Evidence Completeness Gap Analysis Prompt Template

This prompt is for verification analysts and pipeline engineers who need to identify what evidence is missing before a claim can be scored with confidence. It produces a structured gap report listing missing evidence types, their expected impact on confidence, and retrieval suggestions. The core job-to-be-done is distinguishing 'proven false' from 'not yet proven'—a critical signal for downstream scoring models that must avoid silently treating an absence of evidence as a negative determination. Use this prompt after claim extraction and before evidence sufficiency scoring, when your verification workflow requires explicit gap signals to drive retrieval expansion, human-review routing, or confidence suppression.

The ideal user has already extracted discrete, atomic claims from source content and has run an initial evidence retrieval pass that may be incomplete. The required context includes the claim text, the domain or topic area, a summary of evidence already retrieved (with source types), and the target confidence threshold the pipeline must reach. The prompt is designed to work with structured inputs and expects a defined output schema for gap reporting. It is not a general-purpose research assistant; it assumes a verification pipeline context where the cost of missing evidence is measurable. For high-risk domains such as healthcare, finance, or legal compliance, the gap report should feed into a human-review queue rather than auto-triggering additional retrieval without oversight.

Do not use this prompt when you already have a complete evidence set and only need a confidence score—use the Evidence Sufficiency Scoring Prompt Template instead. Do not use it for binary true/false classification without a gap analysis step, as that conflates missing evidence with negative evidence. Avoid this prompt when the claim is inherently unverifiable (opinion, prediction, or subjective framing), where a Fact-Interpretation Separation prompt should run first. Finally, do not use this prompt as a standalone verification step; its output is a diagnostic artifact that must feed into a scoring, retrieval, or review action. The next step after generating a gap report is typically to route high-impact gaps to targeted retrieval, flag the claim for human review, or adjust the confidence score downward with explicit gap documentation.

This template is the core instruction set for an Evidence Completeness Gap Analysis. It is designed to be dropped into an AI harness where the model receives a specific claim and an inventory of already-retrieved evidence. The prompt instructs the model to act as a verification analyst, systematically identifying what evidence is missing, why it matters for confidence scoring, and how to find it. Use this template as the starting point for building a gap detection step in your verification pipeline.

codeCopy
You are a meticulous verification analyst. Your task is to perform an evidence completeness gap analysis on a single claim.

**Claim to Analyze:**
[CLAIM_TEXT]

**Available Evidence Inventory:**
[EVIDENCE_INVENTORY]

**Analysis Instructions:**
1.  **Deconstruct the Claim:** Break down [CLAIM_TEXT] into its core verifiable components (e.g., who, what, when, where, numerical value).
2.  **Map Evidence to Components:** For each component, state whether the [EVIDENCE_INVENTORY] provides direct, indirect, or no supporting evidence.
3.  **Identify Critical Gaps:** List the specific types of evidence that are missing. For each gap, explain:
    - **Missing Evidence Type:** (e.g., primary source document, timestamped record, authoritative database entry, direct quote verification).
    - **Impact on Confidence:** Estimate the potential shift in a confidence score if this evidence were found (e.g., "Low to Medium," "Medium to High").
    - **Retrieval Suggestion:** Propose a specific search query, database, or source type that could fill this gap.
4.  **Assess Unverifiable Components:** Flag any claim components that are inherently unverifiable (e.g., opinions, future predictions, subjective statements) and exclude them from the evidence gap analysis.

**Output Format:**
Return a single JSON object matching this exact schema:
{
  "claim_deconstruction": [
    {
      "component": "string",
      "evidence_status": "direct | indirect | none",
      "source_reference": "string | null"
    }
  ],
  "critical_gaps": [
    {
      "missing_evidence_type": "string",
      "confidence_impact": "string",
      "retrieval_suggestion": "string"
    }
  ],
  "unverifiable_components": ["string"],
  "overall_gap_severity": "low | medium | high | critical"
}

**Constraints:**
- Do not invent evidence. Only use the provided [EVIDENCE_INVENTORY].
- If the inventory is empty, treat all factual components as having "none" status.
- The `overall_gap_severity` must be "critical" if any core factual component has no evidence and no clear retrieval path.

To adapt this template, replace the [CLAIM_TEXT] and [EVIDENCE_INVENTORY] placeholders with your application's data. The [EVIDENCE_INVENTORY] should be a structured list of source summaries, not raw documents, to keep the context window focused. For high-stakes domains like healthcare or finance, add a [RISK_LEVEL] constraint that forces a "critical" severity rating and halts the pipeline for human review if evidence for a regulated claim component is missing. The output schema is designed for direct parsing by a downstream confidence scoring or human-review routing module.

The gap analysis prompt is not a standalone tool; it is a component in a verification pipeline. It should be called after claim extraction and evidence retrieval, but before confidence scoring. The prompt expects a structured claim object and a list of retrieved evidence items, and it returns a gap report. Wire this prompt into your application as a synchronous step in a directed acyclic graph (DAG) or as a tool call within an agent loop. The output should be validated before downstream scoring or human review routing consumes it.

Input assembly: Collect the claim text, claim type, and the full list of retrieved evidence items (with source metadata) from upstream steps. Populate the [CLAIM] and [EVIDENCE_LIST] placeholders. If your system uses a domain taxonomy, inject it into [DOMAIN_CONTEXT] to improve gap classification. Model selection: Use a model with strong reasoning and structured output capabilities (e.g., GPT-4o, Claude 3.5 Sonnet). Avoid smaller or older models that may conflate missing evidence with contradictory evidence. Output validation: Parse the JSON response and validate that each gap object contains the required fields: gap_type, missing_evidence_description, expected_confidence_impact, and retrieval_suggestions. Reject outputs with missing fields or non-enum gap_type values. Retry logic: If validation fails, retry once with the validation error message appended to the prompt as a [PREVIOUS_ERROR] context. If the second attempt fails, log the failure and route the claim to a human review queue with the raw output attached.

Human review integration: Gap reports with expected_confidence_impact rated as high or critical should trigger a review task. Package the original claim, the gap report, and the evidence list into a structured review packet. Do not auto-score claims with critical gaps. Logging and observability: Log the input claim ID, the number of evidence items provided, the number of gaps returned, and the validation status. Track gap detection recall by periodically comparing gap report outputs against human-annotated gap labels on a golden dataset. What to avoid: Do not use this prompt to score confidence directly; it identifies what is missing, not how confident you should be. Do not feed the gap report back into the same model for scoring without first attempting to retrieve the missing evidence types it suggests.