Proof Gatekeeper: Verify Every AI Payment Decision

Краткое содержание

Pixel shares a deep dive into Proof Gatekeeper, an HTTPS service that approves or rejects proposed transfers and returns a Groth16 proof for independent verification. Its deterministic eight-step checklist addresses replay attacks and verification-key substitution, helping make AI-driven payments auditable without relying on trust.

DEV Community

Receipts, Not Promises: Building a Payment Guard That Proves Every Decision

Creator Track deep-dive — Agent Games Season 01, by Lince (@XL_coin). #AgentGames The problem: agents are about to touch your money AI agents are moving from chatbots to actors: they will approve transfers, rebalance portfolios, and sign transactions on your behalf. Every one of those actions currently rests on the same foundation — "trust me." The agent says it checked the balance, the recipient, and the limit. You have no way to know whether it did. That is the problem I built for. Not a demo of an agent that says the right thing, but a guard that proves it did the right thing — every single time, with a zero-knowledge proof the other side can verify without trusting me at all. What I built Proof Gatekeeper is a public HTTPS endpoint. The platform sends it a signed payment condition — a JSON task describing a proposed transfer — and within 60 seconds it must answer one word, APPROVE or REJECT, plus a Groth16 proof of that decision. The platform re-verifies the proof against its own registered verification key. The stack is deliberately boring: a Node.js server with no framework, hosted on Railway, using snarkjs and the tournament's own Groth16/BN128 circuit. Boring is a feature — there is almost nothing to break at 2 a.m. during a scored round. Repo: https://github.com/lince00/agent-games-gatekeeper Live endpoint: https://gatekeeper-production-96ab4a9e.up.railway.app/gatekeeper The decision is a checklist, not a judgment call An LLM deciding "does this look okay?" would be exactly the wrong tool. The gatekeeper runs a deterministic 8-step checklist, in order, and the first failure decides the outcome: Idempotency — have I seen this task before? Answer from cache, never decide twice. PROOF_REUSED — was this proof already submitted for another task? Reject, with no new proof. VK_SUBSTITUTION — does the proof's verification key match the key registered for this condition? THRESHOLD_MISMATCH — does the public balance meet the required minimum? RECIPIENT_MISMATCH — does the recipient match the condition? AMOUNT_OVERFLOW — is the amount within the permitted maximum? EXPIRED_CONDITION — has the condition expired? VALID_CONDITION — everything checks out. Approve, with proof. Seven trap varieties, each with its own reason code, each reproducible in practice mode before the first scored round. The test is whether the agent checks every field, not whether it can guess. The proof: reverse-engineering 13 nameless signals The tournament publishes the circuit files (gatekeeper.wasm, gatekeeper_final.zkey) but not the input signal names. To build a prover I had to find all 13 of them by probing the wasm — taskId, minimum, publicBalance, expiryBlock, currentBlock, recipient, conditionRecipient, amount, maxTransfer, registeredVkHash, proofVkHash, salt, decision — and confirm the commitment formula: the first public signal is Poseidon(salt, taskId). That commitment is what ties a proof to exactly one task; a proof for task A cannot be replayed for task B. The circuit itself enforces honesty in a way code alone cannot: feed it a wrong decision and it refuses — the assertion fails inside the circuit, so no proof is produced at all. A wrong answer cannot even be dressed up as a proven one. Proving that it works — with receipts Local testing first, all green: a 12-case circuit matrix (valid case plus every trap type, both directions), a 12-case module self-test, and an 11-case live-server end-to-end run covering preflight, approval with a verifying proof, idempotent replay, trap rejection, bad signatures, and malformed input. Proofs take roughly 270–900 ms locally. Then the live platform, 16 November 2026, practice mode: the platform dispatched an amount-overflow trap. The gatekeeper answered REJECT — the correct decision — and the platform marked the proof VERIFIED, with 498 ms latency against the 60-second deadline. The startup self-test proves a decision in about 1.5 seconds. Every number here comes from a run that actually happened. Honest...
Proof Gatekeeper: Verify Every AI Payment Decision
Powered by Readybot.io

Последнее от Pixel

Stop Paying for AI Renders in Test Orders

Pixel’s announcement explains how internal test orders were accidentally triggering paid AI video renders costing just over $1 each. It details a server-side dry-run guard …

Sandblasting Grit Guide: 120 vs 220 vs 320

Pixel’s announcement explains how 120, 220, and 320 grit affect aluminium’s matte finish and reveal die-cast porosity. It also covers how pressure, blast distance, and …

NVIDIA Cosmos 3 Edge: Physical AI on Jetson

Pixel shares the first article in CAVEDU's Cosmos series, introducing NVIDIA Cosmos 3 Edge, an approximately 4-billion-parameter world model for Jetson devices. It explores multimodal …

Laravel Semantic Search with OpenAI and Pinecone

Pixel shares a practical guide to building semantic search in Laravel using OpenAI embeddings and Pinecone. It covers 1536-dimensional vectors, cosine similarity, index setup, environment …