FUURAA AI Knowledge Library · Human review protocol

From evidence record to accountable conclusion

A bilingual human-review protocol for AI claims: separate evidence assembly, domain judgment, challenge and final decision to produce a bounded, dated and reopenable conclusion under uncertainty.

Enter the seven review gates ↓

Published4 August 2026Evidence statusFUURAA method synthesis grounded in primary standards and researchScopeHuman evidence review of public AI claims

Boundary first

Structural verification is admission to review—not a conclusion.

Automation can expose missing fields and obvious chronology errors. It cannot inspect every primary source, decide whether an evaluation fits the current decision, seek credible contradiction or own the consequence. Human review makes those judgments explicit and leaves a record.

Applicability boundaryThis protocol is a public research and decision-record method. It does not verify signer identity and does not replace replication, professional safety assessment, legal advice, audit, sector compliance or certification. It discloses no unannounced FUURAA product capability.

Seven review gates

Every gate must leave an input, a judgment and a stop condition.

Review does not begin with belief or disbelief. It begins with a defined decision mandate, and conclusion strength may rise only with evidence fit.

Freeze the review mandate

Core question: Which exact claim and decision is this review allowed to inform?
Evidence to preserve: Record ID, bounded claim, decision owner, deadline and excluded interpretations.
Stop condition: Stop when the claim, subject, comparison or decision can still change without reopening the review.

Declare roles and conflicts

Core question: Who assembled the record, who challenges it and who owns the final decision?
Evidence to preserve: Evidence lead, domain reviewer, challenge reviewer, decision owner and material conflicts.
Stop condition: Do not label a developer self-review as independent review, or let one person silently fill every role.

Admit only reviewable records

Core question: Can every material source, version, date, method, artefact and boundary be inspected?
Evidence to preserve: A structurally verified record plus accessible primary sources and named missing artefacts.
Stop condition: Block the review when identity, chronology, evaluation design, excluded uses or accountable owner is absent.

Test claim–evidence fit

Core question: Does the evidence measure this system, task, population, language, place, workflow and time?
Evidence to preserve: A mapping from each claim clause to direct support, indirect support, contradiction or unknown.
Stop condition: Do not upgrade laboratory, single-language or developer-only results into broad deployment claims.

Seek contradiction and uncertainty

Core question: What credible evidence could reverse, narrow or postpone the conclusion?
Evidence to preserve: Alternative explanations, failed cases, subgroup differences, variation, missing coverage and dissent notes.
Stop condition: Do not close review while a material contradiction is omitted, unexplained or assigned no owner.

Assign a decision-sized evidence state

Core question: What is the strongest bounded conclusion the record can carry today?
Evidence to preserve: Supported, developing, mixed or insufficient; rationale; unknowns; dissent; and rejected stronger wording.
Stop condition: Do not average incompatible dimensions into one confidence score or let urgency raise evidence strength.

Record, expire and reopen

Core question: Who accepts the conclusion, when does it expire and what new event reopens it?
Evidence to preserve: Dated decision, sign-off roles, residual uncertainty, next-review trigger and superseded-record link.
Stop condition: An expired, materially changed or superseded record must not continue to justify a consequential decision.

Separation of duties

Different roles ask different questions so one blind spot does not travel end to end.

01
Evidence lead
Assembles sources and preserves provenance; does not decide alone.
02
Domain reviewer
Checks whether method, metrics and operating context fit the field.
03
Challenge reviewer
Searches for contradiction, leakage, transfer failure and omitted affected parties.
04
Decision owner
Accepts the bounded conclusion, residual uncertainty and reopening obligation.

Decision language

Evidence states describe the current record without false precision.

supported

Supported

Relevant evidence converges, material contradictions are addressed and the wording stays inside the tested boundary.

developing

Developing

Credible support exists, but replication, duration, coverage or field observation remains incomplete.

mixed

Mixed

Credible records disagree or results change materially across methods, groups or operating conditions.

insufficient

Insufficient

The record is too sparse, indirect, outdated or mismatched to support this decision-sized claim.

FUURAA analysisOne system can legitimately hold different evidence states. ‘Outperforms a baseline on one fixed English test’ may be supported while ‘fit for multilingual Southeast Asian public services’ remains insufficient. Review quality depends on matching the conclusion to claim granularity—not producing one confident-looking score.

Printable decision record

Twelve fields make the conclusion, dissent and expiry handoff-ready.

01review_id
Stable ID and the evidence-record version reviewed
02claim + decision
Exact wording and the decision it may inform
03roles
Evidence, domain, challenge and decision owners
04conflicts
Developer, funder, commercial or institutional relationships
05admitted evidence
Primary records and artefacts admitted to review
06excluded evidence
Records excluded and the reason for exclusion
07claim fit
Direct, indirect, contradictory and unknown clauses
08boundary
Version, task, population, language, place and workflow
09state
Supported, developing, mixed or insufficient
10rationale + dissent
Why, what remains unknown and any unresolved dissent
11decision
Accept, narrow, defer, reject or commission more evidence
12expiry + trigger
Expiry date and events that reopen the review

Primary sources and boundaries

Draw from public frameworks without presenting FUURAA method as standards conformance.

Sources checked 4 August 2026. Each source states its role and what cannot be inferred from it.

Primary sourceNIST AI RMF 1.026 January 2023

Connects governance, context mapping, measurement and management across the AI lifecycle.

BoundaryVoluntary and cross-sector; it does not prescribe this review protocol or confer certification.

Open primary source ↗Primary sourceNIST AI RMF PlaybookCompanion to AI RMF 1.0

Provides suggested actions for Govern, Map, Measure and Manage outcomes.

BoundaryThe Playbook states that its suggestions are not a checklist and require contextual selection.

Open primary source ↗Primary sourceNIST AI 600-1 · Generative AI Profile26 July 2024 · page updated 8 April 2026

Adds generative-AI risks and actions around measurement, provenance, information integrity and human oversight.

BoundaryA cross-sector companion resource; actual priorities depend on the named system and use context.

Open primary source ↗Primary sourceW3C PROV-DMW3C Recommendation · 30 April 2013

Models entities, activities, agents, derivation and responsibility for provenance interchange.

BoundaryConsistent provenance does not by itself establish evidence quality, truth or decision fitness.

Open primary source ↗Primary sourceModel Cards for Model Reporting2019

Motivates disclosure of intended use, evaluation conditions, performance and relevant subgroup results.

BoundaryA reporting method cannot independently verify the measurements or make an unsuitable system safe.

Open primary source ↗

Complete evidence path

Build the record, verify its structure, then assign accountable people to reach a conclusion.

When evidence or operating context changes, do not overwrite the old conclusion. Preserve it, create a new version and state which facts, boundaries or judgments changed.

Build an AI claim evidence record Run structural verification Enter AI Evidence Atlas Return to AI Knowledge Library