Skip to main content

Model verification

This page is the technical description. For a plain introduction, see Model verification, with guides for buyers and for providers.

A seller says "this is Claude Opus 5.5". Model verification checks that claim. Without it, a seller could advertise an expensive model, serve a cheaper one, and charge the expensive price.

One method among many​

There are many ways to check a seller and protect model integrity, and no single check covers everything. The KBF audit described here is one of the first verification methods Antseed is launching, and more are coming. Each new method will check a different part of what a seller claims, and together they make it hard to cheat without being caught.

The idea in one paragraph​

Every model has a "knowledge boundary": a set of obscure facts it remembers in a particular way, including the ones it gets slightly wrong. A different model or a smaller sibling has a different boundary; a quantized copy of the same model can differ too little to be told apart. The KBF audit asks a seller a batch of short questions with exact numeric answers (for example "the melting point of tantalum carbide is ___°C") and compares the answers with what the genuine model answers. This is called Knowledge Boundary Fingerprinting (KBF).

How one KBF check works​

  1. A reference is built once per model. The genuine model is queried through a trusted upstream: OpenRouter by default, or the model maker's own API (for example Anthropic or OpenAI) when an auditor sets one up. Questions it answers consistently are kept; questions where cheaper look-alikes give the same answers are dropped, because they cannot tell models apart. The reference also measures the genuine model's own error rate, since even the real model is not perfectly consistent.
  2. The seller is asked this epoch's question set. Each auditor builds a new set of at least 100 questions per model and epoch, asks every seller of that model the same set with varied wording, and never uses it again, so a seller cannot memorize a fixed test.
  3. The set is calibrated. Before any seller is asked, the genuine model answers the set in the same batches sellers will get. How often it misses its own stored answers here sets the limit for that set. If it leaves too many unanswered, the model is not audited that epoch.
  4. A statistical test gives the verdict. The seller's wrong answers are compared with that error rate of the genuine model:
VerdictMeaning
SAMENo evidence of substitution: the seller misses about as often as the genuine model does. This is strong evidence, not proof, that it is the advertised model.
DIFFSignificantly more misses than the genuine model makes. The seller very likely serves a different model.
UNDETERMINEDToo few usable answers to decide either way: fewer than half the questions were answered (errors, timeouts, empty or unreadable answers), or a seller looked different but answered fewer than 80% of the questions. A seller that refuses temperature is not UNDETERMINED for that reason: it is asked again without the field and judged on those answers, and the audit's evidence records the dropped field (omittedSamplingFields). A seller that only accepts another temperature value is not audited.

A missing or unreadable answer is left out of the score, not counted as wrong, so a flaky seller doesn't look like a substituted model.

A reference only goes into use if it has enough statistical power to tell the genuine model apart from the look-alikes it was tested against.

Who checks​

Auditors are addresses approved by the verification contract's owner. An auditor buys from sellers like any other buyer, through the normal network and payments, with no special label on its requests.

The audited models are built into the CLI. Each epoch they are split across the approved auditors, round robin, so every model is audited every epoch by one auditor, and the models move to a different auditor the next epoch. For its models, the auditor checks every seller that advertises them. See Which models you audit.

A DIFF from one auditor is only a report. The contract records it once a second approved auditor, the model's confirmer this epoch (the next auditor in the rotation), audits the same seller with its own questions and also finds DIFF. A SAME is recorded from one auditor, and so is the SAME or UNDETERMINED that clears a recorded DIFF on Base mainnet (see below). So a seller is only marked DIFF when two independent auditors agree.

To run an auditor yourself, see Become a verifier. How audits run covers each step, including where the genuine model's answers come from.

