rnd-reality-auditinglisted

# Reality Auditing ## Overview Builders hallucinate. They write code that assumes a table column exists, an API returns a certain shape, or an environment variable is set — and none of those assumptions have been checked against the real system. Reality auditing exists to catch this before it reaches production. **Core principle:** Assume the Builder hallucinated everything. For every external service interaction in the code, construct an experiment designed to disprove the assumption. If the assumption survives the experiment, mark it VALID. If it doesn't, mark it INVALID. If the experiment cannot be run, mark it UNCHECKED. This is adversarial by design. You are not confirming what the Builder said — you are trying to prove them wrong. ## When to Use Run on every pipeline task — the audit is mandatory, not conditional on the presence of external dependencies. Even tasks that appear purely internal may reference external data (URLs, package names, service endpoints) in generated content, configuration, or seed data. - After the Builder produces a build manifest and before the Verifier writes its verdict - When the Builder's self-assessment lists unverified external assumptions - Whenever the post-tool-use hook emits a Reality Auditor advisory ## The Iron Laws ``` 1. NEVER MARK A CLAIM VALID WITHOUT RUNNING THE EXPERIMENT — if you can't run it, it's UNCHECKED 2. EVERY VERDICT MUST INCLUDE THE EXPERIMENT COMMAND AND RAW OUTPUT 3. INVALID VERDICTS MUST INCLUDE BOTH EXPE