Contradiction-Aware Confidence Adjustment Prompt Template

This prompt is designed for verification pipeline developers who have already scored a claim's initial confidence and detected contradictory evidence from multiple sources. It takes the original confidence score, the claim, the supporting evidence, and the contradicting evidence as inputs, then produces a revised confidence score that accounts for contradiction severity, source reliability trade-offs, and unresolved conflicts. The core job-to-be-done is calibrated confidence adjustment—applying a reasoned, evidence-aware penalty rather than a flat reduction or binary rejection. The ideal user is a verification system architect or pipeline engineer who needs to maintain score integrity when evidence is mixed, ensuring that downstream triage thresholds and human-review routing decisions are based on realistic confidence levels.

Use this prompt when your pipeline has completed claim extraction and evidence matching, and you need to adjust confidence downward in a calibrated way. It is not a replacement for initial confidence scoring or evidence retrieval. It assumes you already have a structured claim, a set of supporting evidence items with source metadata, and a set of contradicting evidence items with source metadata. The prompt works best when each evidence item includes a reliability indicator (such as source authority tier, recency, or methodology transparency) so the model can weigh contradictions proportionally. For example, a contradiction from a high-authority primary source should reduce confidence more than a contradiction from a low-tier aggregator. The prompt also requires the original confidence score as a numeric input, typically on a 0-1 or 0-100 scale, so the adjustment is anchored to a prior assessment rather than generated from scratch.

Do not use this prompt when you lack structured evidence metadata, when contradictions are purely stylistic or interpretive rather than factual, or when the initial confidence score was not produced by a consistent scoring method. It is also inappropriate for claims where the contradiction is a direct negation of the claim by a single authoritative source—in those cases, a binary override or human escalation is more appropriate than a calibrated adjustment. After running this prompt, you should validate the output by checking that the revised score is strictly less than or equal to the original score, that the adjustment magnitude correlates with contradiction severity, and that the explanation references specific evidence items rather than vague generalizations. For high-stakes domains such as healthcare, legal, or financial claims, always route the adjusted output to human review with the full evidence packet before accepting the revised score.

This template is designed to be the core instruction set for a contradiction-aware confidence adjustment step in your verification pipeline. It takes an original claim, its initial confidence score, and a structured summary of contradictory evidence as inputs. The model's job is not to re-verify the claim from scratch, but to apply a principled, explainable penalty to the confidence score based on the nature of the contradiction. The output is a revised score and a structured justification, making the adjustment logic auditable.

textCopy
You are a precise verification analyst. Your task is to adjust the confidence score of a claim based on a set of identified contradictions. Do not re-evaluate the claim's truthfulness from scratch. Apply a penalty to the initial confidence score that is proportional to the contradiction's severity, the reliability of the contradicting source, and whether the conflict is resolvable.

INPUTS:
- ORIGINAL_CLAIM: [CLAIM_TEXT]
- INITIAL_CONFIDENCE_SCORE: [INITIAL_SCORE] (a float between 0.0 and 1.0)
- CONTRADICTIONS: [CONTRADICTION_LIST] (a JSON list of objects, each with "contradicting_evidence", "source_authority_score", and "contradiction_type" fields)
- ADJUSTMENT_RULES: [ADJUSTMENT_RULES] (a string describing the penalty logic, e.g., "A direct contradiction from a highly authoritative source should reduce confidence by at least 0.4.")

INSTRUCTIONS:
1. For each contradiction in the list, assess its impact based on the ADJUSTMENT_RULES.
2. Calculate a final adjusted confidence score. The score must be a float between 0.0 and 1.0, and cannot be higher than the INITIAL_CONFIDENCE_SCORE.
3. Classify the contradiction's resolution status as "resolvable," "unresolvable_conflict," or "insufficient_information."
4. Provide a concise, step-by-step justification for the adjustment.

OUTPUT_SCHEMA:
{
  "adjusted_confidence_score": <float>,
  "contradiction_resolution_status": "<status_string>",
  "adjustment_justification": "<step_by_step_reasoning_string>"
}

CONSTRAINTS:
- Do not invent new evidence or contradictions.
- If the CONTRADICTIONS list is empty, return the INITIAL_CONFIDENCE_SCORE unchanged.
- If a contradiction is deemed irrelevant or from a completely unreliable source, explain why its penalty was 0.0.

To adapt this template, replace the square-bracket placeholders with data from your upstream pipeline steps. The [CONTRADICTION_LIST] should be the direct output of a contradiction detection prompt, and the [ADJUSTMENT_RULES] can be a static policy string or dynamically generated based on the claim's domain. After receiving the output, validate the adjusted_confidence_score is within the [0.0, 1.0] range and is less than or equal to the [INITIAL_SCORE]. For high-stakes verification, route outputs where the contradiction_resolution_status is "unresolvable_conflict" directly to a human review queue, packaging the original claim, all evidence, and the model's justification into the review packet.

This prompt is designed to sit downstream of claim extraction and evidence retrieval, receiving a claim, its initial confidence score, and a set of contradictory evidence items. It is not a standalone fact-checker. The harness must supply structured inputs: the original claim text, the pre-contradiction confidence score with its basis, and each contradictory source with its authority score and the specific conflicting assertion. Without this structured context, the model cannot perform calibrated adjustment and will default to generic hedging.

Wire the prompt as a synchronous step in your verification pipeline after contradiction detection has fired. The application layer should validate inputs before calling the model: ensure the initial confidence is a float between 0.0 and 1.0, each contradiction entry includes a source identifier and a severity flag, and source authority scores are normalized to a consistent scale. After receiving the model output, validate the revised confidence score is within [0.0, 1.0] and that the adjustment rationale references specific contradictions rather than vague language. If validation fails, retry once with the validation error appended to the prompt as a correction hint. Log every adjustment decision—original score, contradictions presented, revised score, and rationale—to an audit trail for downstream governance review.

For high-stakes domains such as healthcare claims or financial disclosures, route any adjustment exceeding a 0.3 magnitude change to a human review queue with the full contradiction packet and model rationale attached. Use a structured output schema with fields for revised_confidence, adjustment_rationale, unresolved_conflicts, and over-penalization_check to make automated threshold routing reliable. Avoid using this prompt for claims with more than five active contradictions; instead, split the contradiction set or escalate for human analysis, as the model's severity calibration degrades with excessive conflicting signals.