Decentralised and trustless​

There is no server that hands out audits, collects results or decides verdicts. The auditors are independent, and what they agree on is worked out from public data and recorded by the contract. Four things make that hold:

  • Nothing to coordinate. Which auditor audits which model, and which auditor confirms which model's DIFFs, follows from the approved auditors (the contract's AuditorSet events), the epoch number and the model list built into the CLI, sorted and rotated round robin. Anyone can recompute the split, and antseed verifier assignment does.
  • Nobody is trusted alone. One auditor's SAME verifies a seller, which only helps the seller. A DIFF, which costs the seller its verified status, is only a report until a second approved auditor audits the same seller with its own questions and also finds DIFF. The second auditor is the model's confirmer, the next in the rotation, so the confirmer changes every epoch and no two auditors can pair up for long. Reports count only in the epoch they were made in and only since the service's last change into or out of DIFF, an auditor's repeated report adds nothing, and a revoked auditor's waiting reports stop counting. Every audit's evidence is on IPFS, so anyone can check what an auditor asked and what the seller answered.
  • Nobody is waited for. The contract takes any approved auditor's audit of any model, so the assignment can't block anything. A seller gets at most two opinions per model and epoch: if a DIFF's confirmer has not answered after 12 hours, the next auditor in the rotation is asked instead, one at a time, and if two auditors disagree nobody else is asked and the SAME stands. Halfway through the epoch, a model with no audit that counts, meaning at least 20% of its sellers got a SAME or DIFF, goes to the next auditor in its rotation that is active this epoch (from three quarters on, to the one after that). A revoked auditor is skipped at once, for confirmations and stand-ins. The stand-in earns the audit's credit. Every auditor reads the time from the chain, so they all switch at the same moment. See If an auditor goes missing.
  • Nothing lasts without being rechecked. A verdict counts for the epoch it was given in and the next one, so a seller stays verified only while auditors keep finding it so, and a missed epoch is bridged by the previous verdict rather than by anyone's say-so.

What is still trusted: the contract owner approves and revokes auditors and sets the report counts (one for SAME and UNDETERMINED, two for DIFF, and the floor for replacing a live DIFF). Those settings are public on-chain, and changing them is visible to everyone.

Public evidence​

Every audit is published to IPFS: the questions asked, the seller's raw responses, the statistics, and where the reference answers came from (which API, which model id and which request settings), with the raw reference requests and responses. Its hash goes on-chain with the result. Anyone can download it and re-run the same kind of check themselves (see Check a model yourself). An auditor that publishes wrong results can be caught from its own evidence and loses its approval.

What goes on-chain​

Text
seller audit ──► evidence on IPFS ──► AntseedKbfAudits ◄── second auditor's audit (for a DIFF)
│ once enough auditors agree
▼
AntseedVerificationResults ──► buyers, points policy

Two contracts share the work:

  • AntseedKbfAudits takes the auditors' KBF submissions, counts the reports for each seller service and records a verdict once enough auditors agree (one for a SAME, two for a DIFF). The owner approves and revokes auditors on it with setAuditor. It earns each auditor credits per audited model: one by default, and the owner can set another default or a per-model amount (creditsPerAudit, setModelCreditsPerAudit). It also decides how long its verdicts count (validityEpochs, 2 by default, at most 4).
  • AntseedVerificationResults holds the recorded verdicts, whatever verification method produced them, and pays verifiers from the verification emissions bucket. A verification method is an audit contract the owner approves on it. KBF is the first audit writing to it; each later method is its own contract next to AntseedKbfAudits, with its own share of the bucket.

Auditor rewards come from KBF's share of the verification bucket. After an epoch ends, each auditor's share follows its credits, capped at 20% of KBF's share per auditor; the part above the cap goes to the emissions reserve in the same claim, as with the other ANTS reward contracts. Rewards are claimed one epoch at a time (claimAuditorReward), and budget nobody claims (for example in an epoch without credits) stays unminted. Rewards stay off until the owner turns them on. Auditors claim with antseed verifier rewards --claim and can publish a display name, icon and link with antseed verifier profile.

