Human Review Workflows

Human review handles ambiguous cases where Deep Mod's confidence falls below your policy threshold. This section explains how to enable and use human review in your policies.

When Human Review Triggers

Human review automatically activates when three conditions are met:

1. Policy Review Mode: Your policy's review mode is set to "Human Review"

2. Ambiguous Result: AI confidence falls below your confidence threshold on one or more rules

3. Live Moderation: Only production runs trigger review (test runs complete automatically)

When these conditions are met, the moderation run status becomes pendingReview instead of completing automatically. Webhooks are delayed until review is finished.

Review Interface

Granular Rule Review

Deep Mod uses a granular review system where reviewers make individual decisions for each ambiguous rule—not a single approve/reject for the entire content.

Layout: The review interface displays:

• Left panel: The original content exactly as submitted

• Right panel: Rule cards for each ambiguous rule requiring a decision

What Reviewers See for Each Rule

Each ambiguous rule is displayed as a card showing:

• Rule condition: The rule text that triggered the ambiguous result

• Confidence score: The AI's confidence percentage for this rule

• Matched content: The specific text segment that contributed to the trigger

• Decision buttons: Approve (✓) or Reject (✗) for this specific rule

Making Review Decisions

1. Review each ambiguous rule card individually

2. Click Approve if the content passes the rule, or Reject if it violates the rule

3. Progress is tracked (e.g., "3/5 rules reviewed")

4. Once all rules have decisions, the derived result is shown

Derived Result Logic: The final moderation result is determined by your rule decisions:

• If any rule is marked as Reject → Final result is FAILURE

• If all rules are marked as Approve → Final result is SUCCESS

Review Notes

Add an optional note (up to 1000 characters) to explain your reasoning. This note is included in the webhook payload and stored in the audit trail.

Review Process

1. Automatic Queue: Ambiguous results enter the review queue with status pendingReview

2. Access Reviews: Team members access pending reviews through the dashboard

3. Rule-by-Rule Decisions: Reviewer approves or rejects each ambiguous rule

4. Add Note: Reviewer optionally adds an explanatory note

5. Submit: Click "Submit Review" to finalize

6. Complete: Final decision replaces ambiguous result, status changes to completed, and webhook is dispatched

After Review Completion

Webhook Delivery

Completed reviews trigger webhooks with the final result. The payload includes:

Example webhook payload:

{
  "policy": "content-policy",
  "result": "failure",
  "reviewed": true,
  "reviewNote": "User uploaded copyrighted content",
  "reviewItems": [
    { "ruleId": 123, "ruleName": "Copyright Violation", "decision": "failure" },
    { "ruleId": 124, "ruleName": "Spam Detection", "decision": "success" }
  ],
  "ruleGroupResults": [ ... ]
}

Audit Trail

All review decisions are permanently recorded with:

• Reviewer identity

• Timestamp

• Individual rule decisions

• Review note (if provided)

Usage Tracking

Reviews are tracked in analytics but don't count toward usage limits since they update existing moderation runs.

Review Mode Settings

Human Review Mode

Ambiguous cases wait for human decision before completing.

• Content evaluation pauses until reviewed

• Webhooks are delayed until review is finished

• Requires dedicated reviewers with dashboard access

• Best for: High-stakes content where accuracy is critical

No Review Mode

All decisions are automated with no human intervention.

• Ambiguous results are sent immediately via webhook with result: "ambiguous"

• Your application handles ambiguous cases programmatically

• Faster processing with full automation

• Best for: High-volume scenarios where you handle edge cases in your own system

Choose review mode based on your accuracy requirements and available review team capacity.

What's Next

• Policy Authoring Best Practices - Design policies that balance automation with review needs

• Plan Usage & Limits - Monitor review usage within your plan limits

Questions about human review? Contact support for guidance on setting up review workflows for your team.