Domain Expert Routing Prompt for Verification

This prompt is designed for verification pipeline operators who need to route extracted claims to the correct domain expert for review. Use it when your system processes claims spanning legal, financial, medical, and technical domains and you need a reliable classification step before assigning human reviewers or specialized auto-verification models. The prompt expects a claim that has already been extracted and decomposed into an atomic, verifiable statement. It is not designed for multi-hop claims, claims embedded in long documents without prior extraction, or claims where the domain is intentionally obscured. Deploy this prompt after claim extraction and before evidence matching or human review assignment.

The ideal input is a single, self-contained factual assertion stripped of rhetorical framing. For example, 'The defendant filed the motion on March 14, 2024' is ready for routing, while 'The legal strategy was flawed from the start' requires further decomposition before this prompt can classify it reliably. The prompt produces a domain label, a confidence score, and a set of required expertise tags. You should configure your pipeline to route claims with confidence below 0.85 to a general triage queue for human review, and claims classified as 'medical' or 'legal' with high confidence directly to domain-specialist reviewers. Never use this prompt as a substitute for actual domain verification; it only decides who should perform the verification.

Before deploying, validate that your upstream claim extraction produces atomic statements with preserved context. If claims arrive as compound sentences or paragraphs, the router will produce unstable classifications. Also ensure your downstream routing logic respects the confidence threshold and never silently auto-routes low-confidence classifications. Start by running this prompt against a labeled test set of 200 claims evenly distributed across your target domains, and measure precision and recall per domain before enabling production traffic. If any domain falls below 90% precision, add few-shot examples for that domain or consider a dedicated classifier model.

This prompt template is designed to be dropped directly into a verification pipeline's routing layer. It takes a claim and its surrounding context, then produces a structured domain classification with confidence scoring and required expertise tags. The output feeds downstream routing logic that assigns the claim to legal, financial, medical, or technical reviewers based on the classification result. Use this template when your verification system handles multi-domain claims and needs consistent, auditable routing decisions before human review begins.

codeCopy
SYSTEM: You are a verification triage classifier. Your job is to analyze a claim and determine which domain expert should review it. You must output valid JSON only, with no additional text.

INPUT:
- Claim: [CLAIM_TEXT]
- Source Context: [SOURCE_CONTEXT]
- Available Domains: [AVAILABLE_DOMAINS]
- Routing Rules: [ROUTING_RULES]

OUTPUT_SCHEMA:
{
  "classification": {
    "primary_domain": "string (one of: legal, financial, medical, technical, general)",
    "secondary_domains": ["string"],
    "confidence": 0.0-1.0,
    "confidence_rationale": "string (explain why this confidence level)"
  },
  "required_expertise": ["string (specific skills needed)"],
  "routing_priority": "string (standard, high, critical)",
  "ambiguity_flags": ["string (any domain overlap or uncertainty)"],
  "escalation_triggers": ["string (conditions that should escalate this claim)"]
}

CONSTRAINTS:
- If confidence is below [MIN_CONFIDENCE_THRESHOLD], set routing_priority to "critical" and flag for multi-reviewer assignment.
- If the claim spans multiple domains, list all relevant domains and set ambiguity_flags accordingly.
- Never classify a claim as "general" if it contains regulated content (legal, financial, medical terminology).
- If the claim involves potential harm or high-stakes decisions, set routing_priority to "high" or "critical" regardless of confidence.

EXAMPLES:
[FEW_SHOT_EXAMPLES]

USER: Classify the claim above according to the schema and constraints.

To adapt this template for your verification pipeline, replace the square-bracket placeholders with your specific inputs. The [CLAIM_TEXT] should contain the atomic claim extracted from your upstream decomposition step. The [SOURCE_CONTEXT] provides surrounding text that helps disambiguate domain boundaries—include at least the paragraph containing the claim. For [AVAILABLE_DOMAINS], list the domains your review team actually covers; this prevents the model from classifying into unsupported categories. The [ROUTING_RULES] placeholder should contain your organization's specific routing logic, such as 'financial claims above $10M go to senior review' or 'medical claims with drug mentions require pharmacology expertise.' The [MIN_CONFIDENCE_THRESHOLD] is your operational cutoff—typically 0.7 to 0.85 depending on your risk tolerance. The [FEW_SHOT_EXAMPLES] section is critical for calibration: include 3-5 examples showing correct classifications for claims that sit at domain boundaries, such as medical-device liability claims that span both medical and legal domains. Test this prompt against a golden dataset of 50-100 expert-labeled claims before deploying, and measure precision and recall per domain to identify systematic misclassifications. If your pipeline handles regulated content, add a human review step for any classification where confidence falls below threshold or where ambiguity_flags is non-empty.

The Domain Expert Routing Prompt is a classification step that must be treated as a critical control point in your verification pipeline. A misrouted claim—sending a medical claim to a legal reviewer or a financial claim to a general queue—creates downstream errors that compound through the verification process. The implementation harness must therefore validate routing decisions before they execute, log every routing event for auditability, and provide escape hatches when the model is uncertain or the domain classification confidence falls below operational thresholds.

Wire this prompt as a synchronous classification service with a strict contract. The application layer should call the model with the prompt template, receive a structured JSON response containing domain, confidence_score, required_expertise_tags, and routing_target, then validate that response before acting on it. Implement a validation guard that rejects any routing decision where confidence_score is below a configurable threshold (start at 0.85 and tune based on production data). For claims that fall below threshold, route to a general triage queue for manual domain assignment rather than guessing. Log the full prompt input, model response, validation result, and final routing decision to your observability platform with a trace ID that connects this routing event to downstream verification outcomes. This traceability is essential for measuring routing precision—you need to compare the model's domain classification against the actual domain expertise required, as determined by reviewer feedback or expert-labeled claims.

Model choice matters here. Use a model with strong zero-shot classification capabilities and low latency, as this prompt sits on the critical path of your verification pipeline. GPT-4o or Claude 3.5 Sonnet are appropriate for high-stakes domains like legal, financial, and medical where misrouting carries regulatory risk. For lower-stakes technical or general content domains, a faster model like Claude 3 Haiku or GPT-4o-mini may suffice. Implement a retry strategy with exponential backoff for transient failures, but cap retries at 2 attempts—a routing decision that can't be made reliably after two tries should escalate to human triage rather than blocking the pipeline. Add a circuit breaker that routes all claims to human review if the model endpoint error rate exceeds 5% in a rolling 5-minute window.

Build eval harnesses that run continuously against a golden dataset of expert-labeled claims across your target domains. Measure precision and recall per domain, not just aggregate accuracy—a model that's 95% accurate overall but only 60% accurate on medical claims is dangerous. Track domain confusion matrices to identify systematic misclassifications (e.g., financial claims routed as legal). Set up alerts when per-domain precision drops below 90% or when the rate of low-confidence classifications spikes, which often indicates prompt drift or a model behavior change. Run these evals before every prompt update and as part of your CI/CD pipeline for the verification service.

For high-risk domains, implement a human-in-the-loop override. Even when the model is confident, allow reviewers to flag misrouted claims and feed those corrections back into your eval dataset. This feedback loop is your primary mechanism for improving routing precision over time. Avoid the temptation to auto-correct routing based on a single reviewer's override—aggregate patterns across multiple reviewers before updating routing rules or retraining classification thresholds. The harness should also enforce domain-specific handling rules: medical claims must never route to non-clinical reviewers, legal claims must preserve attorney-client privilege markers, and financial claims must include relevant regulatory context in the reviewer handoff packet.