# Inference proofs > An open benchmark for provable inference on open-weights models. Systems are compared within one model rung. Passing the six checks does not certify AI safety. Read each system's scope, sample sizes, grades, evidence tiers and source records before citing a result. The deployment article discusses hypothetical agreements, not an enacted treaty. ## Pages - [Checks · Inference proofs](https://www.inferenceproofs.ai/checks): The rules behind a row of the ladder: the six checks, each figure, the order and the glyph's five petals. - [The six checks](https://www.inferenceproofs.ai/checks#h-checks) - [Figures](https://www.inferenceproofs.ai/checks#h-figures) - [Order](https://www.inferenceproofs.ai/checks#h-order) - [Glyph](https://www.inferenceproofs.ai/checks#h-glyph) - [Systems not run](https://www.inferenceproofs.ai/checks#h-not-run) - [AI treaty verification and deployments · Inference proofs](https://www.inferenceproofs.ai/deployment): How proofs of inference could support a hypothetical US-China AI agreement: model commitments, sampled replies, trusted traffic records and verification limits. - [Rerun, enclave or proof](https://www.inferenceproofs.ai/deployment#rerun-enclave-or-proof) - [What a relying party supplies](https://www.inferenceproofs.ai/deployment#what-a-relying-party-supplies) - [In an agreement](https://www.inferenceproofs.ai/deployment#in-an-agreement) - [FlashProve · Inference proofs](https://www.inferenceproofs.ai/flashprove): FlashProve on Llama 4 Maverick 17B-128E Instruct, passed. Cost vs serving: 1,317×. - [Glyph](https://www.inferenceproofs.ai/flashprove#h-glyph) - [Statement](https://www.inferenceproofs.ai/flashprove#h-statement) - [Checks](https://www.inferenceproofs.ai/flashprove#h-checks) - [Cost](https://www.inferenceproofs.ai/flashprove#h-cost) - [Fidelity](https://www.inferenceproofs.ai/flashprove#h-fidelity) - [Sessions](https://www.inferenceproofs.ai/flashprove#h-sessions) - [Proving a forward pass · Inference proofs](https://www.inferenceproofs.ai/forward-pass): The verifier tests the prover's committed values at random points. Matrix products with the committed weights are the cheap part to prove. - [Testing a product at random](https://www.inferenceproofs.ai/forward-pass#testing-a-product-at-random) - [Fiat-Shamir](https://www.inferenceproofs.ai/forward-pass#fiat-shamir) - [From one product to a layer](https://www.inferenceproofs.ai/forward-pass#from-one-product-to-a-layer) - [Lookups](https://www.inferenceproofs.ai/forward-pass#lookups) - [One pass for a whole reply](https://www.inferenceproofs.ai/forward-pass#one-pass-for-a-whole-reply) - [Zero knowledge](https://www.inferenceproofs.ai/forward-pass#zero-knowledge) - [What an accepted proof rests on](https://www.inferenceproofs.ai/forward-pass#what-an-accepted-proof-rests-on) - [Provable inference glossary · Inference proofs](https://www.inferenceproofs.ai/glossary): Definitions for proof of inference, zero knowledge, soundness, model commitments, quantization, fidelity and trusted execution environments. - [A](https://www.inferenceproofs.ai/glossary#h-l-a) - [B](https://www.inferenceproofs.ai/glossary#h-l-b) - [C](https://www.inferenceproofs.ai/glossary#h-l-c) - [E](https://www.inferenceproofs.ai/glossary#h-l-e) - [F](https://www.inferenceproofs.ai/glossary#h-l-f) - [G](https://www.inferenceproofs.ai/glossary#h-l-g) - [H](https://www.inferenceproofs.ai/glossary#h-l-h) - [I](https://www.inferenceproofs.ai/glossary#h-l-i) - [K](https://www.inferenceproofs.ai/glossary#h-l-k) - [L](https://www.inferenceproofs.ai/glossary#h-l-l) - [M](https://www.inferenceproofs.ai/glossary#h-l-m) - [N](https://www.inferenceproofs.ai/glossary#h-l-n) - [O](https://www.inferenceproofs.ai/glossary#h-l-o) - [P](https://www.inferenceproofs.ai/glossary#h-l-p) - [T](https://www.inferenceproofs.ai/glossary#h-l-t) - [Q](https://www.inferenceproofs.ai/glossary#h-l-q) - [R](https://www.inferenceproofs.ai/glossary#h-l-r) - [S](https://www.inferenceproofs.ai/glossary#h-l-s) - [V](https://www.inferenceproofs.ai/glossary#h-l-v) - [W](https://www.inferenceproofs.ai/glossary#h-l-w) - [Z](https://www.inferenceproofs.ai/glossary#h-l-z) - [How a check works · Inference proofs](https://www.inferenceproofs.ai/how-a-check-works): One request for every system, a commitment, a proof and a verifier without the weights, six checks drawn one by one, and the fraction behind the headline figure. - [The request](https://www.inferenceproofs.ai/how-a-check-works#the-request) - [Commit, prove, verify](https://www.inferenceproofs.ai/how-a-check-works#commit-prove-verify) - [Six checks](https://www.inferenceproofs.ai/how-a-check-works#six-checks) - [The