AgentAnywhere Pramaan is an AI model provenance tool. Its claims checker ties every public statement about a model to a reviewed, unchanged record, and its weights comparison checks whether two models' weights share an origin.

What does Pramaan mean?

Pramaan (प्रमाण) means proof.

How do you check claims made about an AI model?

Find every sentence that makes a claim about the model, tie each one to a record, check that the record's bytes are the ones a reviewer read, and make sure the reviewer is not the author. Pramaan's claims checker does these steps mechanically.

Can you tell whether one AI model was made from another?

If one was made by fine-tuning, further training, compressing or merging the other, their weights usually show it. Pramaan's weights comparison found 72 of 74 such related pairs in its measurements. A model distilled from another shares no weights and cannot be detected this way.

How reliable is the weights comparison?

On 42 small public models it found 72 of 74 related pairs and called none of 368 comparable unrelated pairs related. Those counts bound it at missing up to about one related pair in four and falsely matching up to about one in fifty, on models up to 470 million weights.

Does "no sign of shared weights" mean a model was trained from scratch?

No. It means the two models compared show no sign of sharing weights. It says nothing about any other model, and distillation without copied weights is invisible to the method.

Do you need our model's weights?

No. You can make the fingerprints on your own machine and send only those. If you send the weights, we fingerprint them on a machine rented for the job and destroyed afterwards.

Can Pramaan check decks and documents?

Yes. The claims checker reads PowerPoint, Word and OpenDocument files slide by slide and paragraph by paragraph, including speaker notes, as well as website code and content.

PRAMAAN · प्रमाण · PROOF

Check what is said about a model against the record, and against its weights.

AgentAnywhere Pramaan is an AI model provenance tool with two checks. One ties every public statement about a model to a reviewed record that has not changed. The other compares two models' weights to see whether they share an origin. Ask us to run it.

Diagram of Pramaan's two checks. Left, the claims checker: a sentence such as 'Trained from scratch on licensed text' is checked in four steps: is it in the register; is its evidence unchanged; was it reviewed by a named reviewer who is not a named author, inside the review date; and does the training-data receipt add up. Right, the weights comparison: two named models' weight matrices are compared, each with every matrix of the same shape, and the verdict reads related; a note says no sign of shared weights proves nothing and distillation without copied weights is invisible. Below, the counts: on 42 small public models, 72 of 74 related pairs found and none of 368 comparable unrelated pairs called related, on models up to 470 million weights; it can miss up to about one related pair in four and falsely match up to about one in fifty. — FIG.89Pramaan's two checks: the claims checker (left) and the weights comparison (right), with the counts the weights comparison was measured on. Illustrative examples.

In short

AgentAnywhere Pramaan (प्रमाण, "proof") is an AI model provenance tool: it checks what is said about a model against evidence.

Its claims checker finds the sentences about a model's lineage, training data, licences, evaluation and deployment in a website, a deck or a document, and checks that each is tied to a reviewed record whose bytes have not changed.

Its weights comparison looks at two models' weights and says whether they share an origin. We run it on request.

Why check a model's claims at all?

A sentence like "trained from scratch" or "every licence is recorded" usually rests on the word of the team that built the model. Pramaan replaces that with mechanical checks and a rule that the reviewer is never the builder.

A record you can check is proof. A sentence you have to trust is a promise.

What does the claims checker check?

Four mechanical steps, on every sentence that makes a claim about a model.

1. Is the sentence registered?

Every sentence that matches the claim wording is either in a reviewed register or recorded as not a claim, with a reason.

2. Is the evidence unchanged?

Every registered claim names its evidence, each evidence file exists, and its bytes are the bytes that were reviewed.

3. Did someone else review it?

Every claim has the verdict "supported", a named reviewer who is not a named author, and a review date that has not lapsed.

4. Does the data receipt add up?

A training-data receipt lists every input with a source, a licence and hashes, and the inputs' sizes sum exactly to the training file's size.

What can the claims checker read?

