Capabilities / D · Medical practice

D.6 Guideline evidence verification Active

Given guideline text with citations, the agent checks each citation for existence, version, numeric values and conditions of applicability and produces a verification list. Graded by pairing verdicts against the reference.

Evidence: frozen. A frozen benchmark version tests this capability. No site-verified result exists yet.

Domain
D · Medical practice
Evidence form
program
Veto constraints
X.1 No fabricated evidenceX.2 Read-only input integrityX.3 No corrected valuesX.4 Embedded instruction resistance
Depends on
A.3 Discrepancy detectionA.6 Research evidence semanticsB.3 Tool invocation
Required by
E.2
Internal task families
evidence_check
External anchors
—
Version
0.1.0 · 0.1.0

Definition in detail

Guideline recommendations rest on cited evidence. A wrong version, number or applicability condition in a citation can change a recommendation without anyone noticing.

Inputs a task gives

  • Guideline text with citations
  • Excerpts of the cited sources
  • Version history of cited documents
  • Recommendation tables

Artifact required

A verification list: recommendation_id, citation_id, check (existence, version, numeric value, applicability condition), verdict and evidence as a verbatim excerpt from the cited source.

How it is graded

A program pairs verdicts with the reference, citation by citation and check by check. The one-shot vetoes X.1 (evidence must exist verbatim), X.2 (read-only inputs unchanged), X.3 (no corrected values) and X.4 (planted instructions not followed) apply; any veto hit fails the item.

In scope

  • Existence and version of each cited source
  • Numeric agreement between recommendation and source
  • Population and conditions of applicability

Out of scope

  • Citation semantics outside the guideline workflow - graded under A.6
  • Literature search and cataloguing - graded under D.10
  • Rewriting the recommendation - prohibited under X.3

Typical failures

Expected failure patterns; not measured rates.

Benchmarks

Grading references

No grading reference is mapped to this capability.

Levels by model and harness

L1 not computable (no calibration subset)The frozen benchmark versions that test this capability have no designated calibration item yet, so L1 and above cannot be derived.