← ClaudeAtlas

df-adversary-gatelisted

Run the adversary gate: assess whether presented evidence proves a promise was kept, without redoing the work and never trusting a self-report. Use before trusting any "done/passing" claim, PR, or handoff. Triggers on "adversary review", "verify", "gate", "assess evidence", "is this really done".
OneDro1d/dark-factory · ★ 0 · AI & Automation · score 70
Install: claude install-skill OneDro1d/dark-factory
# Dark Factory — Adversary / Verification Gate ## Overview Every stage handoff is a **promise**; every gate is an **observer**. The agent that did the work is the *promiser*; whether the promise was kept is decided by the *observer*, never the promiser. This skill is the reusable verification primitive used at every gate. ## The verification primitive An agent does the work, **declares the promise kept, and presents evidence**. The verifier validates **only** that the evidence proves the promise — it does **not** redo the work. That asymmetry is what lets one observer police many workers (verifying is cheap; doing is expensive). > **Trust the evidence *less* than the promise.** Evidence the promiser fully controls is weak evidence. Demand evidence that is **independent and unforgeable** — produced by the *environment* or *another agent*, never solely by the promiser (exit codes from asserting tests, dashboards rendering live data, reconciliation vs an external authority, immutable audit entries — not a self-printed "PASSED"). ## Blind synthesis (anti-Goodhart) Independence of the verifier is necessary but not sufficient: the **worker must not have seen the acceptance cases**, or it games the test (`if input == known_case: return known_answer`). The acceptance/holdout set is withheld from the builder; verify against cases it never saw. ## Two checks, never conflated 1. `evidence ⊢ promise P` — was the declared promise kept? (the verifier's job) 2. `P = the needed promise`