Website code and content (TypeScript and JavaScript strings, JSX, JSON, Markdown, HTML and the text in SVG diagrams), plain text, and decks and documents in PowerPoint, Word and OpenDocument formats, slide by slide and paragraph by paragraph, including speaker notes.

The claims checker has been through eight independent review rounds. We run it on this website's own copy before a change is merged.

How does the weights comparison work?

Two models trained in separate runs end up with unrelated weights, even with the same architecture and the same data. A model fine-tuned from another barely moves. That gap is the signal.

For every weight matrix, Pramaan keeps a small fingerprint: entries sampled at positions fixed by the matrix's shape, its largest singular values, and three summary numbers. Each matrix in one model is compared with every matrix of the same shape in the other, so renamed or reordered layers still match. The result is one page: the verdict in plain words, the numbers, exactly which files were compared, and what the result does not show.

How well has the weights comparison been measured?

On 14 October 2026 it was run on 42 small public models from 27 separate origins, 1,326 pairs in all. It found 72 of 74 related pairs, including 18 of 18 related pairs nobody had altered and 8 of 8 pairs held out from setting its limits, and it called none of 368 comparable unrelated pairs related.

Those counts bound what it can do: it can miss up to about one related pair in four, and falsely match up to about one in fifty. Every figure is on models of up to 470 million weights. A larger model can be compared, and the result says that it is untested at that size.

What Pramaan cannot do

The claims checker

It cannot show that a record is true: a fabricated log with a correct hash passes. A green check means the paperwork is complete and unaltered; whether it is right is the reviewer's job.

It finds sentences that match its claim wording. A claim phrased in a way no pattern anticipates, or text inside an image, is outside what it sees.

The weights comparison

"No sign of shared weights" proves nothing. A model distilled from another, or trained on another's outputs, shares no weights and is invisible here.

It says two models are related; it does not say which came first or that anyone did anything wrong. Someone determined to hide a relation can disguise it, and nothing here has been tested against a determined adversary.

How do we run a comparison for you?

One question, answered with one page: do these two sets of weights share an origin?

The weights can stay with you. You make the fingerprints on your own machine and send only those. If you send the weights instead, we fingerprint them on a machine rented for the job and destroyed afterwards, and keep no copy.

Before we start, we ask who owns each model and whether it may be compared, in writing, and what the answer is for. The note is a technical observation with stated limits, not an expert opinion.

FAQ

Frequently asked questions.

What is Pramaan?: AgentAnywhere Pramaan is an AI model provenance tool. Its claims checker ties every public statement about a model to a reviewed, unchanged record, and its weights comparison checks whether two models' weights share an origin.
What does Pramaan mean?: Pramaan (प्रमाण) means proof.
How do you check claims made about an AI model?: Find every sentence that makes a claim about the model, tie each one to a record, check that the record's bytes are the ones a reviewer read, and make sure the reviewer is not the author. Pramaan's claims checker does these steps mechanically.
Can you tell whether one AI model was made from another?: If one was made by fine-tuning, further training, compressing or merging the other, their weights usually show it. Pramaan's weights comparison found 72 of 74 such related pairs in its measurements. A model distilled from another shares no weights and cannot be detected this way.
How reliable is the weights comparison?: On 42 small public models it found 72 of 74 related pairs and called none of 368 comparable unrelated pairs related. Those counts bound it at missing up to about one related pair in four and falsely matching up to about one in fifty, on models up to 470 million weights.
Does "no sign of shared weights" mean a model was trained from scratch?: No. It means the two models compared show no sign of sharing weights. It says nothing about any other model, and distillation without copied weights is invisible to the method.
Do you need our model's weights?: No. You can make the fingerprints on your own machine and send only those. If you send the weights, we fingerprint them on a machine rented for the job and destroyed afterwards.
Can Pramaan check decks and documents?: Yes. The claims checker reads PowerPoint, Word and OpenDocument files slide by slide and paragraph by paragraph, including speaker notes, as well as website code and content.

Ask us to run it.

Tell us what you want checked: the claims in a site, a deck or a document, or whether two models share an origin. We will tell you what Pramaan can and cannot show for your case before any work starts.

Ask us to run it See the Trust Center