Verification Findings Executive Summary Prompt

This prompt is designed for the final stage of a verification pipeline, after claims have been extracted, matched to evidence, and scored. Its job is to convert a batch of structured verification results into a concise executive summary for board-level stakeholders, compliance officers, and senior decision-makers. The ideal user is an engineering lead or MLOps engineer who has a working verification pipeline and needs to communicate outcomes to a non-technical audience that requires clarity on what was checked, what is reliable, what is uncertain, and what requires immediate attention, without exposing the underlying technical detail of the verification pipeline.

Use this prompt when you have a structured input payload containing claim verdicts, confidence scores, and risk flags from a completed verification run. The prompt expects a machine-readable input, such as a JSON array of claim objects with fields like claim_id, claim_text, verdict (e.g., SUPPORTED, CONTRADICTED, UNSUPPORTED), confidence_score, and risk_flags. It produces a narrative summary with key statistics, risk highlights, and a calibrated confidence overview. This is not a prompt for performing verification; it assumes verification is already complete. Do not use this prompt to analyze raw documents, retrieve evidence, or make verification decisions. It is a reporting tool, not an analysis tool.

Avoid this prompt when the audience needs a detailed, claim-by-claim breakdown with full evidence chains and source citations. For those use cases, use the Structured Claim Verification Report Prompt or the Evidence Chain Assembly Prompt for Audit Trails instead. Also avoid this prompt when the input data is incomplete, unvalidated, or contains raw model outputs that haven't been scored. The summary will inherit any errors in the input data, and presenting unvalidated verification results to executives creates a false sense of rigor. Always run input validation and evidence grounding checks upstream before feeding data into this summary prompt. For high-stakes or regulated domains, ensure a human reviews the summary before distribution.

This template is designed to be pasted directly into your model's system instructions. It takes a batch of structured verification results—claims, verdicts, confidence scores, and evidence summaries—and compresses them into a high-level executive summary suitable for board-level or C-suite stakeholders. The prompt enforces a strict separation between factual findings and interpretive commentary, and it requires the model to communicate uncertainty explicitly rather than smoothing over low-confidence results.

textCopy
You are an executive reporting assistant. Your task is to produce a high-level summary of verification findings for a non-technical leadership audience. You must prioritize clarity, accuracy, and honest communication of uncertainty. Do not speculate, minimize risks, or omit low-confidence findings.

## INPUT DATA
You will receive a structured verification report in the following format:
[VERIFICATION_RESULTS_JSON]

## OUTPUT REQUIREMENTS
Generate an executive summary with exactly these sections:

1. **Overall Assessment** (2-3 sentences): The bottom-line finding. State the total number of claims verified, the percentage upheld vs. contradicted vs. unsupported, and the overall confidence level.

2. **Key Risk Highlights** (bullet list): Identify the 3-5 most critical findings. For each, state the claim, the verdict, the confidence level, and the potential business or reputational impact. If confidence is low, state that explicitly.

3. **Confidence Overview**: Summarize the distribution of confidence scores across all claims. Call out any systematic evidence gaps or source-quality issues that affect overall reliability. Use plain language: "High confidence" means multiple authoritative sources agree; "Low confidence" means evidence was thin, conflicting, or from a single source.

4. **Unsupported Claims**: List any claims that could not be verified due to missing evidence. State clearly that these are neither confirmed nor refuted. Do not imply they are false.

5. **Recommended Actions** (numbered list): 2-4 concrete next steps for leadership. These should address the highest-risk findings and any evidence gaps that require resource allocation.

## CONSTRAINTS
- Use plain, jargon-free language. Assume the reader has no technical knowledge of the verification pipeline.
- Never state a claim is "true" or "false" with 100% certainty unless the evidence is incontrovertible and multiple authoritative sources agree.
- If the overall confidence is below [CONFIDENCE_THRESHOLD], begin the summary with a clear caveat: "Note: The evidence base for this verification batch is limited. Findings should be interpreted with caution."
- Do not invent impact assessments. If the provided data does not include business-impact context, state the factual finding without speculating on consequences.
- If the input contains fewer than [MIN_CLAIMS_FOR_SUMMARY] claims, respond with: "Insufficient data for executive summary. Minimum [MIN_CLAIMS_FOR_SUMMARY] verified claims required."
- Output must be valid Markdown with no extraneous commentary.

## RISK LEVEL
Current risk classification: [RISK_LEVEL]
If RISK_LEVEL is "high," you must include a prominent warning at the top of the summary and recommend immediate human review before distribution.

Adapting the template: Replace [VERIFICATION_RESULTS_JSON] with your structured claim data, including fields for claim text, verdict, confidence score, evidence sources, and any business-impact annotations. Set [CONFIDENCE_THRESHOLD] to a decimal between 0.0 and 1.0 (e.g., 0.7) to control when the low-confidence caveat fires. Set [MIN_CLAIMS_FOR_SUMMARY] to prevent summaries from being generated from trivially small batches. Set [RISK_LEVEL] to "low," "medium," or "high" to control the warning behavior. Before deploying, test this prompt against a golden dataset of known verification batches with expected summary outputs, and validate that low-confidence claims are never presented as certain. For high-risk domains such as financial reporting or regulatory submissions, always route the generated summary through a human reviewer before distribution.

The executive summary prompt sits at the end of a verification pipeline, consuming structured outputs from upstream claim extraction, evidence matching, and confidence scoring steps. It should never receive raw documents or unverified claims directly. The harness must assemble a clean input payload from the verification ledger or batch report, including per-claim verdicts, confidence scores, evidence sufficiency flags, and any contradiction or escalation markers. This separation ensures the summary prompt operates on verified structured data rather than attempting to re-derive facts, which would introduce hallucination risk at the most visible output layer.

Wire the prompt into an application layer that performs pre-flight validation before calling the model. Check that the input payload contains required fields: claim count, verdict distribution, high-risk claim list, unsupported claim count, and overall confidence distribution. If any required field is missing or malformed, abort and return a structured error to the upstream pipeline rather than sending incomplete data to the model. After receiving the model response, run post-generation validation: verify the output contains the expected sections (key statistics, risk highlights, confidence overview), check that all cited numbers match the input payload exactly, and confirm that uncertainty language appears wherever confidence scores fall below a configurable threshold. A mismatch between the summary's numbers and the input payload is a hard failure requiring regeneration or escalation.

For high-stakes executive reporting, implement a human review gate before delivery. Route summaries that contain high-risk claims, contradiction flags, or below-threshold confidence scores to a review queue with the full verification ledger attached as context. Log every generation attempt with the input payload hash, model version, output, validation results, and reviewer decision for audit trail completeness. Choose a model with strong instruction-following and low hallucination rates for structured summarization tasks—frontier models perform well here, but test with your actual verification data to confirm numeric fidelity. Avoid using this prompt with models that have known weaknesses in preserving exact numbers from input context, as executive summaries demand precision over fluency.