Agent Belief Revision Prompt on Conflicting Evidence

This prompt is for agent developers and infrastructure engineers who need an autonomous agent to update its internal belief state when presented with new evidence that contradicts previously accepted information. The core job-to-be-done is rational belief revision: the agent must not simply overwrite old facts with new ones, but must weigh source reliability, explain the conflict, adjust its confidence, and produce a revised, uncertainty-aware belief state that downstream agents or reasoning steps can safely consume. The ideal user is building a multi-agent system where a research agent, fact-checking agent, or knowledge-base agent must maintain a coherent world model across multiple turns and sources.

Use this prompt when the agent operates in an environment where information quality varies, sources may disagree, and the cost of acting on outdated or incorrect beliefs is high—such as in due diligence, intelligence analysis, scientific literature review, or incident investigation. The prompt requires the following inputs to be effective: the agent's current belief state (structured with claims, confidence scores, and source provenance), the new conflicting evidence (with its own source metadata), and a defined conflict resolution policy (e.g., prefer recency, prefer authority, require human review above a risk threshold). Do not use this prompt for simple fact-update tasks where no conflict exists, or for agents that do not maintain persistent, structured belief states. For stateless Q&A or single-turn retrieval, a standard RAG prompt with source conflict handling is more appropriate.

Before wiring this into a production agent loop, define the evaluation criteria for a successful revision. A valid revision must: (1) explicitly identify the conflicting claims, (2) explain why the conflict exists (source disagreement, temporal change, domain mismatch), (3) adjust confidence scores upward or downward with justification, (4) preserve the revision history for auditability, and (5) flag beliefs that now fall below an actionable confidence threshold. If the agent's revised belief will drive automated actions (e.g., triggering an alert, updating a database, or changing a downstream agent's plan), require a human-in-the-loop review step for any revision where the confidence delta exceeds a predefined boundary or where the risk level is marked high. Start with a well-defined belief schema and a small set of conflict scenarios before scaling to open-ended evidence streams.

This prompt template is designed to be wired into an agent's reasoning loop when it receives new information that contradicts its current belief state. The template forces the model to surface the conflict explicitly, weigh evidence by source reliability and recency, produce a revised belief with an updated confidence score, and explain what changed and why. Use this when your agent ingests data from multiple tools, APIs, or retrieval sources that may disagree. Do not use this for simple fact lookup where no prior belief exists; that workflow belongs to a standard grounding or RAG prompt.

textCopy
You are an agent performing belief revision. Your current belief state is provided below along with new evidence that may confirm, contradict, or extend it.

CURRENT BELIEF STATE:
[BELIEF_STATE]

NEW EVIDENCE:
[NEW_EVIDENCE]

SOURCES AND RELIABILITY:
[SOURCE_METADATA]

CONSTRAINTS:
[CONSTRAINTS]

Your task is to produce a revised belief state following strict rational revision principles:

1. IDENTIFY all conflicts between the current belief state and the new evidence. List each conflict with the specific claims that disagree.
2. WEIGH each piece of evidence by source reliability, recency, and specificity. Prefer more recent evidence from higher-reliability sources when conflicts cannot be resolved.
3. REVISE the belief state to incorporate the new evidence. You may:
   - Update a belief if new evidence is stronger.
   - Retain a belief if existing evidence is stronger.
   - Add a new belief if the evidence introduces a novel claim.
   - Retract a belief if it is directly contradicted by stronger evidence.
   - Flag a belief as CONTESTED if evidence is evenly balanced and no resolution is possible.
4. ADJUST confidence scores for every belief. Increase confidence when corroborated, decrease when contradicted, and set to a low value for contested beliefs.
5. EXPLAIN each revision decision with a brief rationale citing the specific evidence and source that drove the change.

OUTPUT SCHEMA:
Return a valid JSON object with this exact structure:
{
  "conflicts_identified": [
    {
      "current_belief": "string",
      "conflicting_evidence": "string",
      "conflict_type": "direct_contradiction | partial_overlap | new_information"
    }
  ],
  "revised_belief_state": [
    {
      "belief_id": "string",
      "statement": "string",
      "confidence": 0.0-1.0,
      "status": "confirmed | updated | added | retracted | contested | unchanged",
      "supporting_sources": ["source_id"],
      "revision_rationale": "string"
    }
  ],
  "revision_summary": "string",
  "unresolved_uncertainties": ["string"]
}

Do not fabricate evidence. Do not resolve conflicts by ignoring contradictory evidence. If evidence is evenly balanced, mark the belief as contested rather than picking a side arbitrarily.

Adapt this template by replacing the placeholders with your agent's actual data structures. [BELIEF_STATE] should contain the agent's current beliefs as structured objects with IDs, statements, and confidence scores. [NEW_EVIDENCE] should contain the incoming claims with their source IDs. [SOURCE_METADATA] should map source IDs to reliability scores, timestamps, and type labels. [CONSTRAINTS] can include domain-specific rules such as 'prefer peer-reviewed sources over web forum posts' or 'do not revise beliefs about [TOPIC] without human review.' For high-risk domains like healthcare or finance, add a human approval gate before the revised belief state is committed to the agent's memory. Always validate the output JSON against the schema before accepting the revision, and log the full revision record for auditability.

This prompt is not a one-shot query; it is a state transition function inside an agent's reasoning loop. The agent receives new evidence, compares it against its current belief state, and calls this prompt to produce a revised belief state. The harness must therefore manage the before and after states explicitly. Before calling the model, serialize the agent's current beliefs, the new evidence, and all source metadata into the [CURRENT_BELIEFS], [NEW_EVIDENCE], and [SOURCE_METADATA] placeholders. After receiving the response, validate the output against the expected [OUTPUT_SCHEMA] before overwriting the agent's belief state. A failed validation should trigger a retry with the validation error appended to [CONSTRAINTS], not a silent state corruption.

Wire the prompt into a stateful agent service with a strict write path: (1) lock the agent's belief state, (2) assemble the prompt with current beliefs and conflicting evidence, (3) call the model with a low temperature (0.1–0.3) to favor deterministic revision logic, (4) validate the JSON output against the schema—check that revised_beliefs, conflict_explanation, and uncertainty_quantification fields are present and well-formed, (5) log the revision diff for auditability, and (6) commit the new belief state and release the lock. For high-stakes domains such as healthcare or finance, insert a human review gate between steps 4 and 6 when the uncertainty_quantification exceeds a configured threshold or when the conflict_explanation indicates unresolved contradictions. Use a structured log entry that captures the prior beliefs, the triggering evidence, the model's revision, and the reviewer's decision.

Model choice matters here. Use a model with strong reasoning capabilities (GPT-4o, Claude 3.5 Sonnet, or equivalent) because belief revision requires comparing propositions, detecting contradictions, and adjusting confidence scores—tasks where smaller or older models often produce inconsistent uncertainty estimates. Implement retry logic with exponential backoff for transient API failures, but cap retries at 3 attempts. If validation fails on all retries, escalate to a human operator and preserve the unmodified belief state. For multi-agent systems, ensure that the belief state update is atomic: no other agent should read a partially updated belief state. Finally, run periodic eval suites using the provided test cases to detect regression in revision quality, especially around edge cases where evidence is equally weighted or sources have conflicting reliability scores.