Override Decision Comparison Prompt

Governance teams and audit leads use this prompt to perform structured consistency checks on override decisions made by different reviewers or teams. The core job-to-be-done is identifying divergent reasoning in cases with similar evidence, surfacing potential fairness concerns, and preparing evidence for regulatory examinations. This prompt assumes you have already captured structured override records using a template like the Override Justification Structured Log Prompt Template. It is not designed for capturing a single override event; it is a comparative analysis tool for batch review.

The ideal workflow ingests two or more completed override records, each containing the original AI decision context, the human override justification, and the reviewer's identity. The prompt requires these records to be structured and machine-readable. It produces a side-by-side comparison that highlights specific points of divergence in reasoning, flags cases where different conclusions were reached from similar evidence, and identifies potential policy gaps. This is a high-stakes governance workflow, so the implementation harness must include validation checks for input completeness, a human-in-the-loop review step for any flagged divergence, and a clear audit log of the comparison itself.

Do not use this prompt for real-time decision support or as a gate before a single override is executed. It is a retrospective analysis tool for periodic consistency audits, fairness reviews, and pre-regulatory examination preparation. Avoid using it when the input records lack structured justification fields or when the original AI decision context is missing, as the comparison will be unreliable. If you need to check whether a single override complies with policy before execution, use the Override Decision Policy Compliance Check Prompt instead.

This prompt template is designed to be pasted directly into your AI orchestration layer. It instructs the model to perform a structured comparison of two override decisions, focusing on the consistency of their justifications. The goal is to surface divergent reasoning that could indicate fairness issues, policy misinterpretation, or inconsistent governance. Before using this prompt, ensure you have retrieved the complete audit records for both decisions, including the original AI recommendation, the human override justification, and any relevant policy documents.

codeCopy
You are an AI governance auditor. Your task is to compare two override decisions and their justifications to identify inconsistencies and potential fairness concerns.

### INPUT DATA

**Decision A:**
- Original AI Recommendation: [DECISION_A_AI_RECOMMENDATION]
- Human Override Justification: [DECISION_A_OVERRIDE_JUSTIFICATION]
- Reviewer Role: [DECISION_A_REVIEWER_ROLE]
- Case Context: [DECISION_A_CASE_CONTEXT]

**Decision B:**
- Original AI Recommendation: [DECISION_B_AI_RECOMMENDATION]
- Human Override Justification: [DECISION_B_OVERRIDE_JUSTIFICATION]
- Reviewer Role: [DECISION_B_REVIEWER_ROLE]
- Case Context: [DECISION_B_CASE_CONTEXT]

**Relevant Policy:**
[RELEVANT_POLICY_DOCUMENT]

### OUTPUT SCHEMA

You must respond with a single JSON object conforming to this schema:
{
  "comparison_id": "string",
  "decisions": [
    {
      "decision_id": "string",
      "summary": "string (concise summary of the override action)",
      "primary_justification": "string (the core reason for the override)"
    }
  ],
  "similarity_assessment": {
    "are_justifications_consistent": "boolean",
    "consistency_score": "number (0.0 to 1.0)",
    "rationale": "string (explanation of the score)"
  },
  "divergence_analysis": [
    {
      "point_of_divergence": "string (specific area where justifications differ)",
      "decision_a_perspective": "string",
      "decision_b_perspective": "string",
      "potential_fairness_concern": "boolean",
      "concern_description": "string (if concern is true, explain the potential bias or inconsistency)"
    }
  ],
  "recommended_action": "string (one of: 'no_action', 'human_review', 'policy_clarification', 'training_recommended')",
  "action_rationale": "string"
}

### CONSTRAINTS
- Base your analysis strictly on the provided justifications and policy. Do not infer intent beyond what is stated.
- If the case contexts are materially different, acknowledge this in your similarity assessment and do not flag differences as fairness concerns if they are explained by context.
- Flag a potential fairness concern only when similar cases appear to be treated differently without a clear, documented reason.
- If the provided justifications are insufficient for a meaningful comparison, set the consistency_score to 0 and explain why in the rationale.

To adapt this template, replace the square-bracket placeholders with real data from your audit system. The [DECISION_A_CASE_CONTEXT] and [DECISION_B_CASE_CONTEXT] fields are critical; they should contain enough detail for the model to determine if the two cases are truly comparable. The [RELEVANT_POLICY_DOCUMENT] should be the specific policy clause that governs the type of decision being overridden. For high-stakes governance workflows, the output of this prompt should always be reviewed by a human before any action is taken based on its findings.

The Override Decision Comparison Prompt is designed for batch or pairwise analysis, not real-time decision blocking. Wire it into a scheduled governance job or an on-demand audit tool that pulls override records from your audit log database or compliance warehouse. The prompt expects two or more override records with their full justification text, reviewer metadata, decision context, and the original AI recommendation. Before calling the model, normalize the input records into a consistent JSON structure so the prompt receives clean, comparable fields. Do not pass raw log lines or unstructured text directly—preprocess override reasons, timestamps, reviewer roles, and decision types into the [OVERRIDE_RECORDS] array.

Implement a validation layer after the model response that checks the output against the [OUTPUT_SCHEMA]. The schema should require a comparison table with aligned decision elements, a divergence summary array, and a fairness concern flag. Validate that every input record appears in the comparison output, that divergence reasons are grounded in specific text from the justifications, and that the fairness flag includes supporting evidence rather than a bare boolean. If validation fails, retry once with the validation errors injected into the [CONSTRAINTS] field. If the second attempt also fails, route the comparison to a human governance reviewer with the partial output and validation failure log attached. Log every comparison result—including validation passes, retries, and escalations—to your audit trail system with the prompt version, model, and timestamp.

Choose a model with strong reasoning and structured output capabilities for this task. The prompt requires careful alignment detection and divergence explanation, which smaller or faster models may handle inconsistently. If you are comparing more than five override records at once, split the batch into smaller groups and run parallel comparisons, then use a secondary synthesis prompt to merge the results. For high-stakes governance workflows, always require human sign-off on flagged fairness concerns before the comparison is considered final. Never use this prompt to automatically overturn override decisions or to generate compliance findings without human review. The output is an analytical artifact, not a binding determination.