Reviewability is a property of the record around a finding. It requires provenance, visible transformations, stated uncertainty, an accountable reviewer, and a link between the evidence and the action that followed.
The finding is not the whole record
A summary, classification, or score may be useful, but it is not reviewable evidence by itself. A reviewer needs enough context to understand the original question, the permitted source material, the transformation applied, and what the output was intended to support.
A minimum review record
The record should be proportionate to the consequence of the decision. Higher-consequence uses require stronger controls, clearer ownership, and a more complete account of uncertainty and exceptions.
- The operating question and intended use
- Source identity, time, scope, permissions, and provenance
- Material transformations, model or rule version, and scoring standard
- Missing, conflicting, excluded, or uncertain information
- Named review responsibility and permitted intervention points
- The decision or action informed, including the human rationale
- Follow-up evidence and any change to the original conclusion
Test the chain, not just the output
Evaluation should ask whether a reviewer can trace the output to its source, reproduce the stated rule where appropriate, identify material omissions, and understand why the evidence was considered sufficient for the decision. Agreement with an output does not demonstrate that this chain is sound.
Monitoring should also examine whether the source conditions, system behavior, or intended use changed after the original review.
Limits of a reviewable record
A complete record can still contain poor data, flawed assumptions, an unsuitable metric, or human error. Documentation cannot by itself establish fairness, legal compliance, safety, or fitness for a particular use. Those questions require domain-specific review and evidence.
This note is a working definition for product and research review. It is not a certification, assurance standard, legal conclusion, customer result, or claim of operational deployment.
Sources and reference frameworks
- Artificial Intelligence Risk Management Framework (AI RMF 1.0)
National Institute of Standards and Technology. A voluntary framework for incorporating trustworthiness considerations into the design, development, use, and evaluation of AI systems.
- Artificial Intelligence: An Accountability Framework for Federal Agencies and Other Entities
U.S. Government Accountability Office. An accountability framework organized around governance, data, performance, and monitoring, with review questions for system owners and assessors.