A pharmaceutical company submitted a regulatory filing last year. The filing included sections drafted with AI assistance. The company’s quality system required human review before submission. A senior team member reviewed the document and approved it.
Six months later, a regulatory auditor found three claims in the filing that were not supported by the cited source documents. The auditor asked for the review record. The company produced an email showing the senior team member’s approval.
The auditor’s next question: “What did the reviewer check? What standard did they apply? Is there a record of what they found?”
There was no record. The reviewer had read the document. They had approved it. They had not documented what they verified, what they found, or what standard they applied. The signoff existed. The control surface did not.
The Signoff Without a Surface
Most organizations have a human-in-the-loop claim. Few have a human-in-the-loop process.
The phrase “human in the loop” has become a compliance checkbox. A human reviewed it. A human approved it. Therefore, the output is verified.
This logic fails under scrutiny for a specific reason: it conflates approval with verification. Approval is a decision. Verification is a process. The decision without the process is a rubber stamp.
In expert work, the distinction matters more than in most AI applications. A researcher’s paper is reviewed by peers who check specific citations. An expert witness report is challenged by opposing counsel who probes specific methodology claims. A grant proposal is scored by a panel that flags specific inconsistencies. A compliance filing is examined by an auditor who asks for specific evidence of review.
In each case, the external scrutineer is not asking “did someone approve this?” They are asking “what was verified, what was found, and where is the record?”
The human signoff that cannot answer those questions is the signoff that fails under scrutiny. Not because the human was negligent. Because the process gave them nothing to produce.
What a Control Surface Is
The control surface is the structured point between AI production and external delivery where verification happens, decisions are made, and records are created.
A control surface has four components:
1. Defined verification scope. What is being checked? Not “the document.” Specific elements: citation accuracy, claim-source alignment, internal consistency, methodology-design fit, regulatory compliance. The scope determines what the reviewer examines and, equally important, what is outside scope for this review.
2. Defined standard. Against what criteria? A citation is not just “checked.” It is checked for existence (does the DOI resolve?), polarity (does the literature treat this paper as supporting or contradicting?), and alignment (does the paper’s finding match the specific claim?). A methodology section is checked for internal consistency, scope boundaries, and design fit. The standard determines what “pass” and “fail” mean.
3. Decision authority. Who makes the decision? Approve, correct, or escalate. The reviewer who checks citations may not be the same person who approves the final document. The escalation path, what happens when a flagged item cannot be resolved at the reviewer level, must be defined before the review begins.
4. Record production. What artifact does the signoff produce? At minimum: what was checked, what was found, what was corrected, who reviewed it, and when the signoff occurred. This record is the deliverable that makes the entire process auditable.
Without all four, the signoff is a point in time without a process around it. The reviewer reads the document, forms a judgment, and sends an email. The judgment may be sound. The email is not a verification record.
Why Expert Work Demands This
The standard for AI-assisted expert work is not “did someone approve it?” It is “can you defend the approval under adversarial scrutiny?”
Expert work faces a category of scrutiny that most AI applications do not. Consider the four domains:
Research. A peer reviewer checks the citations attached to central claims. If a cited paper does not support the claim, the reviewer questions the entire literature review. The verification record must show that citations were checked for alignment, not just existence.
Litigation. Opposing counsel files a Daubert challenge against the expert report. The challenge probes whether the methodology is reliable and the conclusions are supported. After Kohls v. Ellison, where an AI expert’s report was excluded for hallucinated citations, the verification trail is the defense. 156 documented lawyer sanctions for AI hallucinations confirm that courts are scrutinizing AI-assisted expert work specifically.
Grants. A review panel scores the proposal on significance, approach, and innovation. NIH notice NOT-OD-25-132 states that grants will be terminated if AI content is detected without proper documentation. The verification record is compliance documentation, not optional quality assurance.
Compliance. An auditor asks who verified the AI output in a regulatory filing. The verification record must trace each claim to a source, each source to a check, and each check to a reviewer. Email approvals do not satisfy this requirement.
In each domain, the external scrutineer has the authority to reject, exclude, defund, or penalize the work. The control surface is the layer that produces the evidence to withstand that scrutiny.
The Gate as a Control Surface
A document that does not pass The Gate does not reach any external stakeholder.
The Gate is a pass/fail threshold in the verification process. It operates as a control surface with the four components defined above:
Scope: Every citation, every central claim, every methodology assertion in the document.
Standard: Three-layer verification. Existence (does the source exist?), polarity (is it supported or contested?), alignment (does it support the specific claim?). Plus internal consistency and scope boundary checks.
Decision: The document either passes, passes with corrections (flagged items must be resolved), or fails (structural issues require significant revision before re-verification).
Record: A structured verification report that maps each checked element to its status, each correction to its reviewer, and the final signoff to a named individual and timestamp.
The researcher whose paper passes The Gate submits with a verification record. The litigation team whose expert report passes The Gate files with an exhibit. The PI whose grant proposal passes The Gate has compliance documentation. The compliance officer whose regulatory filing passes The Gate hands the auditor a record instead of an email chain.
The record is the deliverable. Not the approval. The record.
What Happens Without It
Without a control surface, the signoff is invisible. The work is defensible only as long as nobody asks how it was verified.
The consequences of missing control surfaces are predictable because they follow the same pattern across domains:
A researcher submits a paper. A reviewer checks a key citation. The citation does not support the claim. The reviewer asks whether the author verified the reference list. There is no record. The paper is flagged for major revision.
An expert witness submits a report. Opposing counsel identifies a methodology claim that is not supported by the cited literature. The litigation team is asked for the verification record. There is no record. The Daubert motion is granted.
A PI submits a grant proposal. A reviewer notes that the power analysis does not match the study design. The program officer asks for documentation of the internal review process. There is no documentation. The score drops.
A compliance officer submits a regulatory filing. The auditor finds claims not supported by source documents. The auditor asks for the review record. There is an email that says “approved.” The auditor asks what was checked. The conversation shifts from audit to remediation.
In every case, the work may have been done well. The AI output may have been accurate. The human may have read it carefully and caught the important issues. None of that matters if the record does not exist.
The human signoff is not a rubber stamp. It is the control surface that makes AI output defensible.
Building the Surface
The question is not whether your organization needs a control surface for AI-assisted work. It is whether you build it before or after the scrutiny arrives.
Three starting points:
For researchers: Run the Trust Stack on your citation set before submission. The verification report becomes your record. Attach it to your internal review file.
For litigation teams: Run the Contradiction Register on your expert report before filing. The vulnerability scan and signoff become an exhibit for the Daubert defense.
For compliance officers: Establish a verification step between AI-assisted drafting and regulatory submission. Define the scope, the standard, the decision authority, and the record format. The process can be manual or systematic. The record must exist.
The organizations that build control surfaces before the scrutiny are the ones that pass the scrutiny. The organizations that build them after are the ones writing corrective action plans and explaining gaps.
The control surface is the part of the process that makes everything else defensible. Build it.
If your organization uses AI in expert work and the signoff process is “someone approved it,” the control surface does not exist yet. Start with an architectural audit or reach us at [email protected].