Medical Claim Verification Escalation Prompt

This prompt is designed for healthcare AI systems that encounter clinical claims from users, generated text, or retrieved documents. Its job is not to verify the claim. Its job is to detect that a claim requires verification, assess whether the system has sufficient evidence to support it, and if not, produce a structured escalation payload for a qualified human reviewer. Use this prompt when your application operates in a clinical context where an unverified assertion could influence a care decision, a patient's understanding, or a documentation record.

Do not use this prompt for general health curiosity questions, administrative scheduling, or billing inquiries where clinical truth value is not at stake. The prompt assumes you have already extracted the claim and any surrounding context. It does not perform extraction itself. Deploy this prompt only when the output will be consumed by a downstream routing system that can respect the escalation payload's structure, including the risk classification, evidence gap summary, and required reviewer qualifications.

Before integrating this prompt, ensure your pipeline can handle the escalation payload as a blocking or non-blocking event. For high-risk clinical settings, the escalation should block the user-facing response until a human reviewer resolves the claim. For lower-risk documentation assistance, the escalation may log the claim for asynchronous review while allowing the workflow to continue with a visible disclaimer. Never use this prompt to silently discard unverified claims or to generate a definitive clinical answer when evidence is missing.

This prompt template is designed to be placed in your system instructions or as a dedicated classification step before any downstream processing of user-submitted medical content. It forces the model to separate verified clinical statements from unverified claims, personal anecdotes, or speculative health advice. The output is a structured escalation decision that your application can use to either proceed with an AI-generated response or route the content to a qualified human reviewer. Replace every square-bracket placeholder with values that match your organization's clinical governance policies, review queue taxonomy, and risk tolerance before sending this to the model.

textCopy
You are a medical claim verification classifier. Your only job is to analyze user-submitted content, identify medical or clinical claims, and determine whether each claim can be verified against provided evidence or requires escalation to a human reviewer.

## INPUT
User content: [USER_INPUT]

Provided evidence (may be empty): [EVIDENCE_CONTEXT]

## CLASSIFICATION RULES
1. Extract every distinct medical, clinical, or health-related claim from the user content.
2. For each claim, classify it into exactly one of these categories:
   - VERIFIED: The claim is directly supported by the provided evidence with a clear source.
   - UNVERIFIED: The claim cannot be confirmed or contradicted by the provided evidence.
   - CONTRADICTED: The claim directly conflicts with the provided evidence.
   - OUT_OF_SCOPE: The claim requires licensed medical diagnosis, treatment prescription, or individual clinical judgment.
3. If any claim is classified as UNVERIFIED, CONTRADICTED, or OUT_OF_SCOPE, the entire request must be escalated.

## ESCALATION RULES
- Escalate if the user content contains any claim about treatment efficacy, drug dosage, supplement safety, or disease prognosis that is not fully supported by the provided evidence.
- Escalate if the user content requests a diagnosis, interprets lab results, or recommends a specific treatment plan for an individual.
- Escalate if the user content references a specific patient case, personal health history, or identifiable individual.
- Do NOT escalate if the user is asking for general educational information about a medical topic with no claims about specific outcomes or individuals.

## OUTPUT SCHEMA
Return a JSON object with this exact structure:
{
  "requires_escalation": boolean,
  "escalation_reason": string or null,
  "claims_found": [
    {
      "claim_text": string,
      "classification": "VERIFIED" | "UNVERIFIED" | "CONTRADICTED" | "OUT_OF_SCOPE",
      "evidence_gap": string or null,
      "source_required": string or null
    }
  ],
  "recommended_review_queue": "[QUEUE_CLINICAL_REVIEW]" | "[QUEUE_SAFETY_URGENT]" | "[QUEUE_GENERAL_INQUIRY]" | null,
  "confidence": number between 0 and 1
}

## CONSTRAINTS
- Never provide medical advice, diagnosis, or treatment recommendations.
- Never speculate about what missing evidence might show.
- If no evidence is provided, classify all clinical claims as UNVERIFIED.
- If the user content contains no medical claims, set requires_escalation to false and claims_found to an empty array.
- For OUT_OF_SCOPE claims, the evidence_gap field should state "Requires licensed clinical judgment" and source_required should state "Qualified healthcare professional review."
- Set confidence below 0.7 if the user content is ambiguous or contains mixed verified and unverified claims.

