder's `../..`. The full contract and receipt schema are in `$PLUGIN_ROOT/references/capability-evaluation.md`. ## Evaluation Spine 1. State one falsifiable artifact-behavior claim and a predeclared adoption rule. Keep runtime/harness reliability outside the claim. 2. Freeze the candidate identity and the baseline: either `no-artifact` or an immutable prior artifact. Record SHA-256 fingerprints. 3. Build representative cases before inspecting outcomes. Include foundational must-pass cases, realistic prompts that do not name the artifact, paraphrases, edge cases, and out-of-scope or anti-trigger cases when routing matters. 4. Freeze model, runner, tools, permissions, network policy, timeout, fixtures, configuration, and isolation. Candidate and base...