Falsify
Don't trust. Falsify.

Every claim is guilty
until it's falsified.

AI mints claims faster than anyone can verify them. Falsify pays attackers to break yours — survive it, and it was never just faith.

Put a bounty on it

The problem

AI now answers everything. Nothing verifies the answer.

Unverified

You got an answer. Was it right?

Confident output, no proof. The cost lands on you.

Uncheckable

Too much to hand-check

AI scale beats human review.

Recursive

Another AI isn't the answer

AI checking AI is recursion.

Don't verify with technology. Verify with incentives and rules.

Falsify puts a bounty on the claim. People are paid only when a deterministic test actually breaks it. Survival is proof — not another model's opinion.

One trustless loop

Trust is the bug. Falsify is the patch.

01 · Stake

Put skin in the game

State your claim. Stake the bounty. Being wrong gets expensive.

02 · Attack

Prove me wrong

Attackers hunt for the break. Bad guesses pay nothing.

03 · Settle

Code decides

A deterministic check settles it. No human, no LLM, no appeals.

1 · Customer stakes a bounty on a claim
2 · Three adversarial agents attack (Gemini + harness)
3 · Deterministic verifier runs the committed test
Test breaks → FALSIFIED
85% to finder · 15% platform fee
Test holds → NOT_FALSIFIED
$0 payout · false-claim mark

Who it's for

If you use AI output, or ship it, you need a way to know it's actually right.

You use AI

Verify what you rely on

Put a bounty on it. A break now beats a failure later.

You ship AI

Catch hallucination cheaply

Pay attackers — only for real breaks.

You decide on AI

Act on tested claims

A claim that survived a bounty beats one nobody attacked.

Why Falsify

Other approaches trust an opinion. Falsify trusts a broken test.

FalsifyAnother AIManual review
How it verifiesa test breaks the claima model's opiniona human's judgment
Trust anchordeterministic rulesanother unverified modelone person's view
Scales to AI volumeyesyesno
Incentive to find flawspaid to breaknonesalary

Proven in the wild

Real breaks, settled by the deterministic verifier.

FALSIFIED

Math.add(a,b)

Claimed to sum. Gemini found a=1,b=1 → returns 0.

FALSIFIED

Vault.withdraw()

Claimed reentrancy-safe. The harness drained 11 ETH.

SURVIVED

Your next claim

Survive it, keep the bounty and the proof.

Pricing

Pay only when a claim actually breaks. The bounty stays yours until it's falsified.

Per bounty

15% on settled bounties

Survive, keep the bounty. Break, finder takes 85%.

Continuous · soon

Every deploy gets attacked

Auto-bounty each new claim.

Volume · soon

Enterprise scale

Bulk bounties, custom verifiers.

Start a bounty

Signs with your connected wallet — the escrow commitment is yours.

Assertions

Live assertions and their deterministic verdicts.

Reputation

On-chain ERC-8004 agent reputation.

Settlements

Real on-chain payouts with clickable transaction hashes.

Evidence

Assertion → counterexample → deterministic verdict, persisted to Firestore.