What problem does it solve? Agent policies, Skills, and prompts often look correct on paper but fail when a fresh agent interprets them in a cold context. This Skill provides a rigorous methodology to verify that policy changes actually produce the intended behavior, instead of relying on the author's own re-reading or plausible-sounding answers. ## Core Features & Use Cases - Deterministic vs. latent separation: Classify each rule into machine-checkable constraints (file existence, schema, command executability) versus judgment-dependent behavior that requires cold evaluation with a fresh agent. - Grader self-controls: Enforce negative, regression, and positive control fixtures so a grader that cannot separate known-good from known-bad answers is never trusted as quality evidence. - Comparative latent evaluation: Run paired baseline-vs-candidate comparisons with condition parity, blind judging, identity pinning, and hard-failure release gates. - Context budget monitoring: Track always-on instruction growth against a checked-in baseline via bash evals/policy-evaluation/context-budget.sh. - Use Case: After rewriting a routing rule in your agent policy, use this Skill to decide whether the change needs a repo-local eval, build scenario fixtures with controls, run a cold-context eval on a fresh agent, and report completion evidence including model identity and known limitations. ## Quick Start Ask the agent to evaluate whether your latest policy or prompt change is correctly interpreted by a fresh agent using deterministic checks and cold-context evals.
