forge-calibratelisted

# forge-calibrate A reviewer you have never measured is a reviewer you are trusting on vibes. This skill turns "the review feels inconsistent" into two numbers — **recall** against planted findings and **agreement** between repeated runs — and a list of what the reviewer never sees. Run it before and after changing a lens, an agent prompt, a rule card, or the confidence gate; without the before/after pair, a prompt change is a guess. ## Argument - `--runs N` — how many times each case is reviewed (default 3; 5 for a decision you'll act on). - `--case <name>` — only that golden case. - `--focus=<lenses>` — passed through to the review. - `seed` — copy the bundled golden set into the project (see Golden set) and stop. ## Golden set A case is a directory with `before/` and `after/` trees (the diff is `diff -ruN before after`), an `expected.txt` of finding keys (`<lens>:<file>:<tag>`, one per line, `#` comments allowed), and optionally a `context/` with the spec / standards / `rules.txt` the case depends on. - Bundled: `${CLAUDE_PLUGIN_ROOT}/skills/forge-calibrate/golden/` — eight cases, one per failure class: swallowed error, scope creep + unrequested config, hollow test, silent breakage of an untouched caller, rule-card violation that step 0 must catch by ID; and three **gate cases** for `forge-gatekeeper` — hardcoded live token (step 0 must catch it), missing object-level authorization (IDOR), irreversible migration shipped with its code change. - Project: `<