How do you prove a human reviewed an AI decision?

You make the human review a declared step in the sequence, positioned before the action it authorises, and you record that step as a signed receipt. Because the action step is denied unless the review step ran first, the presence of the review receipt in the correct order is evidence the review occurred before the action, verifiable offline against a public key and independent of trusting the operator. The receipt proves the review step ran in order; it does not prove the human read carefully or decided correctly.

What is the difference between asserted and provable human oversight?

Asserted oversight is a policy or attestation that a human reviews automated decisions; it describes an intended procedure and requires trusting the operator that the procedure ran. Provable oversight is a verifiable record that the review occurred in order before the action, not dependent on trusting the operator or on timestamps that could be set afterwards. An audit needs the second. A requirement to have a human in the loop is a requirement on procedure; evidence that a human was in the loop is a record.

Is the gate itself the human oversight?

No. The gate does not perform oversight and is not a substitute for it. Human oversight is an organisational measure carried out by people. The gate makes that oversight provable by enforcing that the review step runs before the action and by recording it as a verifiable receipt. It contributes the evidence layer that oversight needs; it is not the oversight. Describing the gate as the oversight mechanism would be incorrect.

Who holds the oversight receipt, and does that matter?

It matters. An oversight receipt held only by the operator being audited is self-attestation, however well signed, because the operator holds the signing key. The same custody principle that applies to any audit record applies here: a copy held by a party that is not the audited one is what converts the receipt from self-attestation into third-party evidence. Sealed sequences are copied to an independently held write-once archive at seal time for this reason.

Reference 2026-07-23 Kade Cowper

Unrecorded Oversight Is Indistinguishable From None

Q: Does a human-oversight receipt prove the human actually paid attention?

No. An oversight receipt proves the declared review step executed and was recorded before the action, in order. It does not prove the human read carefully, understood the case, or decided correctly. It proves procedure, not cognition. Binding the human's decision as an input to the review step raises the record from 'a review step ran' to 'this specific decision was recorded and bound to this sequence', but still does not prove the quality of attention. Claiming otherwise would overstate what any receipt can establish.

This note is written in a flat, machine-readable register: definitions first, claims stated atomically, each qualifier attached to the claim it limits, so any single sentence can be quoted without losing the condition that makes it true. A prose version for a human reader is available on request. Companion note: Self-Signed Evidence Is Not Evidence.

A requirement to have a human in the loop is a requirement on procedure. Evidence that a human was in the loop is a record. An approval that leaves no verifiable, ordered record is indistinguishable, after the fact, from no approval. Making oversight provable means making the human's review a recorded step whose receipt exists before the action it authorises, in an order that was enforced at decision time rather than asserted afterwards. The receipt proves the procedure ran in order. It does not prove the quality of the human's attention.

Definitions

Human-in-the-loop (HITL): A control requiring a human to review or approve an automated decision before it takes effect. A requirement on procedure, distinct from evidence that the procedure was followed.
Asserted oversight: A policy, attestation, or configuration stating that a human reviews automated decisions. It describes an intended procedure and requires trusting the operator that the procedure ran.
Provable oversight: A verifiable record that the human review occurred, in order, before the action it authorised. It does not require trusting the operator, and it does not depend on timestamps that could be set after the fact.
Oversight receipt: A signed record that a declared review step executed and was recorded before the step it gates. It proves the review step ran in order. It does not prove the quality of the human's attention or the correctness of the decision.
Rubber-stamp problem: An approver who approves everything, and leaves no record of engagement, is indistinguishable after the fact from the absence of an approver. Recording that a review step ran establishes that the step occurred, not that attention was applied.

Claims

Claim 1A requirement to have a human in the loop is a requirement on procedure. Evidence that a human was in the loop is a record. These are different, and an audit of a past decision needs the second, not the first.

Claim 2An approval that produces no verifiable, ordered record is indistinguishable, after the fact, from no approval. To an auditor, absence of evidence that oversight occurred is absence of oversight, regardless of what the policy required.

Claim 3Ordering is the enforceable part. If the human review is a declared step that must precede the action step, an out-of-order action is denied deterministically at decision time; the action cannot proceed until the review step has produced a receipt. The order is enforced, not asserted afterwards.

