ProvanisAI PROVENANCE GATEWAY
AI provenance gateway

Every AI request.
Its decision.

Provanis sits between your people and the model, decides every request, and gives you a receipt you can hand to a client.

6
fields in every record
97%
precision, measured
5+
languages of IDs detected
0
raw prompts stored
The gap

A dashboard counts tokens. A client asks a question.

What was sent

Nobody records the prompt that left, so nobody can show it later.

Who allowed it

A token count cannot name the person who said yes.

What you can show

Without a receipt, the answer is reconstructed from memory.

The route

Four steps. One request.

Nothing reaches the model before step three decides.

1

Ingress

Route resolves, key checked, identity recorded.

2

Inspect

Typed segments, local detectors, no cloud.

3

Decide

Allow, redact, review or block.

4

Record

Six fields written before the reply returns.

It sits in the path, not on the side

Your people and agents on one side, the models on the other. One base URL per tool.

Live product

The console, in your language

These are real screens — the language switcher changes the product, not just this page.

Provanis console

The receipt

You can prove what was sent.

Six fields, hash-chained, verifiable offline — without trusting this server.

1

Identity

Who sent it, and how strongly proven.

2

Message

The typed segments it decomposed into.

3

Decision

Allowed, redacted, blocked, or by whom.

Edit a row, break the chain

Every event carries the hash of the one before it. Tampering is detectable, not just discouraged.

It learns

Feedback makes it sharper.

Mark a false flag, and the next similar request is allowed. Report a miss, and it goes to whoever writes the rules.

Mark as false detection

A review that was not necessary becomes a suppression precedent. High-severity identity numbers are never auto-allowed.

Mark as missing detection

A report for the rule authors — it never changes past decisions on its own.

It learns your judgement, not a model's

Every human decision becomes a precedent, cited in the receipt, spot-checked, and auto-disabled if it drifts.

What it is not

Not a DLP. Not a firewall. Provenance.

CapabilityDLPFirewallLLM routerProvanis
Inspects the prompt——partly✓
Decides per request——routing✓
Named approval with expiry———✓
Per-request receipt, verifiable offline———✓
Sees files, email, USB✓partly——

The receipt for your AI

That is the category, in one line. If the boundary fits your firm, the pilot takes an afternoon.

Measured

Accuracy you can argue with.

On a public corpus, scored against Microsoft Presidio. The misses are printed, not hidden.

97.6%
precision
41.5%
recall, with NLP
96%
benign left alone
484
tests passing

Method and harness are published — re-run it on your own prompts before you sign.

Pilot

One afternoon, one of your prompts.

Bring the request you would least like to explain tomorrow. That is the one that shows whether the receipt is real.

FAQ

Asked before the pilot.

Does it cover desktop chat?

No. A gateway only sees traffic that goes through it. Desktop chat with no endpoint, files, email and clipboard are out of scope — we say so before you buy.

Does it make me compliant?

No. It produces evidence. It does not certify a model, a prompt or a firm.

Where does my data live?

On your own infrastructure. Detectors run locally. Raw prompts are not stored.

How many languages?

The console ships in English, Traditional and Simplified Chinese, and Japanese. Detection covers HK, CN, TW, JP, KR and US identifiers.