Evidence Sufficiency Report Generation Prompt Template

Use this prompt when you have a batch of individual claims that have already been scored for evidence sufficiency and you need to produce a consolidated, stakeholder-ready report. The primary job-to-be-done is transforming structured verification data—confidence scores, source authority weights, corroboration counts, and contradiction flags—into a narrative summary that decision-makers, compliance reviewers, or downstream governance systems can act on. The ideal user is a verification pipeline operator or engineering lead who has completed the scoring phase and now needs a traceable, audit-ready output that explains what was checked, against what evidence, and with what level of certainty.

Do not use this prompt for real-time claim scoring, evidence retrieval, or raw fact-checking. It assumes that the heavy lifting of extraction, matching, and confidence calibration has already occurred upstream. If you feed it unscored claims or raw documents, it will hallucinate evidence assessments rather than report on them. The prompt requires structured input containing at minimum: a claim identifier, the claim text, a sufficiency score, an evidence inventory with source references, and any contradiction or uncertainty flags. Without these fields, the report will lack the traceability that makes it useful for audit and review workflows.

This prompt is appropriate for high-stakes domains—legal, financial, medical, or scientific verification—only when paired with a human review step. The generated report can accelerate reviewer understanding by pre-structuring the evidence landscape, but it must not be treated as a final determination. Always validate that the report faithfully represents the input scores without introducing unsupported conclusions or omitting low-confidence claims. If your pipeline routes claims to human reviewers based on confidence thresholds, use this prompt to generate the review packet, not to make the routing decision itself.

This prompt template is designed to be the final stage in a claim verification pipeline. It consumes a batch of claims that have already been scored for evidence sufficiency, source authority, and corroboration, and synthesizes them into a stakeholder-ready report. The template assumes you have upstream processes for claim extraction and scoring; its job is to produce a consistent, auditable narrative and structured output that can be delivered to compliance teams, editorial boards, or product managers. Do not use this prompt on raw, unscored claims, as it will hallucinate evidence assessments to fill the gap.

textCopy
You are an evidence verification reporter. Your task is to produce a structured Evidence Sufficiency Report from a set of pre-scored claims. Do not re-score the claims. Synthesize the provided scores and evidence summaries into a clear, actionable report.

## INPUT
[SCORED_CLAIMS_JSON]

## OUTPUT SCHEMA
You must respond with a single JSON object conforming to this structure:
{
  "report_id": "string",
  "generated_at": "ISO 8601 timestamp",
  "executive_summary": {
    "total_claims": "integer",
    "sufficient_evidence_count": "integer",
    "insufficient_evidence_count": "integer",
    "contradicted_count": "integer",
    "overall_assessment": "string (2-4 sentence summary of the evidence landscape)"
  },
  "evidence_inventory": [
    {
      "source_id": "string",
      "source_description": "string",
      "authority_tier": "high | medium | low",
      "claims_supported": ["claim_id", ...]
    }
  ],
  "confidence_breakdown": [
    {
      "confidence_level": "high | medium | low | unverified",
      "claim_count": "integer",
      "claim_ids": ["claim_id", ...]
    }
  ],
  "action_items": [
    {
      "priority": "critical | high | medium | low",
      "action": "string (specific, assignable task)",
      "related_claim_ids": ["claim_id", ...]
    }
  ],
  "human_review_required": "boolean",
  "review_questions": ["string (specific question for a human reviewer)", ...]
}

## CONSTRAINTS
- Do not alter any claim_id, source_id, or score value provided in the input.
- If a claim has a confidence score below [CONFIDENCE_THRESHOLD], it must be listed in the action_items with a priority of at least 'high'.
- The executive_summary.overall_assessment must not introduce new facts or evidence not present in the input.
- If any claim is flagged as contradicted, human_review_required must be set to true.
- The evidence_inventory must deduplicate sources by source_id.
- All action_items must be specific and assignable (e.g., 'Retrieve Q3 earnings transcript to verify claim CL-0042', not 'Find more evidence').

## RISK LEVEL
[RISK_LEVEL]

## EXAMPLES
[FEW_SHOT_EXAMPLES]

To adapt this template, replace the square-bracket placeholders with your pipeline's data. [SCORED_CLAIMS_JSON] should be a JSON array of claim objects, each containing at minimum a claim_id, claim_text, confidence_score, evidence_summary, and contradiction_flag. [CONFIDENCE_THRESHOLD] is a float between 0.0 and 1.0 that defines your triage boundary for human review. [RISK_LEVEL] should be set to high, medium, or low to adjust the tone and urgency of the report's language. [FEW_SHOT_EXAMPLES] is optional but strongly recommended; provide 1-2 example input-output pairs to stabilize the report structure, especially for the action_items and overall_assessment fields. Before deploying, validate the output JSON against the schema and run a completeness eval to ensure no input claim is dropped from the report.

The Evidence Sufficiency Report prompt is designed as the final stage in a claim verification pipeline. It consumes structured scored claims—each with a confidence level, evidence inventory, source authority assessments, and contradiction flags—and produces a stakeholder-ready report. This prompt should not be called in isolation. It expects upstream outputs from claim extraction, evidence matching, and confidence scoring prompts. The implementation harness must validate that all required inputs are present before invoking the report generation step, and must enforce output schema compliance before the report reaches a human reviewer or downstream system.

Wire this prompt into your application as a post-scoring assembly step. The harness should: (1) collect scored claims from the confidence scoring pipeline, ensuring each claim object includes claim_id, claim_text, confidence_score, evidence_items (with source, authority_tier, recency_date), contradiction_flags, and sufficiency_tier; (2) inject these into the [SCORED_CLAIMS] placeholder as a structured JSON array; (3) set [REPORT_SECTIONS] to specify which sections are required (e.g., executive_summary, evidence_inventory, confidence_breakdown, action_items); (4) define [OUTPUT_SCHEMA] as a strict JSON schema with fields for report_id, generated_at, sections[], and completeness_checks; (5) set [RISK_LEVEL] to high for regulated domains, which triggers mandatory human-review routing. After model invocation, validate the output against the schema. If validation fails, retry once with the error message injected into [PREVIOUS_ERRORS]. If the retry also fails, escalate to a human operator with the partial output and error context.

For production deployments, log every report generation attempt with the prompt version, model used, input claim count, output schema compliance status, and any retry or escalation events. Implement a review queue for reports where completeness_checks.any_gaps == true or where the RISK_LEVEL is high. The human reviewer should confirm that the executive summary accurately reflects the evidence inventory and that action items are correctly prioritized. Do not auto-publish reports without this review gate when the risk level is elevated. For model choice, use a model with strong instruction-following and structured output capabilities (e.g., GPT-4o, Claude 3.5 Sonnet). Avoid smaller or older models that may drop required sections or hallucinate evidence details not present in the input. If your pipeline uses RAG for evidence retrieval, ensure the evidence items passed to this prompt are already grounded and cited—this prompt synthesizes, it does not retrieve.