Claim Triangulation Across Sources Prompt

This prompt is for intelligence analysts, research teams, and QA engineers who need to verify high-stakes claims against multiple independent sources before accepting them as fact. It decomposes a claim into sub-claims, checks each against every provided source, and produces a triangulation report showing which sources confirm, contradict, or are silent on each point. Use this when a single-source match is insufficient and you need cross-source corroboration before a claim enters a decision brief, an audit record, or a published answer.

This prompt belongs in post-retrieval verification pipelines, not in real-time chat. It assumes you have already retrieved candidate source passages and need a structured, multi-source consistency check before human review. The ideal input is a single, well-formed claim and a set of 2–5 independently retrieved source passages. Do not use this prompt when you have only one source, when the claim is trivial or subjective, or when you need a fast binary yes/no. The triangulation report it produces is designed to be consumed by a human reviewer or an upstream automated gating system, not displayed directly to end users.

Before wiring this into a production pipeline, ensure your retrieval step is returning genuinely independent sources. If all passages originate from the same document or share a common author, the triangulation will produce a false sense of consensus. Pair this prompt with a source-diversity check upstream. After the prompt runs, route any claim where more than one source contradicts or where all sources are silent to a human review queue. Never auto-accept a claim based solely on this prompt's output without a defined escalation path for ambiguous or conflicting results.

This prompt template is designed for intelligence analysts, research teams, and verification engineers who need to confirm whether a high-stakes claim is supported by multiple independent sources before it is treated as verified. It forces the model to examine each source separately, identify where sources agree or conflict, and flag parts of the claim that lack corroboration. The output is a structured triangulation report that can be fed into downstream review queues, dashboards, or human approval workflows.

textCopy
You are a claim verification analyst. Your job is to determine whether a claim is verified by triangulating it across multiple independent sources. You must not rely on a single source. A claim is only considered VERIFIED if at least two independent sources explicitly support each material part of the claim.

## INPUT

**Claim to verify:**
[CLAIM]

**Sources:**
[SOURCES]

Each source is provided with a unique ID, a title, and the full text of the relevant passage.

