evidence-claim-mismatch

Review and repair mismatches between claims and evidence snippets in an existing dismech disease entry. Use for evidence-claim-mismatch issues, evaluation-queue findings, or a requested claim/snippet

SKILL.md

GitHub

Review evidence–claim mismatches

Find claims whose selected snippets do not support what the YAML asserts. Expand the literature search to include more papers where necessary. Evaluation results are leads for review, not instructions to agree with a model.

Read the claim in context

  • Read the current disease YAML and the issue discussion, including linked PRs. Issue examples are saved evaluation inputs and may lag behind edits. Skip resolved findings; do not recreate old problems to match a snapshot.
  • Use the linked saved assessments to go beyond the issue's few examples. Assess the disease/subtype and inherited context as well as the assertion's terms, descriptions, qualifiers, support/refutation, and directness. / means the whole claim; other paths identify individual aspects of that claim.
  • A snippet may support the term but not a bundled description or qualifier. Distinguish overclaiming from a claim that is simply less detailed than its source. A broader, correctly supported claim does not need to repeat every detail of the excerpt.
  • Nested claims with their own evidence are separate assertions. In particular, evidence on a pathophysiology node need not justify its downstream edges. Check those edges' evidence separately. Consider complementary evidence on the assertion before deciding that the KB claim itself is unsupported.
  • Treatment/medical-action claims may describe hypotheses about an action's effect on a mechanism; do not automatically read them as clinical recommendations.

Repair the evidence or claim

Read the cached reference to understand the passage, while asking whether the selected snippet conveys the needed support. The task is claim–snippet fit, not rating the source. Extend or replace a snippet to make its context clear; ... and [editorial context] are available when faithful to the source. Do not add an unsupported qualifier inside editorial brackets. Follow dismech-references for quoting, cache generation, and validation.

Retain correct claims and limitations. Revise an assertion only when the evidence justifies doing so; seek additional literature if the existing sources do not cover it. PARTIAL can reflect weak, incomplete, or ambiguous support and is not by itself an error. Record the reason when a flagged claim should remain.

Follow CLAUDE.md for history records and validation. Make targeted YAML edits without unrelated serialization churn. Summarize the substantive corrections and retained findings in the curation history. Link the PR with Closes #<issue> so merging completes the queued task. If review finds no changes necessary, explain the findings on the issue and close it without a content PR.