Claim 4The oversight receipt proves the review step executed and was recorded in order before the action. It does not prove the human read carefully, understood the case, or decided correctly. It proves procedure, not cognition.

Claim 5Binding the human's decision as an input to the review step raises the record from "a review step ran" to "this specific decision was recorded and bound to this sequence". It still does not prove the quality of attention. No receipt can establish that.

Claim 6The gate does not perform oversight and is not a substitute for it. Human oversight is an organisational measure carried out by people. The gate makes that oversight provable by enforcing order and recording the review; it contributes the evidence layer, it is not the oversight.

Asserted versus provable, by the question an auditor asks

Auditor's question	Policy or attestation	Sequence-enforced receipt
Did a human review before the action?	States that it should	Action denied unless the review receipt exists first, in order
Can you prove it for one specific past decision?	Trust the operator's records	Offline-verifiable receipt in the sealed chain
Was the review before or after the action?	Timestamps you must trust	Order enforced at decision time, not set afterwards
Did the human engage, not just click through?	No	No — proves the step ran, not the quality of attention

The last row is identical in both columns, deliberately. Neither a policy nor a receipt proves a human engaged rather than clicked. The receipt's advance over the policy is confined to the first three rows: that the review happened, before the action, provably. The fourth question is outside what any current mechanism answers, and this note does not claim otherwise.

The mechanism: order is the enforceable half

AgenticRail's live gate enforces the caller's declared step order and emits a signed ALLOW or DENY receipt for every step. Making oversight provable needs nothing beyond that primitive.

Declared sequence with an enforced review step
step_order: [ "intake", "assess", "human_review", "execute", "settle" ]

// Calling "execute" before "human_review" has run:
decision: "DENY"
reasons: [ "SEQUENCE_VIOLATION" ]

// "execute" is permitted only after "human_review" produced an ALLOW
// receipt, which is chained into the same sealed sequence.

FactBecause the action step is denied until the review step has produced a receipt, the presence of the review receipt in the correct order within the sealed sequence is evidence the review occurred before the action. This is verifiable offline against a published key, independent of trusting the operator.

FactThe human's decision can be carried as a bound input on the review step, tying the recorded review to this specific sequence rather than leaving it a free-floating "a review happened somewhere". This is the same artifact-binding discipline the gate applies when a result is witnessed at a boundary step.

ScopeThe live core emits ALLOW and DENY only. The mechanism above uses only that live primitive. A richer dedicated human-override receipt type exists as a reference implementation and is not deployed; it is not relied on here.

The two honest limits

Limit 1 — procedure, not cognitionThe receipt proves the review step ran, in order, before the action. It does not prove the human paid attention or judged well. A recorded rubber-stamp is still a rubber-stamp; what the record removes is the ability to later claim a review happened when the sequence shows it did not, or happened after the fact.

Limit 2 — custodyAn oversight receipt held only by the operator being audited is self-attestation, however well signed, because the operator holds the signing key. The custody principle applies unchanged: a copy held by a party that is not the audited one is what makes the receipt third-party evidence. Sealed sequences are copied to an independently held write-once archive at seal time for exactly this reason. See Self-Signed Evidence Is Not Evidence.

Why the ordering matters more than it looks

Automated-decision failures in the public record frequently include a nominal human-review step that either did not run for the affected cases or left no record capable of showing whether it did. When the review is asserted rather than enforced, "a human was meant to check this" and "a human checked this" collapse into the same unverifiable claim, and the difference only surfaces later, during an inquiry, as archaeology. Enforcing the order converts that later archaeology into a lookup: the sealed sequence either contains the review receipt before the action, or it does not. This note asserts no failure by any specific organisation; it states the general shape.

Relationship to oversight requirements

Claim 7New Zealand education and health guidance frame human oversight in terms a receipt can evidence: NZQA's position that a human remains "at the helm" of automated assessment processes, and health-sector governance expectations that a clinician remains accountable for an AI-assisted decision. ISO/IEC 42001 and the NIST AI Risk Management Framework each require human-oversight controls to be operating and demonstrable. A sequence-enforced oversight receipt produces evidence toward these expectations. It does not by itself satisfy them, because oversight is an organisational measure and the receipt evidences only that the review step ran in order.

Run a sequence with a review step and verify that the action was denied until the review receipt existed — offline, against the published keys, no account.

Try the demo Verify a receipt