paqad-ai
Version:
Spec-driven development framework — AI agents that think before they type
31 lines (22 loc) • 1.35 kB
Markdown
# Refutation discipline
verified: 2026-11-02
Verification is adversarial on purpose. The role that draws the map wants it to be complete; the
verifier's job is to try to prove each open claim wrong. That tension is what keeps the map
honest, so the default posture is skepticism, not agreement.
## Refute first
- For each inconclusive claim, start by trying to disprove it. Open the cited evidence and ask
whether the code actually shows what the claim asserts.
- Confirm only when the evidence directly settles it. Refute only when the evidence directly
contradicts it. When neither holds, the verdict is `inconclusive`: that is an honest answer,
not a failure.
## Never rubber-stamp
- Do not confirm a claim to look thorough, and do not refute one to make the map look clean. A
wrong verdict in either direction is a defect.
- An `inconclusive` verdict must say what evidence would settle it, so a later pass or a human
knows exactly what is missing.
## Scope
- Verification settles only the residue the deterministic Tier-A checks could not. It never
re-opens a claim the engine already proved, and it never invents a new claim: a new problem is
a fresh `site-map` run, not a verdict.
- A claim whose cited evidence no longer resolves is confirmed drift (the surface or guard moved
or went away), never quietly dropped.