WorkflowType Safe WorkflowCommunityNeeds review

AI Content Quality Review

Evaluate AI-assisted content against a brief, evidence, and publication rules with an explicit review gate.

An editorial workflow for scoring completeness, evidence, brand fit, and policy risk without treating a model judgement as final approval.

Provider: Type Safe Workflow
Reviewed: September 23, 2026
Project status: Community
Verification: Needs review
Categories: content, operations
Compatibility: JEV, Laya, Bespoke Nimble, Markdown

Inputs

✓Content brief
✓Draft
✓Evidence links
✓Editorial policy

Outputs

✓Quality scores
✓Blocking issues
✓Review route
✓Revision notes

Workflow steps

1Freeze the brief, audience, required claims, prohibited claims, and disclosure rules.
2Score completeness, evidence, clarity, brand fit, and policy risk independently.
3Use deterministic checks for missing links, required headings, and disclosure text.
4Block unsupported factual claims and policy-sensitive findings regardless of average score.
5Send low-confidence or conflicting dimensions to a qualified editor.
6Record the final editor decision separately from model output.

Decision tasks

✓Score
✓Verify

Workflow implementation guide

This editorial pattern requires local evaluation before operational use. Empty tested-version fields mean the directory has not run it against a pinned model release.

Business problem

AI-assisted drafts can look fluent while missing the brief, evidence, disclosure, or policy requirements needed for publication.

Who this is for

Editorial, growth, and compliance teams reviewing AI-assisted marketing or knowledge content.

Model selection notes

Select by agreement with trained editors on each rubric dimension. Do not combine a strong style score with factual or policy approval.

Tested model versions

Unknown

Industry: Content, Marketing
Difficulty: intermediate
Estimated setup time: 3 to 6 hours for rubric and sample review
Compatible model candidates: jev, laya, bespoke-nimble
Required integrations: Content repository, Evidence checker
Recommended model: No single model recommended before local evaluation

Detailed execution steps

1Freeze the brief, audience, required claims, prohibited claims, and disclosure rules.
2Score completeness, evidence, clarity, brand fit, and policy risk independently.
3Use deterministic checks for missing links, required headings, and disclosure text.
4Block unsupported factual claims and policy-sensitive findings regardless of average score.
5Send low-confidence or conflicting dimensions to a qualified editor.
6Record the final editor decision separately from model output.

Confidence policy

Calibrate each rubric dimension separately. Factual support and policy risk remain blocking checks instead of being averaged away.

Human review policy

A qualified editor makes every publication decision and verifies factual claims, disclosures, and sensitive-topic guidance.

Privacy notes

Drafts can contain embargoed or customer information. Use the minimum required text and follow the content-access policy.

Cost notes

Estimate cost per reviewed draft, including claim extraction and editor time. Avoid repeated full-document calls when only one section changed.

Common failure modes

✓Fluent language receives a high score despite unsupported claims.
✓One overall score hides a blocking policy issue.
✓The reviewer sees model output before forming an independent judgement.

Troubleshooting

✓Separate evidence, policy, and style decisions into independent fields.
✓Blind a sample of reviewers to model output and compare disagreement rates.

Structured request example

{"state":{"brief":"Compare three sourced decision models","draft":"..."},"questions":{"evidence":{"type":"score","criteria":["unsupported","partial","fully sourced"]},"policy_risk":{"type":"noul"}}}

Keep blocking checks independent from averages

typescript · TypeScript 5.9

Provider version must be pinned before testing

export function publicationRoute(result: {
  averageScore: number;
  unsupportedClaim: boolean;
  policyRisk: number;
  confidence: number;
}) {
  if (result.unsupportedClaim || result.policyRisk >= 0.4) return 'block';
  if (result.confidence < 0.85 || result.averageScore < 0.8) return 'review';
  return 'editor-approval';
}

Implementation guidance

Keep the declared inputs and outputs visible in code review. Validate every model response before it reaches a database, API, or customer-facing action.