Reviewer forms are structured records used by human reviewers to document how a specific GenAI run was assessed. In RAIDT, they capture scores, evidence pointers, decisions, disagreements, and rationale in a way that is tied to the run record rather than treated as detached commentary. Their function is to make human judgement inspectable, comparable across runs, and usable within a governance process.

Conceptually, reviewer forms sit between informal review notes and formal audit records. They are more structured than ad hoc comments, because they require defined fields, criteria and decision statements; but they are more operational than a high-level policy document, because they are completed in relation to an actual run, model output, task context and evidence set. This is why they fit naturally inside RAIDT, which treats the run as the unit of governance.

In GenAI governance, organisations often claim that there is human oversight, but the evidence for that oversight is weak. A reviewer may approve or reject an output without leaving a reconstructable explanation of what was reviewed, against which criteria, with what level of confidence, and on the basis of which evidence. Reviewer forms address that gap. They convert human oversight into a recorded governance artefact that can travel with the evidence pack and inform the five-pillar score profile.

Reviewer forms also differ from generic quality-assurance checklists. A checklist may confirm that a step was completed, whereas a RAIDT reviewer form records the reasoning, evidence trail and judgement applied to a specific run. That makes it relevant not only to compliance but also to contestability, post-run review, corrective action, and organisational learning.

Reviewer forms solve a recurring operational problem in responsible AI governance: human oversight is frequently asserted but poorly evidenced. Without a structured review instrument, organisations struggle to show who reviewed a run, what they saw, how they judged quality or risk, why they accepted or rejected the output, and what happened when reviewers disagreed. This weakens accountability and makes retrospective review difficult.

The concept also avoids a common confusion between human presence and meaningful human oversight. A person looking at an output is not enough. Governance requires a record of the judgement process, including the criteria used and the basis for the decision. Reviewer forms help organisations move from a principle such as ?a human remains in the loop? to an operational reality in which that human intervention is documented and reviewable.

If reviewer forms are missing, several risks appear. Decisions become hard to explain, disputed cases become hard to resolve, scoring becomes inconsistent across reviewers or teams, and the organisation loses the ability to learn systematically from borderline or problematic runs. In practice, this means weaker audit readiness, weaker defensibility in front of supervisors or regulators, and weaker evidence for continuous improvement.

Key idea: Reviewer forms matter because they turn human oversight from an assertion into run-level evidence that can be reviewed, challenged and used for governance.

• The identity, role, and authority of the reviewer for a specific run.
• The run identifier, task context, and output under review.
• The rubric criteria or assessment dimensions used in the review.
• Scores, ratings, or qualitative judgements linked to those criteria.
• Evidence pointers showing what artefacts or logs informed the judgement.
• Acceptance, rejection, revision, or escalation decisions.
• Disagreement notes where multiple reviewers interpret the run differently.
• The rationale explaining why the reviewer reached that conclusion.
• Follow-up actions such as corrective action, monitoring flags, or post-run review triggers.

Audience question

Why use reviewer forms when an expert can simply approve the output directly?

Answer

The concern behind this question is usually that reviewer forms look bureaucratic or duplicative. If a qualified human has already checked the output, it can seem unnecessary to ask that person to complete additional documentation. The direct answer is that approval alone is not enough for governance. A later reviewer, supervisor, auditor, or policy stakeholder needs to know what was reviewed, what standard was applied, what evidence was considered, and why the final decision was justified.

A practical example is a clinician reviewing a GenAI-assisted discharge summary. If the clinician simply clicks ?approved?, the organisation knows only that a review happened. If the clinician completes a reviewer form, the organisation can see whether the review addressed factual accuracy, omission risk, patient-safety concerns, source-document consistency, and any edits made before release. That produces a record that can be revisited if a problem emerges later.

RAIDT handles this better than a generic AI governance approach because it does not treat the review as a detached compliance ritual. The reviewer form is tied to a specific run, integrated into the evidence pack, and capable of influencing the score profile across pillars such as Responsibility, Auditability and Traceability. In other words, RAIDT makes reviewer judgement operational, not merely symbolic.

Consider a public-service team using a GenAI system to draft benefit-eligibility explanation letters for citizens. One run produces a letter that is fluent and apparently helpful, but the output omits an important explanation of appeal rights. The run-level issue is not just whether the model was generally useful; it is whether this particular run produced an output that could mislead a recipient in a legally and procedurally significant way.

The evidence needed includes the run ID, the prompt and input context, the generated letter, the relevant policy or legal guidance, the reviewer form, and the reviewer?s rationale for revision or rejection. The reviewer form records that the output was understandable but incomplete, notes the missing appeal-rights explanation, links that judgement to the relevant guidance, and marks the run for corrective action and template adjustment.

