The approval happened. Can you prove it?
Somewhere in your systems there is a log line that says a human approved an AI agent's action — a payment released, a record changed, a message sent to a customer. The log line is the whole of the evidence. It records that an account was signed in and a button was pressed. It does not record who was actually there, whether they meant this specific action, or whether they understood what they were approving. When the action turns out to be wrong — or fraudulent — the log line is asked to carry weight it was never built to carry, and it folds.
This is the gap the phrase proof of human intent names: not a record that approval was clicked, but evidence that a specific human, at a specific moment, performed a deliberate act of intent toward a specific action — evidence that still holds up when the vendor, the session, and the account are all gone.
Why the click collapsed the evidence
The signature used to be an event. A person, in the presence of witnesses, performed a practiced physical act on a specific document — and the act itself was the evidence, examinable long afterward. Digital approval went backwards: the typed name and the click-to-approve collapsed that event into a database entry, a claim written by the same system that benefits from it being believed. AI agents make the collapse dangerous, because the scale changed. One compromised session or one rubber-stamped queue no longer mis-approves one document; it can mis-approve a thousand agent actions before anyone reads a log.
Human-in-the-loop only means something if the human's presence in the loop is provable. Otherwise "a human approved it" is a sentence a system says about itself.
What proof of human intent actually requires
Protocol One's standard answers with its first and fifth laws of Identity Physics: no identity without an event, and every chain starts and ends with a human — intent opens the chain, confirmation closes it, machines carry the middle. Applied to an AI approval, that means the approval is captured as a human event, not a click:
Intent, stated as an act. The approver performs the first part of a Tri-Signature event — three handwritten signatures inside one 10-second burst: the first signature is the intent to act, made toward this action and no other. A performed signature is evidence in a way a pressed button is not: it has dynamics, timing, and a hand behind it.
Identity, bound to the same event. The second signature is who is performing it, captured in the same burst, so intent and identity cannot be separated and recombined across sessions.
The action, fingerprinted. What is being approved is sealed with DocLock at the moment of approval — its fingerprint and counts fixed — so the proof covers this payment, this record, this message, and cannot be quietly re-pointed at a different one afterward.
When, attested. The confirmation signature closes the event, and the time is sealed at that confirmation — when becomes part of the proof, not a timestamp a server can be asked to restate.
The result is a Forever Seal in its Simple form: one flat, complete evidence package for the approval — who, what, and when — that anyone holding it can verify offline, with no account, no vendor server, and no permission. That last part is not a feature detail. An approval proof that can only be checked inside the vendor's own dashboard is the log line wearing a better suit.
The toolkit that operationalizes it
The AI Agent HITL Governance Toolkit is the working shelf for this pattern: how to place the human event in an agent workflow so that no action executes until who, what, where, and when are verified first — Protocol One's second law, no action without proof — and how to keep the resulting evidence packages so an auditor, an insurer, or a court is reading sealed events rather than system logs. The wider library, including the document-proof and workstation-verification pieces, is gathered in the Learn hub at protocolone.store/pages/learn.
Honest limits
Proof of human intent proves a human performed a deliberate approval event toward a specific action. It does not prove the approval was wise, informed in substance, or free of pressure — a person can sign intent toward a bad decision, and the seal will faithfully prove they did. It is evidence about presence and intent, not a judgment on either.
It is also not a compliance certificate. Human-oversight requirements — in the EU AI Act and elsewhere — are legal questions with legal tests; a sealed approval event is strong raw material for meeting them, and no product page can honestly promise it satisfies them. Have counsel map the pattern to your obligations before you rely on it.
And the capture quality varies by device: a stylus records richer signature dynamics than a fingertip on glass. The honest claim is that the sealed event preserves the evidence an examiner needs — never that any single check is unforgeable or court-proof.
Protocol One LLC — Standards Office, 3707 E Southern Ave, Suite 1023, Mesa, AZ 85206. {^}