Proof over trust
A zero-knowledge proof ties an answer to one committed set of model weights. We port the prover and never touch the judge: every proof is checked by the unmodified upstream verifier.
The evidence layer for AI agents. Cryptographic proof of what a model computed, made on hardware you can buy, and receipts that state exactly how much they prove.
AI agents now act: they call tools, move data and approve changes. The record of what they did is mostly logs. We believe every action deserves evidence, and every piece of evidence should say how much it proves.
WHY makes AI show its receipts. Every line of a receipt says how much it proves.
The agent asked to delete files
SEENThe safety check said “block”
PROVENThe block was carried out
SEENThe receipt has not been altered since
SEALEDSEENsomeone saw itSEALEDany change would showPROVENshown with mathematics
A log line says what an operator claims. A proof says what a computation did. Our work starts from four rules.
A zero-knowledge proof ties an answer to one committed set of model weights. We port the prover and never touch the judge: every proof is checked by the unmodified upstream verifier.
A proof about some input is not a proof about your tool call. A public encoder turns the call into the model's input, and the checker re-encodes it byte for byte.
Observed, tamper-evident, attested, proven, validated. Each claim gets the level of its weakest link, and the checker refuses anything graded higher than its evidence.
Software only: no secure enclave, no datacenter. Our proofs are made on a 16 GB GeForce RTX 5070 Ti in an ordinary desktop PC, and checked on a CPU.
Open tools for checking AI, not trusting it. Everything ships with its logs, its failed tests and its limits.
Anthropic's open-source FlashProve prover, ported to a store-bought RTX 5070 Ti with 12 small patches. Zero changes to the verifier, the protocol or the proof format.
Open source · CheckerSigned receipts for an agent's tool calls, and an independent checker that grades every claim by its weakest link. The receipts and logs are in the repository, and the proofs are in its release.
Experiment · Guard modelA 260K-parameter guard we trained and converted ourselves. It answers allow or block for a tool call, and each answer is proved in about 2 to 3.6 seconds.
Method · VocabularyOne vocabulary for how much a record proves, from “a gateway saw it” to “independently validated”. Never a bare “verified”.
Proofs are expensive, so we aim them where they pay: small models whose decisions someone else must be able to check.
A small guard model allows or blocks each agent action, and proves which decision it made for that exact call.
Demonstrated on synthetic dataLet a counterparty check a decision without seeing the model. The proof system is designed to hide the weights; we have not yet verified that for our port.
Open questionRecords an auditor, insurer or counterparty can re-check for themselves, each stating its level and why it is not higher.
Working prototypeWhich model answered? A commitment ID ties every proof to one set of weights, so a model cannot be swapped quietly between decisions.
Demonstrated on tiny modelsPatches, scripts, logs, checksums and proofs are public. Anyone can rebuild the verifier and re-check our results.
Open sourceHow much does it actually prove? Five levels, never merged. A claim gets the level of its weakest link.
Solid: demonstrated · Dashed: not yet
A trusted gateway saw it happen and wrote it down. Most agent records today stop here.
send_email with these arguments,” or “the block was enforced.”Hashed, signed or chained, so you can tell if it changed. A fact about what the gateway said, not about the world.
A validated hardware or provider attestation. It moves trust from the operator to a chip vendor; it does not remove it.
A valid proof of an exactly specified computation. It says which answer a committed model computed for this exact input. It does not say the answer is correct.
BLOCK for this delete_files call.”Independent tests show the result meets explicit acceptance criteria. The only level that says a decision was right.
It never trusts the issuer. Try to fool it.
Our checker recomputes every link itself: it runs the verifier, re-encodes the tool call and re-applies the decision map. Each forgery below is signed with a key the checker trusts, so a valid signature earns nothing.
A replay of six of the twelve cases in our public tamper suite, with the checker's messages shortened. All twelve are caught: see the log.
Credits, not partners: the open projects and standards our results stand on. No endorsement is implied.
Measured, not estimated. Every dot is a proof.
Every number here comes from a logged run on one desktop PC, and the logs are public. Nothing on this page is a benchmark claim.
See the evidenceFresh proofs from our GPU port, accepted by the unmodified verifier.
Flipped bits, wrong IDs, another model's key. All refused.
Each proved answer equals the ordinary model's.
Re-run from a fresh clone of the public repository.
Proving 64 requests at once at Llama 4 Maverick's layer width, on a 16 GB card. Synthetic weights, 4 layers.
Evidence ↗To prove one request on tiny models. A guard decision takes about 2 to 3.6 seconds.
Evidence ↗On held-out synthetic tool calls, and 94.1% on names it never saw. A synthetic policy: a demo, not a product.
Evidence ↗The port is 12 small prover patches, +34 / −28 lines. The judge stays exactly as upstream wrote it.
Evidence ↗Everything we have run so far. Each proving and evidence experiment has a stated success condition, a stop condition, raw logs and a list of what it does not show. Negative results stay in.
Which experiences deserve to change an agent's future behaviour? A controlled simulation of 18 memory policies. Our original score did not beat the designated baseline, and we published that.
Code and paperFlashProve's CUDA prover targets a datacenter H100. With 12 small patches it runs on an RTX 5070 Ti: 3 of 3 fresh proofs accepted by the unchanged verifier, 18 of 18 tampered inputs rejected.
A clean clone, the published patches and a separately built verifier reproduce the result on the same machine. The verifier binary is bit-identical across three builds.
The production build proves the next token of TinyStories260K. The answer is token 310 in every run: the same as FlashProve-L4's published result on different hardware, and the floating-point model's own choice.
Layers at Llama 4 Maverick's full width commit in about 43 s and prove in 3 to 6 s. Full batches of 64 requests peak at 9.3 GiB. All 750 proofs in the batch study were accepted.
Research note 01Under fixed seeds our build is byte-deterministic, which makes a byte-for-byte comparison with unmodified upstream possible. Upstream's lookup test passes on our card, and all 18 of its cheating-prover cases are refused.
A receipt format with a level on every claim, and a checker that recomputes the level and refuses overclaims. 12 of 12 tamper cases caught, including a dishonest issuer who swaps proofs.
A public encoder closes the gap between “the model judged some input” and “the model judged this call”. The whole claim now grades L4, and a reused proof from another call is caught.
Our own converter reproduces FlashProve-L4's model file byte for byte. The guard scores 99.2% on held-out synthetic calls; 40 of 40 decisions are proved; 39 of 40 are right, and the wrong one is on the record.
Research note 02Independent replication on another machine. A byte-for-byte comparison with unmodified upstream on an H100. Native Linux and RTX 30/40 cards. Larger guards on real policies. Verifiers for L3 and L5.
Help us replicateWhat this does not show.
We think this section matters more than the results. If a claim is not on this page or in a repository log, we are not making it.
Don't take our word for it.
Clone the repositories, run the checker, rebuild the proofs on your own GPU. Then tell us what broke. If you run agents in production and need decisions your auditors can check without trusting you, talk to us.
Replicate our results