Independent evidence infrastructure for consequential AI decisions
An approval is not oversight.
PromptID is an independent assessor. We examine one consequential decision that AI helped produce, determine whether a competent human genuinely governed it, and issue a finding backed by a sealed evidence record. We assess. We never build or fix what we assess.
- Instrument
- HOP, the Human Oversight Protocol
- Version
- v0.1
- Unit of assessment
- One decision or decision class
- Maintained by
- PromptID Inc.
What we look at
A decision, as it happens
An algorithm flags a prior authorization for denial. A clinician opens the case, reads the recommendation, and signs. The determination reaches the member that afternoon.
Nothing in that sequence is unusual. None of it tells you whether the clinician could have said no.
Health-plan utilization management
An algorithm recommends denying a prior authorization. A clinician signs the determination before it is transmitted.
Credit adverse action
A model declines an application. A credit officer approves the adverse action notice before it goes out.
Insurance claims denial
A model scores a claim and recommends denial. A reviewer signs off before the denial letter issues.
HOP is assessed against one decision or decision class, because that is the level at which liability attaches. If yours is not listed, the test is whether a statute, a regulator or a court expects a human in the loop, and whether you would be asked to prove one was there.
Scope check
Does this apply to you?
Six questions you can answer without talking to us. They are the ones an assessment starts from, and you can run them against your own workflow this afternoon.
Consequence
Does an AI system materially shape a decision that affects someone's money, coverage, credit, care or employment?
A named human
Does a specific person approve or sign it before the consequence takes effect?
A live option
Could that person have reached a different outcome, and would overriding have been survivable for them?
The record
Could you produce what that person saw at the moment they decided? Not the policy. What was on the screen.
The record, later
Could you produce it for one specific decision made fourteen months ago?
The duty
Would a statute, a regulator or a court expect a human in the loop here?
Questions four and five are the ones that decide a determination. A control you cannot evidence is, under challenge, the same as a control that was not there.
The problem
Proof of human oversight does not exist
Three conditions hold wherever AI shapes a consequential decision. None of them is a technology failure. All of them are evidence failures.
A click is not judgment
System logs show that a person approved something. Not that they had the information, the authority, or the time to change it.
Rubber-stamping is indistinguishable from review
Both produce identical records. The difference surfaces in litigation, an audit or a claim dispute, and by then it cannot be remedied retroactively.
Nothing measures oversight sufficiency
Generic AI governance frameworks assess the AI system. Almost nothing tests whether a human could have changed one specific decision.
Standing law
What already applies to you
Neither of these waits on future regulation. One is California law in force today. The other is how Europe's highest court already reads a nominal human step. Both attach to the human, not to the model.
California AB 316
Removes the defense that an AI system acted autonomously. Liability reaches any party that developed, modified or used the system.
Ch. 672, Stats. 2025 · Civ. Code § 1714.46
The SCHUFA ruling
Where a decision draws strongly on an automated output, it remains an automated decision under GDPR Article 22. Formal human involvement does not convert it into a human one unless that involvement is functional.
OQ v Land Hessen · C-634/21 · 7 Dec 2023
The EU AI Act adds a second driver on a known clock. High-risk obligations apply from 2 December 2027 for standalone systems and 2 August 2028 for AI embedded in regulated products. A December 2025 executive order directs the Department of Justice to challenge state AI statutes; none has been brought against AB 316. Regulation strengthens the case. It is not the case. The exposure is liability, and liability does not wait on any single statute.
The instrument
What HOP tests
Four conditions, each a necessary one, each anchored to external authority rather than to a rubric we invented. Every condition, and the authority behind it, is published. The scoring is not.
Positioned
Could they still stop it?
Informed
Did they have what they needed to disagree?
Authorized
Could they override, and survive doing it?
Capable
Did they know the matter, and did they have the time?
The finding
What a determination says
Not a maturity score, a colour or a percentage. One of three findings, in plain words, supported by a frozen and versioned record of the evidence it rests on.
The human was positioned, informed, authorized and capable. The evidence supports it.
An oversight condition was not met. The record shows which one and where it broke.
Oversight may have happened. Nothing survives that could show it. Under challenge that is the same answer as no.
Independence
Nobody grades their own work
Every assurance market ends up separating who does the work from who judges it. AI decisions have not made that separation yet. Most assessment is done by the organization that built the workflow, or by a firm that wants to sell the remediation. PromptID exists to close that gap, which is why we never build or fix what we assess.
Who maintains it
People who measure, not a governance consultancy
HOP is built on binding law, audit-evidence doctrine and published human-factors research. It was written by two people who have spent their careers on one question. Can this person do the work, or do they only hold the credential? One came at it from research. One from 20+ years of judging it at scale.
Dr. Joseph Miller
- DoctoratePhD, Cornell. Research affiliations spanning Berkeley.
- ResearchSupported by DARPA, the National Science Foundation, the Sloan Foundation and the Kavli Foundation.
- IndustryFive years at Bridgewater, leading an intelligence team building algorithmic behavior models.
- MethodPublished researcher on verifying human capability in AI-mediated work (EdArXiv, 2026).
- ContributionThe measurement architecture and the evidence discipline.
April Elias
- Practice20+ years assessing whether people could perform in roles where being wrong is expensive. Executive search, mid-level through C-suite, at some of the world's largest companies.
- MethodStructured evaluation of capability against evidence, at volume and under time pressure, across four technology disruptions.
- FocusWhether a person can do the work, as distinct from whether they hold the credential that says they can.
- ContributionThe competence and capability conditions in HOP.
PromptID has filed formal public comments with the NIST Center for AI Standards and Innovation on the security of AI agent systems, arguing that competent human oversight is a security control current frameworks name but do not specify, and that evidence of it must be behavioral, tamper-resistant and independently verifiable rather than self-reported.
Who authorized this?
Someone will ask that about a decision you have already made. You will have an evidence package or you will have nothing. Twenty minutes on a call and we will tell you whether HOP applies to your decision class.