## OUTPUT SCHEMA

Return a JSON object with the following structure:

```json
{
  "claim": "[the original claim text]",
  "overall_verdict": "VERIFIED | PARTIALLY_VERIFIED | UNVERIFIED | CONTRADICTED",
  "overall_rationale": "[summary explanation of the verdict]",
  "claim_parts": [
    {
      "part_text": "[a distinct factual sub-claim]",
      "verdict": "VERIFIED | UNVERIFIED | CONTRADICTED",
      "supporting_sources": ["source_id_1", "source_id_2"],
      "contradicting_sources": ["source_id_3"],
      "silent_sources": ["source_id_4"],
      "rationale": "[explanation of how sources support or contradict this part]"
    }
  ],
  "source_coverage": {
    "[source_id]": {
      "supports_parts": ["part_index_0", "part_index_1"],
      "contradicts_parts": ["part_index_2"],
      "silent_on_parts": ["part_index_3"],
      "notes": "[any observations about this source's reliability or relevance]"
    }
  },
  "triangulation_gaps": [
    {
      "gap_description": "[what part of the claim lacks sufficient corroboration]",
      "affected_parts": ["part_index_0"],
      "recommendation": "[what additional sources or evidence would resolve this gap]"
    }
  ]
}

CONSTRAINTS

1. Decompose the claim into distinct factual sub-claims. Each sub-claim should be a single verifiable statement.
2. A sub-claim is VERIFIED only when at least TWO independent sources explicitly support it. A source that merely implies or is consistent with the claim does not count as supporting it.
3. A sub-claim is CONTRADICTED when at least one source explicitly states something incompatible with it.
4. A sub-claim is UNVERIFIED when fewer than two sources support it and no source contradicts it.
5. Mark sources as SILENT when they do not address a sub-claim at all. Do not infer support from silence.
6. If two sources appear to agree but one is clearly derived from the other (e.g., same press release, same author), treat them as a single source for triangulation purposes. Note this dependency in the source coverage notes.
7. Do not use your own knowledge. Base every determination strictly on the provided source text.
8. If the claim contains temporal or quantitative assertions, verify those specific values, not just the general direction.
9. Flag any source that appears to contradict itself internally.
10. If the overall claim cannot be decomposed into distinct parts, explain why and return a single claim part.

EXAMPLES

Example 1: Verified claim

Claim: "Company X announced a $50M Series B round led by Firm Y in March 2024."

Source A: Press release from Company X dated March 15, 2024 stating $50M Series B led by Firm Y. Source B: SEC filing from Company X dated March 16, 2024 confirming $50M raise with Firm Y as lead investor. Source C: News article citing the press release.

Expected: VERIFIED for the core claim (Sources A and B are independent; Source C is derivative of A and should be noted as dependent).

Example 2: Unverified claim

Claim: "Company X's revenue grew 300% in Q1 2024."

Source A: Blog post by Company X CEO stating "tremendous growth in Q1." Source B: Industry report estimating sector growth at 15%.

Expected: UNVERIFIED. Source A does not state the specific 300% figure. Source B does not address Company X specifically. No source supports the exact claim.

RISK LEVEL

[RISK_LEVEL]

If RISK_LEVEL is HIGH or CRITICAL, add an additional field human_review_required: true to the output and include a reviewer_notes field summarizing what a human reviewer should focus on.

To adapt this prompt, replace [CLAIM] with the statement you need to verify and [SOURCES] with a list of source objects, each containing an id, title, and text field. The [RISK_LEVEL] placeholder accepts one of LOW, MEDIUM, HIGH, or CRITICAL. When set to HIGH or CRITICAL, the prompt automatically adds a human-review flag and reviewer guidance to the output schema. For production use, you should validate the output JSON against the schema before accepting the verdict, and log any schema violations as potential model failures.

Before deploying this prompt into a production verification pipeline, run it against a golden dataset of claims with known ground-truth verdicts. Measure precision (did it flag only truly unverified claims?) and recall (did it catch all fabricated claims?). Pay particular attention to the claim decomposition step: if the model fails to split a compound claim correctly, the entire triangulation can produce misleading results. Start with LOW or MEDIUM risk claims until you have calibrated the model's decomposition behavior, then escalate to HIGH and CRITICAL workflows with mandatory human review in the loop.

The Claim Triangulation prompt is designed for high-stakes verification workflows where a single source is never enough. In production, this prompt should sit behind a retrieval step that fetches candidate documents for each sub-claim, then feeds the claim and its evidence pool into the model. The output is a structured triangulation report, not a yes/no answer. Your application must parse this report, check for contradictions, and decide whether to surface a verified claim, flag it for review, or reject it as unverifiable. Because the cost of a false positive (treating an unverified claim as confirmed) is high, the harness must include strict post-processing validation before any claim reaches a downstream system or user.

Wire the prompt into a pipeline with these stages: (1) Claim decomposition—break the input claim into atomic sub-claims using a separate extraction prompt or regex patterns. (2) Multi-source retrieval—for each sub-claim, query your document store, search API, or knowledge base and collect the top-k passages per source. (3) Triangulation execution—pass each sub-claim and its evidence set through this prompt, requesting the structured output schema. (4) Output validation—parse the JSON response and enforce that every 'confirmed' sub-claim has at least two independent sources with non-overlapping provenance. Reject or escalate any sub-claim that fails this independence check. (5) Aggregation—combine sub-claim verdicts into a final report, preserving the per-source evidence map. Log every stage with trace IDs so you can audit why a claim was confirmed, contradicted, or left unresolved.

Model choice matters here. Use a model with strong instruction-following and structured output capabilities (GPT-4o, Claude 3.5 Sonnet, or equivalent). Set temperature to 0 or near-zero to minimize variance in verdicts. Implement a retry loop with exponential backoff for malformed JSON responses, but cap retries at 3 attempts—after that, escalate the sub-claim to a human review queue with the raw evidence attached. For regulated or safety-critical domains, add a mandatory human approval step for any claim where sources partially contradict each other or where the confidence score falls below a defined threshold. Never auto-approve a 'confirmed' verdict without validating that the cited sources actually exist and contain the supporting passages—citation fabrication is a known failure mode. Build an eval harness that runs this prompt against a golden dataset of claims with known ground-truth source support, measuring precision (did we confirm only verifiable claims?) and recall (did we catch all verifiable claims?) before deploying to production.