The protocol

HOP, the Human Oversight Protocol

A decision-level determination of whether a competent human was positioned, informed, authorized and capable of materially changing a consequential AI-mediated decision before consequence occurred, and whether sufficient evidence exists to reconstruct that oversight afterward.

The four conditions

HOP evaluates four conditions in sequence. Each is a necessary condition. None is sufficient alone. Each condition is anchored to external authority rather than to an internally invented rubric.

What is published and what is not: every condition, and the authority behind it, is set out below. The dimensions, thresholds and gate logic behind a determination are controlled methodology and are not published. We disclose the construct we test, never the scoring mechanics. Publishing the mechanics would let a client engineer the result, which is the failure mode every self-assessment framework already has.

Positioned

Was the human placed at a point in the workflow where intervention was still possible before consequence occurred, and could the process actually be stopped?

Anchored to
Cal. Civ. Code § 1714.46
CJEU C-634/21
EU AI Act Art. 14

Informed

Did they receive what they needed to disagree, including the basis for the output, its confidence, and its documented failure modes?

Anchored to
EU AI Act Art. 14(4)
EDPS TechDispatch 2/2025

Authorized

Did they hold actual authority to override, and was overriding a survivable act inside the organization rather than a career cost?

Anchored to
CJEU C-634/21
EU AI Act Art. 26(2)

Capable

Were they competent in the specific matter, and was the workload compatible with judgment rather than throughput?

Anchored to
EU AI Act Art. 26(2)
EU AI Act Art. 14(4)(b)

Table of authorities

An instrument that cites nothing is a rubric with opinions in it. Every criterion in HOP traces to binding law, regulatory guidance, audit-evidence doctrine or published research. This is the crosswalk.

Authority	What it establishes	Where it binds
California AB 316Ch. 672, Stats. 2025 · Civ. Code § 1714.46 · in force	Removes the defense that an AI system acted autonomously. Liability reaches any party that developed, modified or used the system.	All four conditions. Removes autonomy as a defense, so the organization that used the system remains answerable for the outcome.
CJEU — OQ v Land HessenC-634/21 · ECLI:EU:C:2023:957 · 7 Dec 2023	Nominal human involvement does not remove a decision from meaningful-oversight requirements. Involvement must be functional.	Positioned and Authorized. Separates functional involvement from ceremonial sign-off.
EU AI Act, Article 14Reg. (EU) 2024/1689, as amended · binds providers · high-risk from 2 Dec 2027	Effective human oversight. Overseers must be able to understand the output, interpret it correctly, and intervene or interrupt.	Positioned and Informed. Sets the oversight the system must make possible. The duty to staff it sits with the deployer under Article 26.
EU AI Act, Article 26(2)Reg. (EU) 2024/1689, as amended · binds deployers	Deployers must assign human oversight to natural persons who are competent, trained and supported with the necessary authority.	Authorized and Capable. Supplies the competence and authority criteria.
EDPS TechDispatch #2/2025Human Oversight of Automated Decision-Making · 23 Sep 2025	Regulatory criteria for what makes oversight of automated decision-making effective rather than nominal.	All four conditions. Informs criteria design across the instrument.
PCAOB AS 1105, as amendedRelease 2024-007 · adopted 12 Jun 2024	System-generated information does not self-authenticate. It must be independently evaluated before it is relied on as evidence.	The evidence model. This is why an approval log is not, by itself, proof of oversight.
IAASB ISA 500 · AICPA AU-C 500Audit evidence	Sufficiency and appropriateness of audit evidence, and the discipline of binding conclusions to the evidence obtained.	Evidence caps. A finding cannot exceed what the evidence supports.
Human-factors researchPublished literature on automation bias and vigilance	Documented degradation of human scrutiny over automated recommendations, particularly under time pressure and high volume.	Capable. Supplies the workload and vigilance criteria.

Evidence model

HOP distinguishes two kinds of evidence item, and they fail differently. Collapsing them is the most common error in AI governance assessment, because it hides the more expensive problem.

Type one

Control conditions

Whether the safeguard existed at all. A hold that pauses issuance pending review. Authority to override without escalation. A reviewer qualified for the matter. If a control condition was absent, the oversight did not happen.

Type two

Evidence records

Whether anything survives that could prove it. Per-case timestamps rather than batch. A log of what was displayed at the moment of decision. Override records retained through the limitation period. If these are absent, the oversight cannot be reconstructed.

A control failure and an evidence failure need different remedies. One is a workflow problem. The other is an evidence-architecture problem, and it is the one organizations discover last, because nothing surfaces it until someone demands the record.

Determinations

A HOP Determination is not a maturity score, a colour or a percentage. It resolves to one of three findings, stated plainly, and is supported by a frozen, versioned Decision Record. Every determination pins the method version it was issued under.

Finding 01

Sufficient

The human was positioned, informed, authorized and capable. The evidence proves it, and the record is complete enough that the basis of the finding can be reconstructed years from now.

Finding 02

Insufficient

At least one oversight condition was not in place. This is a control failure. The record does not need to improve. The workflow does.

Finding 03

Insufficient due to evidence gap

The controls may have been in place. Nothing survives that could prove it. Under challenge this is the same answer as no. It is also the finding that most evidence architectures are currently set up to produce.

Independence rules

Independence is a structural bar, not a policy preference. It is the reason a determination survives contact with someone who is motivated to discredit it.

No implementation: PromptID does not build, configure or remediate the controls it assesses. We identify gaps and reassess. We do not fix them for a fee.
No contingent outcome: Fees are not contingent on the finding. An assessor who is paid more for a favourable determination is not an assessor.
Evidence-capped: Findings are bound by the evidence produced. A client cannot argue its way to a stronger result.
Separation of duties: The person who collects evidence is not the sole person who determines against it.
Measured reliability: Agreement between assessors is measured, not assumed. Assessors work the same evidence independently and their determinations are compared. The instrument’s agreement statistics are published.
Frozen record: The record supporting a determination is versioned and sealed, so the basis of a finding cannot be edited after the fact without the edit being obvious.

Scope of the instrument

HOP has a deliberately narrow unit of assurance. Knowing what an instrument does not test is part of knowing what its findings are worth.

What it assesses: Human oversight of one consequential AI-mediated decision or decision class, at the level where liability attaches.
What it does not: HOP does not certify the AI system, validate a model, or assess an organization's overall AI maturity. Those are different questions and they need different instruments.
Where it applies: Workflows in which a human retains consequential oversight. Where authority is delegated to an autonomous system with no human approving each transaction, a different protocol is required.

Bring us a decision

The fastest way to know whether HOP applies is to describe one decision class and let us tell you plainly. Twenty minutes.

See a specimen determination Book a call