Source-to-Claim Alignment Verification Prompt

This prompt is for audit, compliance, and AI engineering teams who need to verify that every factual claim in a generated text is directly supported by a specific source passage. The primary job-to-be-done is traceability at scale: you have a model-generated output (an answer, a summary, a report) and a set of source documents, and you need to prove that the output is grounded. The ideal user is a developer or QA engineer integrating this verification step into a CI/CD pipeline for a RAG application, a compliance officer reviewing AI-generated audit evidence, or a product manager defining acceptance criteria for a cited-answer feature. The required context includes the full generated text, the complete set of retrieved or provided source passages, and a defined alignment tolerance for your use case.

Do not use this prompt when you need a quick, probabilistic hallucination score without passage-level traceability. This prompt performs fine-grained claim decomposition and source mapping, which is computationally and token-intensive. It is overkill for real-time chat moderation or low-risk content generation where a simpler LLM-as-judge eval will suffice. It is also not a substitute for a retrieval step; it verifies alignment between claims and already-provided sources. If your sources are missing or irrelevant, use a retrieval expansion or evidence gap identification prompt first. For regulated industries, the output of this prompt is a decision-support artifact, not a compliance certification—human review of misaligned or stretched claims is mandatory before any filing or submission.

Before using this prompt, ensure you have a clear definition of what constitutes an 'aligned' versus 'stretched' versus 'misaligned' claim for your domain. A financial audit team might require exact figure matching, while a market intelligence brief might accept paraphrased summaries as aligned. Define your alignment distance scoring thresholds upfront. After running this prompt, route high-distance or misaligned claims to a human review queue and use the output to trigger a regeneration or correction workflow. Never silently discard alignment failures in production.

The following prompt template is designed to be dropped into an audit or compliance harness. It takes a generated text block and a set of source passages, then produces a structured alignment report. Every factual assertion in the generated text must be traced to at least one source passage. Claims that cannot be mapped, or that stretch the source beyond what it supports, are flagged with an alignment distance score. This prompt is not a general-purpose fact-checker; it assumes the source passages are authoritative and complete for the domain.

textCopy
You are an evidence-audit assistant. Your task is to verify that every factual claim in the [GENERATED_TEXT] is supported by the provided [SOURCE_PASSAGES]. You must produce a structured alignment report.

## INPUTS
- GENERATED_TEXT: [GENERATED_TEXT]
- SOURCE_PASSAGES: [SOURCE_PASSAGES]
- ALIGNMENT_THRESHOLD: [ALIGNMENT_THRESHOLD]

## INSTRUCTIONS
1. Extract every discrete factual claim from [GENERATED_TEXT]. A factual claim is a statement that can be verified against a source. Ignore opinions, stylistic language, and transitional phrases.
2. For each claim, search [SOURCE_PASSAGES] for the best supporting passage. A supporting passage must contain the same factual content, not merely related topics.
3. Assign an alignment score from 0 to 100 for each claim-source pair:
   - 100: Verbatim or near-verbatim match.
   - 75-99: Strong support with minor rephrasing.
   - 50-74: Partial support; the source implies but does not state the claim.
   - 25-49: Weak support; the source is topically related but does not confirm the claim.
   - 0-24: No support; the claim cannot be found in any source.
4. Flag any claim with a score below [ALIGNMENT_THRESHOLD] as MISALIGNED.
5. Flag any claim that combines information from multiple sources in a way that creates a new assertion not present in any single source as STRETCHED.
6. If a claim has no supporting passage, set the source to "NONE" and the score to 0.

## OUTPUT SCHEMA
Return a JSON object with the following structure:
{
  "claims": [
    {
      "claim_id": "string",
      "claim_text": "string",
      "best_source_id": "string | null",
      "best_source_text": "string | null",
      "alignment_score": number,
      "flag": "ALIGNED | MISALIGNED | STRETCHED",
      "explanation": "string"
    }
  ],
  "summary": {
    "total_claims": number,
    "aligned_claims": number,
    "misaligned_claims": number,
    "stretched_claims": number,
    "average_alignment_score": number
  }
}

## CONSTRAINTS
- Do not invent source passages. If a claim cannot be matched, set best_source_id to null and flag as MISALIGNED.
- Do not skip claims. Every factual assertion in [GENERATED_TEXT] must appear in the output.
- If [GENERATED_TEXT] contains no factual claims, return an empty claims array with a summary noting zero claims.
- Preserve the exact text of claims and source passages. Do not paraphrase in the output fields.

To adapt this template, replace the square-bracket placeholders with your application's inputs. [GENERATED_TEXT] should be the model output under audit. [SOURCE_PASSAGES] should be a JSON array of objects, each with an id and text field, representing the retrieval corpus or document set. [ALIGNMENT_THRESHOLD] is an integer between 0 and 100 that defines the cutoff for flagging a claim as misaligned; start with 50 for initial testing and adjust based on your tolerance for partial matches. The output schema is strict JSON, which means your application harness should validate the structure before accepting the result. If the model returns malformed JSON, use a schema repair prompt from the Output Repair and Validation Prompts pillar rather than retrying this prompt blindly.

The Source-to-Claim Alignment Verification Prompt is not a standalone chat instruction; it is a verification node inside a larger evidence-grounding pipeline. The harness must supply the generated text and the full set of retrieved source passages as structured inputs, then parse the model's alignment report into a machine-readable schema. This prompt is designed for post-generation audit, meaning it runs after an answer or document has been produced, not during initial generation. The harness should treat this as a synchronous verification step with a strict output contract: every factual claim must be mapped to a source passage, and any unmapped or low-alignment claim must be flagged for human review before the output reaches an end user or compliance record.

Wire the prompt into an application by first extracting discrete claims from the generated text using a lightweight claim-splitting preprocessor or by instructing the model to perform extraction and mapping in a single pass. The harness should validate the JSON output against a schema that requires claim_text, source_passage_id, alignment_score, and flag fields for each claim. Implement a retry wrapper: if the model returns malformed JSON, missing required fields, or an alignment score array that does not match the claim count, re-invoke the prompt once with the validation error message appended to the [CONSTRAINTS] block. After two failed retries, escalate the entire output for human review rather than looping indefinitely. For high-stakes compliance workflows, log every alignment report—including scores, flags, and the raw prompt—to an immutable audit store so that reviewers can trace why a claim was marked as stretched or unsupported.

Model choice matters here. Use a model with strong instruction-following and structured output capabilities, such as GPT-4o or Claude 3.5 Sonnet, and set the temperature to 0 or near-zero to maximize deterministic alignment scoring. If your application already uses a RAG pipeline, pass the same retrieved chunks that were used during generation as the [SOURCE_PASSAGES] input; do not re-retrieve or expand the context during verification, as that would invalidate the alignment comparison. The alignment distance scoring should be treated as a signal for review prioritization, not as an absolute truth. Set a threshold—such as alignment_score < 0.7—that triggers mandatory human review, and configure the harness to route flagged claims into a review queue with the source passage and generated text side by side. Avoid using this prompt on streaming or real-time outputs without buffering the full generated text first; partial claims produce unreliable alignment maps.