Can you prove what your AI did?

Two questions get very hard the moment AI is in the loop, and both are really the same question wearing different clothes:

• Did this come from where it claims? — the deck, the report, the image an agent produced.
• What did the agent actually do? — at 3am, in a workflow no human watched.

Today the honest answer to both is “trust me.” Logs are mutable. Screenshots are fakeable. A dashboard your vendor controls is only as good as your faith in the vendor. None of that survives an audit, a dispute, or a regulator — and it definitely doesn't survive an adversary with an incentive to rewrite history.

Make it provable, not promised

The fix is old and boring and works: cryptographic commitments. You don't need to publish the content to prove things about it — you publish a hash, and hashes are cheap, private, and impossible to forge backwards.

VAIBot applies this in two places.

Content receipts

The Provenance API takes a piece of content, hashes it (keccak256 for anything anchored on-chain, sha256 for local and offline use), and anchors that hash on a public chain. What comes back is a receipt anyone can verify: this exact artifact existed, in this exact form, at this time, and has not been altered since. The content itself never leaves your hands — only the fingerprint does. Anchoring is batched by a background service so you get on-chain finality without paying per-item gas.

Action receipts

The other half is what the agent did. Every decision the guard makes — allow, ask, deny — lands in a hash-chained, tamper-evident audit log, and governance decisions produce receipts of their own. Because each entry commits to the one before it, you can't quietly delete or reorder the awkward parts; a break in the chain is detectable. “What happened and in what order” stops being a matter of trust.

Why this is worth the trouble

• Compliance — show that your agent operated under an enforced policy, with a record you didn't author after the fact.
• Disputes — settle “the AI produced this” / “no it didn't” with a receipt instead of a screenshot.
• Authenticity — give downstream consumers a way to check that an AI artifact is the one you actually signed off on.
• Tamper-evidence — a log that can't be edited without leaving a mark changes the incentives of everyone who touches it.

What a receipt does and doesn't claim

Precision matters here, so: provenance proves integrity, timing, and origin— that a specific artifact or event existed, when, and that it wasn't altered. It does notprove the content is correct, safe, or true. It is an authenticity layer, not a truth oracle. But “we can prove exactly what was produced and exactly what was done” is a foundation almost no AI stack has today, and most of the trust problems above can't be solved without it.

Governance decides what your AI is allowed to do; provenance proves what it did. Get started, or read the enforcement side in the circuit-breaker post.