What does an Inferstamp proof actually prove?

That a specific answer follows from a specific set of model weights and a specific input, under the exact arithmetic of the circuit. The weights and the input stay private. Only short fingerprints of them appear in the receipt.

What does it not prove?

It does not prove the answer is correct, fair or useful, that the input describes reality, or that the provider did nothing else besides the proved computation. It also says nothing about how the input was collected or preprocessed.

Does checking a proof require access to the model?

No. Verification needs only the public receipt, the agreed fingerprints and a public verification key. Model weights are never sent to the verifier, which is why a provider can prove a result without disclosing its model.

How long does verification take?

Checking is fast and roughly constant: tens of milliseconds in a browser for the small circuit on this site, measured on 17 September 2026. Producing a proof is the expensive side, and that cost grows with model size.

Does this work for large language models?

Not today. The working prototype proves a small fixed-point linear model. Proving a full LLM forward pass is far beyond what proving systems do economically right now. The realistic near-term targets are small scoring, ranking and eligibility models.

Proof that the answer came from
the model you agreed on.

An AI answer is just a claim. Inferstamp attaches a receipt that anyone can check in about a second — without trusting the provider, and without ever seeing the model's weights.

Check a proof yourself How it works ↗

Three steps. Nothing taken on faith.

Agree on the model

Both sides lock in which model and which input are in play. Each becomes a short fingerprint. The weights stay with whoever owns them.

Get the answer with a receipt

The provider runs the model and returns the result plus a small proof file. The receipt carries no weights, no input values, no secrets.

Check it yourself

The check is arithmetic, not reputation. Swap the model, alter the input or edit the answer and it fails — on your machine, in milliseconds.

Try to fool it.

Below is a real proof, recorded earlier from test data. The check itself runs now, in your browser. Change the answer and watch it break.

The claim

Test data

Example Claimed answer

Type any number here, then run the check again.

The agreed fingerprints

Model: Loading…
Input: Loading…
Recorded: Loading…

Also rewrite the answer inside the proof Makes the check fail on the mathematics rather than on the mismatch.

The check

● Runs on your device

Loading the example

Fetching the public files. No private data is needed.

Nothing is sent to a server

This checks a result recorded earlier, so it proves the mathematics — not that a provider ran the model a second ago. Why that distinction matters ↗

Look inside the receipt

Fingerprints, the result, the request details and the proof itself. No weights, no input values, no secrets.

Download this receipt ↓

Where a receipt pays for itself.

Directions we are exploring with clients.
All use cases ↗

You paid for the big model

An invoice says one model; a fingerprint shows which one actually answered. Version labels are promises — a commitment can be checked.

Decisions people can contest

Scoring, pricing and eligibility calls where someone may later ask what produced this number, and deserves more than a log line.

A model you cannot show

Prove a score follows from your weights while keeping them private. The receipt carries a fingerprint, never the model.

Updates you want to trust

When a model is retrained, check that the new weights came from the agreed update rule instead of accepting a new checkpoint on trust.

The idea in 56 secondsSOUND + CAPTIONS

What it does not prove.

A proof ties an answer to a model and an input. It says nothing about whether that answer is good — a proved prediction from a bad model is still a bad prediction. It cannot tell you the input was true, that the data was collected honestly, or that the provider did nothing else on the side.

Today's working version proves a small numeric model, not a language model. We would rather say that up front than have you discover it in week three of a project.

The full mechanics, limits and measurements ↗

Would this hold up
in your workflow?

Start with one model and one decision. We work out what you could actually prove, what it would cost, and whether something simpler would do the job.

€500 fixed-scope assessment

A working session on one model and one decision
A written statement of what can and cannot be proved
An honest comparison with signed receipts, re-running the model, and hardware attestation
A prototype scope and estimate, with the assumptions written down

No production integration is included, and we will tell you when proofs are the wrong tool. If we build the prototype afterwards, the assessment fee comes off that work.

Proof that the answer came fromthe model you agreed on.