ground-truth-gateslisted
Install: claude install-skill firaen22/claude-rulebook
# Ground-Truth Gates
**The rule: no gate → no change to judgment-bearing behavior.** Behavioral
prompting ("be careful with this regex") does nothing measurable. A harness that
runs the REAL function over labeled REAL data converts "sounds right" into a
falsifiable pass/fail — the highest-ROI artifact in any project; build it BEFORE
tuning prompts, regexes, or classifier logic, not after the regression.
If this skill ever conflicts with `~/.claude/harness/`, the harness wins (R8).
## The three gates
| Gate | Locks | Mechanism |
|---|---|---|
| **Golden** | a classifier's decisions | labeled real cases run through the real function (no mocks), scored cost-asymmetrically |
| **Replay** | a pure transform's output | frozen baseline, diffed on every run; any drifted line fails |
| **Project** | everything at once | one command (`run-all.sh`) = typecheck + golden + replay; run it before claiming "done" |
## Quick start — bundled runnable template
The template lives in this skill's `scripts/` directory and is **run-verified**
(all pass AND fail paths executed 2026-07-07; the commands below are the exact
ones that ran). Zero dependencies beyond `node` ≥18.
```bash
cd ~/.claude/skills/ground-truth-gates/scripts
# whole project gate (typecheck optional: export TYPECHECK_CMD="npx tsc --noEmit")
bash run-all.sh # → ALL GATES PASS, exit 0
# golden gate alone
node golden-gate.mjs --module ./example/classifier.mjs --cases ./example/cases.jsonl
# r