The RAIDT pillars affected are Responsibility, because a documented human decision is required; Auditability, because the review can be reconstructed; Interpretability, because the rationale explains what made the output unacceptable; Dependability, because repeated omissions can be tracked across runs; and Traceability, because the judgement is linked back to the run and the supporting evidence. In governance-readiness terms, the reviewer form turns a potentially disputable review event into a defensible record.

Reviewer forms link to RAIDT in four ways.

First, they operationalise RAIDT?s central claim that governance should be based on evidence about a specific run rather than on broad assurance statements about a system in general.
Second, they connect human judgement directly to the run-level evidence, so that oversight is attached to the exact output, context and decision being governed.
Third, they feed the evidence pack and help justify elements of the RAIDT score profile by recording how reviewers interpreted quality, risk, adequacy and acceptability.
Fourth, they strengthen reviewability, contestability, audit readiness and organisational learning by leaving a structured trail of judgement that others can revisit later.

Reviewer forms ? Run-level review ? Evidence pack ? RAIDT score profile ? Governance readiness

Reviewer forms have their strongest effect on Responsibility, Auditability and Traceability, but they also support Interpretability and Dependability when used consistently.

Responsibility

Reviewer forms clarify who exercised judgement, under what authority, and with what decision outcome. They help show that responsibility is not abstractly assigned but operationally enacted for a particular run.

Example evidence / implication:

• Named reviewer role and approval authority recorded against the run.
• Acceptance, rejection, escalation, or revision decision explicitly documented.

Auditability

Reviewer forms are a core audit artefact because they preserve the reasoning process behind a human review. They allow later inspection of what criteria were applied and what evidence supported the decision.

Example evidence / implication:

• Review criteria and scores recorded in a consistent structure.
• Evidence pointers and rationale available for retrospective reconstruction.

Interpretability

Reviewer forms can make the basis of human judgement more interpretable by explaining why an output was considered acceptable, risky, misleading, incomplete, or in need of revision.

Example evidence / implication:

• Free-text rationale explains which parts of the output drove concern.
• Reviewer notes identify which criteria were most influential in the decision.

Dependability

Reviewer forms support dependability by revealing recurring failure modes across runs and by making review outcomes comparable over time. This helps organisations identify instability, drift, or persistent weaknesses.

Example evidence / implication:

• Repeated reviewer flags show patterns of omission, hallucination, or unsafe phrasing.
• Structured review outcomes inform monitoring and corrective-action priorities.

Traceability

Reviewer forms strengthen traceability because they link judgement to the exact run, evidence sources, reviewer, and downstream action. They make it possible to trace not only what the system produced, but how the organisation responded.

Example evidence / implication:

• Run ID, timestamp, reviewer identity, and linked artefacts stored together.
• Escalation or corrective-action trail connected back to the original review event.

In general AI governance, reviewer forms may mean a simple checklist or sign-off document showing that a human looked at an output. In RAIDT, reviewer forms mean a run-linked governance artefact that records judgement in a structured way, ties that judgement to evidence, and allows the decision to inform evidence packs, score profiles and follow-up actions.

The RAIDT meaning is more operational because it is not satisfied by the statement that ?someone reviewed this?. It asks how the review was documented, what evidence informed it, whether disagreement was captured, whether the decision can be reconstructed, and how the review contributes to governance readiness. That makes reviewer forms part of an evidence architecture, not just an administrative convenience.

Misunderstanding

Reviewer forms are just paperwork added after the real review has already happened.

Correction

In RAIDT, the reviewer form is part of the real review, because it is the mechanism that makes the review inspectable and usable. For example, two reviewers may both reject a GenAI-generated policy summary, but for different reasons: one because it is factually inaccurate, the other because it is procedurally incomplete. Without a reviewer form, those distinctions are lost. With a reviewer form, the organisation can see what kind of problem occurred, how the judgement was reached, and what corrective action is appropriate.

Reviewer forms do not prove that the reviewer was correct, unbiased, careful, or sufficiently expert. They record judgement; they do not guarantee judgement quality. A poorly designed form can also create false confidence if reviewers complete it mechanically or if the rubric is weak, ambiguous, or disconnected from the real task.

Reviewer forms also do not replace the underlying evidence. They should point to outputs, prompts, logs, source documents, or policies, not substitute for them. In addition, they are less effective when reviewers are not trained, when role authority is unclear, or when forms are used inconsistently across teams.

RAIDT handles these limitations by treating reviewer forms as one part of a wider governance package: they work best when linked to run artefacts, scoring logic, post-run review, and corrective-action pathways. In other words, the form is necessary for accountable review, but it is not sufficient on its own.