## EXAMPLES
[FEW_SHOT_EXAMPLES]

## RISK LEVEL
Current risk threshold: [RISK_LEVEL]
- LOW: Escalate only OUT_OF_SCOPE claims.
- MEDIUM: Escalate UNVERIFIED and OUT_OF_SCOPE claims.
- HIGH: Escalate everything except fully VERIFIED claims with high-confidence evidence match.

After pasting this template, replace the placeholders with your operational values. [QUEUE_CLINICAL_REVIEW], [QUEUE_SAFETY_URGENT], and [QUEUE_GENERAL_INQUIRY] should map to your actual review queue identifiers or routing keys. [FEW_SHOT_EXAMPLES] should contain 2-4 representative examples showing correct classification and escalation decisions for your domain. [RISK_LEVEL] should be set dynamically by your application based on user context, regulatory requirements, or session risk scoring. The [EVIDENCE_CONTEXT] placeholder should be populated from your retrieval system or left as an empty string when no grounding evidence is available. Always test this prompt against a golden dataset of known medical claims before deploying, and monitor the escalation rate to ensure you are not over-escalating benign educational queries or under-escalating clinical risk.

This prompt is designed to be a decision gate in a healthcare AI pipeline, not a conversational response. It should be called after a model generates or encounters a clinical claim, and its output must be parsed by application code to determine whether the claim is routed to a human review queue or allowed to proceed. The prompt's primary output is a structured JSON object containing an escalation_decision enum (ESCALATE, FLAG_FOR_REVIEW, PASS), a risk_level, and a list of evidence_gaps. The application layer must treat this output as the single source of truth for routing logic.

Integration flow: When a clinical claim is detected, the application assembles the prompt by injecting the [CLAIM_TEXT], [SOURCE_CONTEXT], and [CLINICAL_POLICY_BOUNDARIES] placeholders. The model response must be validated against a strict JSON schema before routing occurs. If the response fails schema validation, the application should retry once with an explicit repair prompt that includes the validation error. If the retry also fails, the claim must be escalated to human review by default—never pass an unvalidated claim through. Log every decision, including the raw model output, validation result, and final routing action, to an immutable audit trail.

Retry and fallback logic: Use a model with strong JSON mode or structured output support (e.g., GPT-4o with response_format or Claude with tool-use extraction). Set temperature=0 to minimize variance. If the model returns a valid JSON structure but the escalation_decision field contains an unexpected value, treat it as an ESCALATE event. Implement a circuit breaker: if more than 5% of requests in a rolling window fail validation, halt automated processing and alert the operations team. For high-throughput systems, consider a two-stage architecture where a cheaper classifier model performs initial triage and only ambiguous cases reach this detailed escalation prompt.

Human review routing: When the decision is ESCALATE or FLAG_FOR_REVIEW, the application must package the original claim, the model's evidence_gaps list, the risk_level, and the recommended_reviewer_role into a structured ticket for the review queue. Do not expose the raw prompt or model output directly to reviewers; instead, present a formatted summary that highlights the specific gaps and the clinical policy boundary that triggered escalation. Include a link to the full audit log. Reviewers should be able to confirm, overturn, or request more evidence, and their decision must be logged back into the system to close the feedback loop for future calibration.

What to avoid: Never use this prompt's output to automatically reject a claim or deny information to a user without human confirmation when risk_level is HIGH or CRITICAL. Do not skip validation even if the model appears to return well-formed JSON consistently—schema drift is a leading cause of silent failures in production. Avoid routing evidence_gaps directly into a follow-up prompt that attempts to fill those gaps automatically, as this can create hallucination loops where the model invents supporting evidence to pass its own verification check.