Research Verification

Overview

Before reporting a finding, ask an independent verifier to classify each claim against the actual evidence as supported, contradicted, unsupported, or overstated. This is a lightweight claim gate, not a replacement for multi-source reconciliation on high-stakes research.

Procedure

1. Preserve the exact claim, its producer, and the primary evidence available.
2. Select a verifier different from the producer and suitable for the subject.
3. Supply the task, rubric and primary evidence, especially for numbers, dates, quotations and precise technical assertions; retain an initial evidence-based judgment before showing the producer's evaluative summary, then assess that summary separately as attributed information rather than ground truth.
4. Require a verdict and reasoning for each claim. Hold unsupported claims; correct or flag contradicted and overstated claims.
5. Escalate contested or reliability-critical research to independent source reconciliation.

Rules

• Independence is essential: a model's self-review is not the same as independent verification.
• Plausibility-only checks are weak for precise facts; evidence context is the preferred mode.
• Choose a verifier with appropriate safety and subject-matter limitations for sensitive topics.
• Keep the initial evidence judgment separate from summary-assisted review; the research-team protocol cites both the adversarial-report harm and the honest-report improvement, so this rule is not a claim that all summaries degrade accuracy.

Pitfalls

• Calling an unsupported number verified because it sounds plausible.
• Using the producing worker as the only verifier.
• Treating a local claim check as a complete evidence review.

Verification checklist

• Producer and verifier differ.
• Primary evidence was provided where available.
• Unsupported and overstated claims were withheld or clearly labeled.