Question 1

How does Jev contribute to the risk score?

Accepted Answer

A typed Jev model reads the context the agent acted on for injected or manipulated instructions; its findings contribute penalties and caps to a code-controlled 0-100 score, and every verdict is signed as an ES256 attestation with a one-hour TTL.

Question 2

What does the signed attestation actually assert?

Accepted Answer

Each JWS lists which checks the provider actually ran -- sanctions, domain, on-chain facts, feeds, simulation, kit watch and model -- alongside caller claims recorded as asserted. The request hash lets a verifier prove the verdict matches the exact request it answers.

Question 3

What does a check cost, and is there a free tier?

Accepted Answer

None. Every evaluation is paid: $0.001 per check on prepaid credits, $0.0035 per call via x402 on Base, or $0.005 when you send a transaction for simulation. Prepaid credits avoid the payment round trip and are the cheapest and fastest route. The provider is live at x402check.xyz with discovery, a DID document and JWKS under .well-known.

Question 4

What do the deterministic checks cover?

Accepted Answer

A match on the OFAC SDN XML, 1,056 digital-currency addresses, scores 0 and critical with no model call at all. MetaMask's phishing list of about 100k hosts and ScamSniffer addresses cap the score at 20, the latter with a 7-day publication lag. A community-flagged domain caps at 40 only when the provider's own look-alike analysis corroborates it.

Question 5

What does the kit watch see that the published lists do not?

Accepted Answer

It reads every Ethereum and Base block as produced and records three families. Look-alike wallets delegating via EIP-7702 to an address-poisoning executor cap at 20, wallets delegating to a labelled sweeper family at 20 or to unlabelled auto-forwarding code at 40, and new contracts running drainer-kit code at 30, which recognises redeployed drainers before their address is ever listed.

Question 6

What does it explicitly not do, and how do I call it?

Accepted Answer

It does not see laundering or other transaction-graph behaviour, and it cannot flag an unknown drainer address that is simply sent funds unless the kit watch has seen it. A clean verdict means no check fired, not that a counterparty is safe, and caller-supplied screening or authorization fields are recorded as asserted and can never lower the score. Call it with POST /v1/risk-check carrying wallet, optional chain, domain and a context field of up to 4,096 characters that is untrusted by design, with keys, seed phrases and credit tokens redacted before the model sees them.

Question 7

What did the reported scale run look like?

Accepted Answer

A 540-call run reported 99.76% at a threshold of 65 to 75 with zero false positives, at roughly $0.00005 per decision and p50 near 400 ms. The README is explicit that measured limits live in docs/EVIDENCE.md, that the demo video predates v0.2.0 and shows superseded v5 corpus numbers, and that a clean verdict means no check fired rather than safe.

X402check

About this project

Highlights

Questions about this entry

Reactions & coverage

More Jev resources like this

From the Jev community