Manual implementation

A researcher or small team can apply reviewer forms manually by using a standard template in Markdown, PDF, spreadsheet, or form software. The reviewer records the run ID, the output reviewed, the assessment criteria, the decision, and the rationale, then stores the completed form alongside the run evidence.

Semi-automated implementation

Semi-automated implementation can pre-populate reviewer forms with metadata such as run ID, model version, task type, prompt version, timestamps, and linked artefacts. Structured review fields, dropdown criteria, and standardised evidence references improve consistency while still leaving the substantive judgement to the reviewer.

Fully automated implementation

At scale, a governance platform or orchestration layer can generate reviewer forms automatically for selected runs, route them to the correct reviewer role, enforce completion before release, record disagreements, trigger escalation, and attach the resulting judgement to dashboards, monitoring systems, and evidence packs. The automation is in the workflow, routing and record-keeping; the human judgement itself remains accountable human oversight.

Within the RAIDT project, reviewer forms are useful across foundations, empirical validation and policy translation. In Paper 08 Foundations, they help show how RAIDT converts broad governance principles into concrete review artefacts at the level of the run. In Paper 09 Empirical Validation, they can support inter-reviewer comparison, disagreement analysis, and evidence about how governance decisions are actually made in practice. In Paper 10 Policy Pathways, they provide a realistic mechanism by which organisations can demonstrate that human oversight is documented rather than merely declared.

They are also relevant to the run-level evidence pack, because reviewer forms help explain why a run received a particular treatment or score; to the scoring rubric, because reviewer observations can support or qualify pillar assessments; and to governance interventions, because recurring reviewer findings can point to where templates, prompts, workflows, escalation rules, or deployment controls need to change. For supervision and viva preparation, this item is especially valuable because it provides a concrete answer to the question of how RAIDT operationalises human oversight.

Q1. Why are reviewer forms necessary if there is already a human-in-the-loop?

Because ?human-in-the-loop? is a governance claim, not evidence. Reviewer forms make the human judgement visible, structured and reviewable at the level of a specific run.

Q2. Do reviewer forms slow work down too much in practice?

They can add friction if badly designed, but that is precisely why structure matters. A good reviewer form captures only the fields needed for accountable review and can be pre-populated or routed efficiently in semi-automated and automated settings.

Q3. Are reviewer forms mainly for compliance rather than quality improvement?

No. They support compliance, but they also generate learning. Repeated reviewer notes can reveal systematic failure modes, weak prompts, unclear policies, or training needs that would otherwise remain hidden.

Q4. Could reviewer forms become a box-ticking exercise?

Yes, if they are detached from evidence, poorly designed, or never revisited. In RAIDT, their value comes from being linked to run artefacts, scoring logic, disagreement handling, and corrective action.

Q5. What makes a RAIDT reviewer form different from a normal approval form?

A RAIDT reviewer form is tied to a specific run, records evidence pointers and rationale, and contributes to a wider evidence pack and score profile. It is therefore part of governance reconstruction, not just a sign-off record.

• human oversight documentation in AI governance
• structured review forms for high-stakes decision support systems
• audit trails for generative AI review processes
• contestability and documented decision rationale in AI systems
• inter-reviewer consistency in AI-assisted professional judgement
• accountability artefacts for human-in-the-loop AI
• evidence-based governance of generative AI in organisations
• sociotechnical review workflows for AI deployment
• operationalising responsible AI through review documentation
• assurance case design for AI oversight and audit readiness

Reviewer forms are the practical mechanism that turns human oversight into evidence inside RAIDT. Instead of simply saying that a person reviewed a GenAI output, RAIDT asks what they reviewed, what criteria they applied, what evidence they used, what decision they made, and why. The reviewer form records that judgement in a structured way and links it to the specific run. That matters because governance depends on reconstructable decisions, not just on informal assurance statements. It also means reviewer disagreement, escalation, and corrective action can be documented rather than hidden. In the RAIDT model, reviewer forms strengthen responsibility, auditability and traceability most directly, while also improving organisational learning. They are therefore not just paperwork; they are a core governance artefact for making review visible, contestable and operational.

Reviewer forms are structured records of human judgement for a specific run because RAIDT makes oversight credible only when it is captured as run-level evidence.

• S8.01 ? Manual implementation
• S8.02 ? Semi-automated implementation
• S8.03 ? Automated orchestration
• S8.04 ? Gating
• S8.05 ? Monitoring
• S8.06 ? Post-run review
• S8.07 ? Corrective action
• S8.08 ? Cloud deployment
• S8.09 ? Local deployment
• S8.11 ? Reproducibility pack
• S8.12 ? Paper-code separation

• Audience question: Why forms? Answer: they make human judgement inspectable and consistent.