Evidence Threshold Decision Prompt Template

This prompt is for autonomous or semi-autonomous agents that reach a decision point and must choose between proceeding, escalating to a human, or gathering more evidence. It applies configurable evidence sufficiency thresholds to each decision, producing a structured decision with rationale and specific evidence gaps when the threshold is not met. Use this prompt inside an agent's planning or execution loop whenever an action has a pre-defined evidence bar that must be cleared before the agent is allowed to commit. This is not a general-purpose decision prompt; it is specifically designed for workflows where the cost of a wrong decision is high, the evidence is incomplete, and the decision logic must be auditable.

The ideal user is an engineering lead or AI builder wiring decision logic into a production agent. Required context includes the decision to be made, the available evidence items with source metadata, the configured sufficiency threshold for this decision class, and the available actions (proceed, escalate, gather evidence). Without this structured context, the prompt degrades into an unbounded reasoning task that produces inconsistent, un-auditable outputs. The prompt expects evidence items to carry fields like source, recency, relevance_score, and content. If your evidence pipeline cannot produce these fields, implement the Evidence Sufficiency Evaluation Prompt Template first to normalize raw observations into structured evidence before reaching this decision point.

Do not use this prompt for low-stakes decisions where a wrong choice is cheap to reverse. Do not use it when the agent has only one available action, when the evidence is always complete, or when the decision can be reduced to a simple threshold on a single numeric score. In those cases, a deterministic rule in application code is cheaper, faster, and easier to test. This prompt earns its cost only when the decision involves multiple evidence dimensions, trade-offs between speed and certainty, and an audit trail requirement. If you find yourself calling this prompt for every micro-decision in a loop, reconsider whether the loop itself should be restructured to batch decisions or defer to a higher-level planner.

This prompt template is the core decision engine for an agent that must decide whether to proceed, escalate, or gather more evidence. It forces the model to apply explicit evidence thresholds to each decision point, producing a structured decision with a clear rationale and a specific list of evidence gaps when the threshold is not met. Use this template when your agent operates in a domain where acting on insufficient evidence carries material cost, such as approving a financial transaction, executing an irreversible infrastructure change, or making a clinical recommendation.

textCopy
You are an evidence-driven decision agent. Your task is to evaluate a proposed action against a configurable evidence sufficiency threshold and produce a structured decision.

## INPUT
[ACTION_DESCRIPTION]
[AVAILABLE_EVIDENCE]
[CONTEXT]

## THRESHOLD CONFIGURATION
- Required Evidence Categories: [REQUIRED_EVIDENCE_CATEGORIES]
- Minimum Source Count per Category: [MIN_SOURCE_COUNT]
- Minimum Source Reliability Score (0.0-1.0): [MIN_RELIABILITY_SCORE]
- Contradiction Policy: [CONTRADICTION_POLICY] (options: "block", "flag_for_review", "ignore")
- Missing Evidence Policy: [MISSING_EVIDENCE_POLICY] (options: "block", "escalate", "proceed_with_warning")

## OUTPUT SCHEMA
Return a single JSON object with the following structure:
{
  "decision": "proceed" | "escalate" | "gather_evidence",
  "confidence_score": 0.0-1.0,
  "rationale": "string explaining how the evidence was weighed against the threshold",
  "evidence_gaps": [
    {
      "category": "string",
      "description": "string describing what specific evidence is missing",
      "criticality": "blocking" | "warning",
      "suggested_collection_method": "string"
    }
  ],
  "contradictions_found": [
    {
      "sources": ["source_a", "source_b"],
      "description": "string describing the contradiction",
      "resolution_suggestion": "string"
    }
  ],
  "override_log": null | {
    "reason": "string",
    "authorized_by": "string",
    "timestamp": "ISO8601"
  }
}

## CONSTRAINTS
- Do not invent evidence. If a required category has no evidence, list it as a gap.
- If the contradiction policy is "block" and contradictions exist, the decision must be "escalate" or "gather_evidence".
- If the missing evidence policy is "block" and critical gaps exist, the decision must not be "proceed".
- Confidence score must reflect evidence completeness, source reliability, and contradiction severity.
- If an override is present, log it explicitly and still report the evidence-based assessment.

## EXAMPLES
[FEW_SHOT_EXAMPLES]

## RISK LEVEL
[RISK_LEVEL] (options: "low", "medium", "high", "critical")

To adapt this template, replace each square-bracket placeholder with your specific inputs. The threshold configuration block is the most critical section to customize: define your required evidence categories based on your domain's risk taxonomy, set minimum source counts and reliability scores that match your operational tolerance, and choose contradiction and missing-evidence policies that align with your escalation paths. The few-shot examples should include at least one case where the threshold is met, one where it is not met due to missing evidence, and one where a contradiction triggers escalation. Wire the output JSON into a downstream validator that checks for schema compliance and policy adherence before the decision is acted upon. For high-risk or critical risk levels, always route the decision through a human review queue before execution, regardless of the model's confidence score.

The Evidence Threshold Decision prompt is designed to be called at specific decision gates within an agent's execution loop, not as a standalone chat interaction. It should be invoked whenever the agent reaches a point where the next action is irreversible, costly, or depends on uncertain information. The prompt expects a structured input containing the proposed action, available evidence items with their confidence scores, and a configured threshold policy. The output is a machine-readable decision object that your application code can parse and act on without additional model interpretation.

To integrate this prompt into an application harness, wrap it in a function that accepts three required inputs: action_description, evidence_items (an array of objects with source, content, confidence, and freshness_timestamp fields), and threshold_policy (a configuration object specifying minimum_confidence, required_evidence_count, and irreversibility_level). The function should construct the prompt by injecting these values into the template's [ACTION], [EVIDENCE_LIST], and [THRESHOLD_CONFIG] placeholders. After receiving the model response, parse the JSON output and validate it against a strict schema that requires decision (enum: PROCEED, ESCALATE, GATHER_MORE), rationale (string), confidence_score (float 0-1), and evidence_gaps (array of strings). If parsing fails, retry once with the validation error injected into the [PREVIOUS_ERROR] placeholder. If the retry also fails, default to ESCALATE and log the raw response for debugging.

For production deployments, implement a decision override registry that logs every instance where the application code overrides the model's decision. Common override scenarios include: operator-in-the-loop approvals for high-irreversibility actions, time-bound auto-escalation when the agent has been gathering evidence beyond a configured timeout, and safety-policy blocks for actions on an explicit deny list. Each override should be logged with the original model decision, the override reason, the operator identity (if human), and a timestamp. This audit trail is essential for tuning threshold policies and defending the system's behavior during reviews. Additionally, wire the evidence_gaps output into your agent's evidence collection module so that GATHER_MORE decisions automatically trigger targeted retrieval or clarification requests rather than generic re-prompting.

Model choice matters for this prompt. Use a model with strong instruction-following and structured output capabilities, such as GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro. Avoid smaller or older models that may conflate confidence language or hallucinate evidence items not present in the input. Set temperature to 0 or a very low value (0.1 maximum) to reduce variance in threshold judgments. If your application requires explainability for compliance, store the full prompt and response pair alongside the parsed decision. For high-throughput systems, consider caching decisions for identical (action, evidence, threshold) tuples with a short TTL, but invalidate the cache immediately if any evidence item's freshness timestamp exceeds the policy's maximum age. The most common production failure mode is the model returning PROCEED when evidence is superficially relevant but factually insufficient—mitigate this by including a counter-example in the prompt's [EXAMPLES] section showing a correct GATHER_MORE decision when evidence quantity is high but relevance is low.