figure](https://www.inferenceproofs.ai/how-a-check-works#the-figure) - [Provable inference benchmark · Inference proofs](https://www.inferenceproofs.ai/): Compare proof-of-inference systems on open-weights models: six checks, proving cost vs serving, proof size, verification time and fidelity. - [Llama 4 Maverick 17B-128E Instruct](https://www.inferenceproofs.ai/#h-llama-4-maverick) - [Llama 3.1 8B Instruct](https://www.inferenceproofs.ai/#h-llama-3.1-8b) - [GPT-2](https://www.inferenceproofs.ai/#h-gpt-2) - [Systems not run](https://www.inferenceproofs.ai/#h-notrun) - [Learn about proofs of inference · Inference proofs](https://www.inferenceproofs.ai/learn): Learn what a proof of inference shows, how verification works, and how to interpret benchmark costs, model fidelity and AI treaty proposals. - [For policy readers](https://www.inferenceproofs.ai/learn#h-policy) - [For prover teams and engineers](https://www.inferenceproofs.ai/learn#h-engineering) - [Glossary](https://www.inferenceproofs.ai/learn#h-terms) - [Mandate · Inference proofs](https://www.inferenceproofs.ai/mandate): Who runs Inference proofs, what it commits to and how to check its figures. - [Who runs it](https://www.inferenceproofs.ai/mandate#h-who-runs-it) - [Ties](https://www.inferenceproofs.ai/mandate#h-ties) - [What we do](https://www.inferenceproofs.ai/mandate#h-what-we-do) - [What we do not do](https://www.inferenceproofs.ai/mandate#h-what-we-do-not-do) - [How to check us](https://www.inferenceproofs.ai/mandate#h-how-to-check-us) - [Changes and corrections](https://www.inferenceproofs.ai/mandate#h-changes-and-corrections) - [How we compare · Inference proofs](https://www.inferenceproofs.ai/method): Every system gets the same request. Those passing the six checks, with a stated security level, are ranked by cost vs serving inside one model. - [Fixed before the runs](https://www.inferenceproofs.ai/method#fixed-before-the-runs) - [The checks](https://www.inferenceproofs.ai/method#the-checks) - [The figure](https://www.inferenceproofs.ai/method#the-figure) - [Order and the glyph](https://www.inferenceproofs.ai/method#order-and-the-glyph) - [Judgement and changes](https://www.inferenceproofs.ai/method#judgement-and-changes) - [Getting a row](https://www.inferenceproofs.ai/method#getting-a-row) - [What a proof shows · Inference proofs](https://www.inferenceproofs.ai/proof-of-inference): A proof shows a reply is the committed model's greedy reply to its prompt, not whose model it is or that the reply was served. - [Outside the proof](https://www.inferenceproofs.ai/proof-of-inference#outside-the-proof) - [Today's systems](https://www.inferenceproofs.ai/proof-of-inference#today-s-systems) - [Passed](https://www.inferenceproofs.ai/proof-of-inference#passed) - [SASH · Inference proofs](https://www.inferenceproofs.ai/sash): SASH on Llama 3.1 8B Instruct, passed. Cost vs serving: 62,160×. - [Glyph](https://www.inferenceproofs.ai/sash#h-glyph) - [Statement](https://www.inferenceproofs.ai/sash#h-statement) - [Checks](https://www.inferenceproofs.ai/sash#h-checks) - [Cost](https://www.inferenceproofs.ai/sash#h-cost) - [Fidelity](https://www.inferenceproofs.ai/sash#h-fidelity) - [Sessions](https://www.inferenceproofs.ai/sash#h-sessions) - [Scale · Inference proofs](https://www.inferenceproofs.ai/scale): Every system run, on a log axis: cost vs serving, proof size, verify time. - [Cost vs serving](https://www.inferenceproofs.ai/scale#h-multiple) - [Proof size](https://www.inferenceproofs.ai/scale#h-proof-size) - [Verify time](https://www.inferenceproofs.ai/scale#h-check-time) - [The statement · Inference proofs](https://www.inferenceproofs.ai/statement): The ladder's verifiers check a committed, quantized copy's greedy reply. No row has a proved error bound that covers what it runs. - [Each system's statement](https://www.inferenceproofs.ai/statement#each-system-s-statement) - [The committed weights](https://www.inferenceproofs.ai/statement#the-committed-weights) - [Integer arithmetic](https://www.inferenceproofs.ai/statement#integer-arithmetic) - [Checked tokens](https://www.inferenceproofs.ai/statement#checked-tokens) - [The seams](https://www.inferenceproofs.ai/statement#the-seams) - [The error bound](https://www.inferenceproofs.ai/statement#the-error-bound) - [Claim cards](https://www.inferenceproofs.ai/statement#claim-cards) ## Records - [Source repository](https://github.com/ai-verification-observatory/inference-proofs) - [Level rules](https://github.com/ai-verification-observatory/inference-proofs/blob/main/levels/L1.yaml) - [Specification](https://github.com/ai-verification-observatory/inference-proofs/blob/main/docs/SPEC.md) - [Sitemap](https://www.inferenceproofs.ai/sitemap.xml)