Missing Evidence Citation Interruption Prompt

This prompt is designed for compliance and audit engineering teams who need a programmatic, pre-release gate to prevent AI-generated content containing unsubstantiated claims from reaching end-users or downstream systems. Its primary job is to act as a post-generation validation step: you feed it a piece of generated text alongside the evidence that was used to produce it, and the prompt returns a structured report identifying every factual assertion that lacks a direct source citation. Use this when your product or internal workflow must satisfy strict auditability requirements, regulatory evidence standards (such as those in finance or healthcare), or an internal policy that mandates traceable sourcing before any content is published, stored, or acted upon.

This prompt is not a replacement for a robust retrieval or evidence-ranking system. It assumes that a retrieval step has already occurred and that you have a set of candidate evidence chunks. Its sole function is to verify that the final generated output is faithfully grounded in that provided evidence. You should integrate this prompt into a validation harness that runs automatically after content generation but before the content is released. The harness should be capable of receiving the structured interruption report, logging it for audit trails, and either blocking the release, routing the content for human review, or triggering a retry with more targeted evidence retrieval. Do not use this prompt for real-time, low-latency chat applications where a multi-second validation delay is unacceptable, or for purely creative tasks where factual grounding is not a requirement.

Before implementing, ensure your pipeline can clearly separate the 'generated content' from the 'source evidence' as distinct inputs. A common failure mode is providing the model with a single, large block of text that mixes both, which degrades its ability to accurately match claims to sources. Instead, structure your input with a dedicated field for the final draft and a separate list of evidence chunks, each with a unique identifier. The next step is to copy the prompt template into your validation service and define the exact JSON schema you expect as output, which will allow you to programmatically parse the list of uncited statements and enforce a hard block if the list is not empty.

This prompt template is designed to be used as a system instruction or a validation step in a compliance-sensitive AI workflow. Its job is to interrupt the model's output when a generated claim, decision, or summary lacks a required source citation. Instead of allowing an ungrounded statement to reach a user or a downstream system, the prompt forces the model to stop and produce a structured report that identifies exactly which statements are uncited and what kind of evidence is needed. This is critical for audit, legal, and clinical review workflows where traceability is non-negotiable.

textCopy
You are an evidence-citation auditor. Your only job is to review the provided [OUTPUT_TEXT] against the [CITATION_REQUIREMENTS] and determine whether every factual claim, decision, or summary statement is properly grounded with a source citation.

[CITATION_REQUIREMENTS]:
- Citation format required: [CITATION_FORMAT, e.g., "[Source: document_id, section]"]
- Minimum citation threshold: [THRESHOLD, e.g., "Every factual claim must have at least one citation"]
- Exemptions: [EXEMPTIONS, e.g., "Common knowledge statements, procedural instructions"]

[OUTPUT_TEXT]:
"""
[INSERT THE TEXT TO AUDIT HERE]
"""

[AVAILABLE_SOURCES]:
[LIST OF SOURCE IDENTIFIERS AND THEIR DESCRIPTIONS]

Your task:
1. Extract every factual claim from the [OUTPUT_TEXT].
2. For each claim, check if it is supported by an explicit citation that matches the [CITATION_FORMAT] and references one of the [AVAILABLE_SOURCES].
3. If ALL claims are properly cited, respond ONLY with: "CITATION_CHECK_PASSED: All claims are grounded."
4. If ANY claim is uncited or has an invalid citation, STOP. Do not proceed. Respond ONLY with a structured interruption report in the following JSON format:

{
  "status": "HALTED_MISSING_CITATIONS",
  "uncited_claims": [
    {
      "claim_text": "The exact uncited statement.",
      "claim_location": "Approximate position in the text (e.g., 'Paragraph 2, Sentence 3').",
      "required_evidence_type": "What kind of source would ground this claim (e.g., 'clinical trial result', 'contract clause', 'financial record')."
    }
  ],
  "invalid_citations": [
    {
      "claim_text": "The statement with the bad citation.",
      "provided_citation": "The citation that was found.",
      "failure_reason": "Why it failed (e.g., 'Source not in AVAILABLE_SOURCES', 'Wrong format')."
    }
  ],
  "instruction": "Generation halted. Please ground the identified claims with valid citations from the available sources before release."
}

[CONSTRAINTS]:
- Do not generate any text outside the JSON report or the single pass message.
- Do not guess or invent citations.
- If the [OUTPUT_TEXT] is empty, respond with {"status": "HALTED_EMPTY_INPUT", "instruction": "No text provided for audit."}

To adapt this template, replace the square-bracket placeholders with your specific values. Define [CITATION_FORMAT] precisely, such as [Source: doc_id, para]. Set the [THRESHOLD] to match your risk tolerance; a strict audit might require every sentence to be cited, while a summary might only need citations for quantitative data. Populate [AVAILABLE_SOURCES] with the identifiers the model is allowed to reference. The [EXEMPTIONS] field is critical for preventing false-positive halts on transitional or instructional language. After pasting this prompt, test it with a mix of fully cited, partially cited, and uncited texts to calibrate the strictness before wiring it into a production harness.

This prompt is designed to sit between a content-generation step and a publication or release step in a compliance-sensitive pipeline. It should be invoked after the model produces a claim, decision, or summary but before that output reaches an end user, a database, or a downstream system. The harness receives the generated text and the expected citation sources, passes them into the prompt, and then evaluates the structured output to decide whether to proceed, retry, or escalate. The prompt itself is stateless—it does not remember prior runs—so the harness must supply all required context on each invocation.

Implement the harness as a post-generation gate with a strict JSON output contract. After the primary model generates text, call a separate model instance with this prompt, passing the generated text as [GENERATED_OUTPUT] and the available evidence list as [AVAILABLE_SOURCES]. Parse the response into a structured object with fields: status (PASS or FAIL), uncited_claims (array of claim strings), and missing_evidence_types (array of evidence categories needed). If status is FAIL, do not release the output. Instead, log the failure, attach the uncited claims to the review item, and route to a human review queue. Implement a retry loop with a maximum of two attempts: on the first FAIL, feed the uncited claims back to the generation model with a request to either cite or remove them. If the second attempt also fails, escalate to human review without further retries. This prevents infinite loops while giving the system one chance to self-correct.

For model selection, use a model with strong instruction-following and JSON output reliability. GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro are suitable. Avoid smaller or older models that may hallucinate citations or produce malformed JSON. Set temperature to 0 or near-zero to maximize deterministic behavior. Enable structured output mode if the provider supports it (e.g., OpenAI's response_format with a JSON schema, or Anthropic's tool-use with a defined schema). Validate the response programmatically before acting on it: confirm the status field is exactly PASS or FAIL, that uncited_claims is an array of non-empty strings, and that missing_evidence_types is an array. If validation fails, treat it as a FAIL and escalate to human review with a note that the interruption prompt itself produced an invalid response.

Log every invocation with the input text hash, the source count, the output status, and the timestamp. This creates an audit trail for compliance reviewers. For high-risk domains such as healthcare, legal, or finance, do not allow automatic retry—route all FAIL results directly to human review. In lower-risk domains, you may configure the harness to automatically strip uncited claims from the output and release only the cited portions, but this requires additional post-processing logic to reconstruct a coherent final text. Test the harness with known-good and known-bad inputs before deployment: a fully cited paragraph should return PASS, while a paragraph with fabricated statistics should return FAIL with those statistics listed in uncited_claims.