Make “trust our dashboard” something no AI-governance vendor gets to ask for.
Every AI agent that can spend money, send a message, or call a tool on someone’s behalf is a decision that either got checked before it ran or didn’t. Most of the industry answers with a log — a record of what already happened, kept in a database only the vendor can read and only the vendor can edit. That’s not governance, it’s a receipt for the damage. Our goal isn’t to be the AI governance vendor everyone trusts. It’s to be the one nobody has to — because the proof travels with the decision, not with our word.
Four things this company doesn’t compromise on.
Authorize before, not audit after
A log tells you what an agent already did. Provingx decides whether the call is allowed to happen at all, before it reaches the model provider — the block is the product, not the paragraph explaining the damage afterward.
Verify us, don't trust us
Every decision is Ed25519-signed. Anyone can check it offline with nothing but a public key and our published verifier — no Provingx service, no Provingx opinion, involved in the checking.
Anchor evidence somewhere we don't control
The transparency log's root is periodically anchored to the Bitcoin blockchain. To rewrite what a receipt says happened, you would have to rewrite Bitcoin. We picked a ledger we can't quietly edit on purpose.
Publish what breaks
A public page lets anyone try to sneak a call past a real passport or tamper a real signed receipt, against the exact code running in production. Every attempt is counted. We'd rather you find a gap on that page than trust a claim on this one.
These aren’t values on a wall. They’re rules our test suite enforces.
A belief is easy to state and easy to quietly abandon under a deadline. What we actually hold ourselves to is written as engineering practice, checkable in our own commit history rather than just in this paragraph.
Before we trust a regression test as evidence a bug is fixed, we revert the fix, confirm the test goes red for the reason we expect, then restore it. Several bugs in our own history survived a passing test that was quietly asserting nothing.
A response capped at N rows that doesn't say it's capped is a wrong answer wearing a right one's clothes. We fetch one past the limit and tell the caller when an answer is a page, not the whole truth — audits and evidence exports included.
When something can't be verified — a model list we can't reach, a signature we can't check — the honest answer is a refusal, not a plausible-looking guess standing in for one. An invented success is worse than a visible failure.
AI agents were being handed real authority — the ability to spend money, send messages, take actions on someone’s behalf — and the governance around that amounted to a dashboard and a shrug.
“Trust us, we log everything” isn’t authorization. It’s a paper trail for the damage, written after the damage is already done. I wanted the version that decides BEFORE the call runs, signs the decision, anchors the evidence somewhere I can’t quietly edit it later, and lets anyone check it without me in the loop at all.
That’s the whole bet Provingx is built on. Not trust me. Verify me.
— Amit, founder of Provingx
See it hold, not just hear us say it does.
Free to start, no card. Or open the transparency log first and verify a real receipt yourself — offline, with nothing but a public key.