These replace the single AntseedKbfVerification contract (migration M003) with the same KBF rules. On Base mainnet the current contracts are AntseedVerificationResults 0x139ab1722a62a23371f491220902c460E9fAeAb2 and AntseedKbfAudits 0x86B92DC48EF7c5a12de6e44409E83C8EF36657Ff (release 003-verification-audits), which the CLI and SDK default to. The points policy that reads them is AntseedVerificationPointsPolicy 0xD58393a752b3b7bdBea6231E0e24C11403e9f32f (release 003-verification-policy). They replaced the earlier 0xF2C8482aD2A348854698E93f41FE3cc32307687c and 0xe57aB03423A3835571F7d68AC1E23CF8a2763343 pair, which they are not compatible with. The old contracts stay on-chain but are no longer read.

AntseedVerificationResults keeps one result per seller service (one model offer of one seller):

  • The latest result wins, whichever approved auditor submitted it. A new SAME replaces an old DIFF and the other way round; an UNDETERMINED clears the earlier verdict.
  • Results expire on their own. Each verdict is recorded with the epoch it stops counting in, chosen by the audit that recorded it. A KBF verdict counts for the epoch it was given in and the next one (about two weeks; the KBF owner can set up to four), so sellers must keep being checked. A seller that fixes its model recovers at the next audit, but after a DIFF that audit is two epochs later: see the next point.
  • For reputation, a DIFF only affects that one service. A seller's other verified models keep their reputation. Rewards work per seller instead: see How results affect rewards.
  • Replacing a DIFF takes one auditor by default. Within KBF, a SAME or UNDETERMINED replaces a live DIFF once it has its own required reports (one each) and at least diffReplacementReports of them. That floor is 1 by default (and on Base mainnet); the owner can raise it, e.g. to 2 so one auditor alone cannot lift a DIFF. If the owner clears a wrong result, reports made before the clear stop counting. The auditors' CLI adds a hold the contract doesn't enforce: a seller isn't audited in the epoch after its DIFF was recorded, so the DIFF stands through that epoch. See After a DIFF.
  • Auditors cannot audit sellers they own.

How buyers use results​

Buyers read each seller service's recorded result from AntseedVerificationResults. A service with a live DIFF keeps 10% of its effective reputation, so automatic routing skips it under the default minimum trust score. Only a recorded DIFF counts: a DIFF from one auditor that is still waiting for the second changes nothing, and the service keeps its earlier result until then. SAME, UNDETERMINED, expired and missing results change nothing. The penalty lifts on its own when the DIFF expires or a later result replaces it (one auditor's SAME or UNDETERMINED under the default diffReplacementReports of 1). See Routing.

How results affect rewards​

Verdicts also change ANTS rewards, through AntseedVerificationPointsPolicy, a points policy registered on Base mainnet. Where reputation looks at one service at a time, the points policy looks at the whole seller agent: it counts the agent's live SAME and DIFF services and gives every settlement with that seller one multiplier, for the seller's points and the buyer's points alike.

  • Each SAME service adds 2%, up to +50% in total.
  • DIFF services cost 4% × (DIFF count)²: −4% for one, −16% for two, −36% for three, −64% for four, and all points from five on.

So a seller with 6 SAME services and no DIFF earns 112% of its points, and one with 6 SAME and 1 DIFF earns 108%. As with reputation, only recorded, live verdicts count, and the effect follows the verdicts: it applies from the next settlement after a verdict is recorded and ends when the verdict expires or is replaced. Points already recorded do not change. See Reward policies for how this policy fits with the others.

Check a model yourself​

Anyone can re-run the latest published audit against a seller or any OpenAI- or Anthropic-compatible endpoint, paying for the questions themselves:

  • antseed verifier check <model> in the CLI finds the latest on-chain audit and runs it, through your buyer proxy or against any URL.
  • @antseed/model-verifier (its antseed-verify command) does the same without the rest of Antseed.
  • The antseed-verify-model agent skill asks for an endpoint, has you set its API key, runs the check and explains the result.