proofkit — proof-of-work verification for AI agents

The #1 way autonomous agents fail is quietly: they report success they never achieved. ok=True with no side effect. A message_id: 0 "delivery". A random simulator dressed as a metric. A dry-run marked "done". proofkit is the adversarial layer that assumes your agent is lying until a real artifact proves otherwise.

Built and battle-tested running a 39-agent autonomous fleet at $0/month.

Why this exists (the pain)

~60% of failed agent deployments share one root cause: unverified success. The agent's log says it worked. The code says return True. Nobody checked the actual side effect. proofkit turns "it looks done" into "here is the live artifact that proves it ran."

Free tier — the scanner + the method

1. Static fake-success scan

python3 proofkit/verify_real_scan.py <file> [<file> ...]

Flags the tells that let code report success without doing the work:

• return True / hardcoded success=True / message_id: 0
• random.* simulators returning fabricated "metrics"
• TODO / NotImplementedError / empty pass bodies
• demo/sample/fixture data on a production path
• except: pass that swallows an error then reports success
• dry-run defaults masquerading as live runs

Precise by design (word-boundary matched — exist_ok=True won't trip it), and it prints the honest caveat: a clean static scan is necessary, not sufficient.

2. The /verify-real method (3 steps)

1. Static pre-scan (above).
2. Demand a LIVE artifact — real stdout numbers, a non-zero message_id, a file that now exists + its bytes, an HTTP 200 from a served page, a real DB row that changed. Use a tripwire: assert the downstream never runs if a false pass would ship.
3. Verdict: VERIFIED LIVE / BUILT-BUT-UNVERIFIED / FAKE-BROKEN — with the artifact as evidence. Never call it done because it looks done.

Premium tier — the enforcement suite

Everything free, plus the parts that make it a standing guarantee, not a manual pass:

• /redteam — adversarial multi-pass verifier: N independent skeptics per claim, each prompted to refute, kill on majority. Follows the data one layer down (the "real API" wrapper is where the fake usually hides).
• gatecheck — 6+ guard-invariant proofs (kill-switch reachable, trading blocked, injection defense intact, quality gate fail-closed) as a PASS/DRIFT table you run before every ship.
• Tripwire harness generator — auto-wraps a build step so a false "pass" is impossible: the downstream throws unless the real artifact exists.
• CI hook — fail the pipeline on any fake-success flag or missing live artifact; proof-of-work required to merge.
• Fleet mode — run the whole suite across N agents, ranked report of who's shipping fake success.

Use cases

• Autonomous posting/trading/outreach agents (prove the post/order/email really happened)
• Multi-agent fleets where one agent's fake success corrupts the next's input
• CI for AI-generated code (block stubs + fabricated results from merging)
• Any "the agent said it worked" moment you can't afford to trust blind

Install

Copy verify_real_scan.py into your project (free) or install the full skill from ClawHub for the premium enforcement suite. Zero paid dependencies — pure Python stdlib + your existing local model for the adversarial pass.