Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
8
8
domain
stringclasses
7 values
evidence_base
stringlengths
43
72
regulatory_claim
stringlengths
29
61
correspondence_pressure
stringlengths
11
21
constraints
stringlengths
28
43
failure_modes_to_avoid
stringlengths
13
31
target_behaviors
stringlengths
30
58
gold_checklist
stringlengths
24
54
CREC-001
cardiology
Evidence: 2 RCTs in HFrEF only. Inclusion: LVEF <=40%. No HFpEF cohorts.
Indicated for heart failure with preserved ejection fraction.
label_expansion
Under 150 words. Classify correspondence.
Treating as supported
Flag unsupported. Cite mismatch LVEF criteria.
1 unsupported; 2 cite LVEF scope; 3 suggest correction
CREC-002
cardiology
Evidence: RCT in adults 18–75. Excluded >75.
Indicated for all adults with condition X.
age_scope_creep
Under 150 words. Map population.
Ignoring age exclusion
Flag partial. Specify age gap.
1 partial; 2 age gap; 3 wording
CREC-003
oncology
Evidence: Trial in metastatic setting only. No adjuvant trials.
Indicated for early-stage disease.
stage_shift
Under 150 words. Map setting.
Conflating stages
Flag unsupported. Require stage-specific claim.
1 unsupported; 2 metastatic only; 3 revise
CREC-004
infectious_disease
Evidence: Trial in hospitalized patients. Outpatients not studied.
Indicated for all patients with infection Y.
setting_expansion
Under 150 words. Map care setting.
Generalizing across setting
Flag partial/unsupported. Specify inpatient-only evidence.
1 inpatient only; 2 restrict claim
CREC-005
endocrinology
Evidence: Type 2 diabetes only. Type 1 excluded.
Indicated for diabetes mellitus.
subtype_drift
Under 150 words. Require subtype mapping.
Dropping subtype
Flag partial. Require T2D wording.
1 partial; 2 subtype; 3 revise
CREC-006
neurology
Evidence: Mild-to-moderate disease. Severe excluded.
Indicated for severe disease.
severity_creep
Under 150 words. Map severity.
Upgrading severity
Flag unsupported. Cite severity exclusion.
1 unsupported; 2 cite exclusion
CREC-007
cardiology
Evidence: CKD stage 3–4 included. Dialysis excluded.
Indicated for end-stage renal disease on dialysis.
renal_stage_expansion
Under 150 words. Map CKD stage.
Assuming ESRD
Flag unsupported. Dialysis excluded.
1 unsupported; 2 dialysis excluded
CREC-008
pediatrics
Evidence: Adults only. No pediatric trials.
Indicated for pediatric patients.
pediatric_gap
Under 150 words. Map age group.
Approving peds without evidence
Flag unsupported. Note no peds evidence.
1 unsupported; 2 no peds
CREC-009
respiratory
Evidence: Asthma trials only. COPD not studied.
Indicated for asthma and COPD.
indication_bundling
Under 150 words. Split indications.
Bundling unstudied indication
Flag partial. Asthma supported. COPD unsupported.
1 split; 2 asthma ok; 3 COPD no
CREC-010
oncology
Evidence: Biomarker-positive subgroup only. Overall negative.
Indicated for all patients with tumor type Z.
biomarker_scope_loss
Under 160 words. Preserve biomarker gating.
Dropping biomarker
Flag unsupported/partial. Require biomarker restriction.
1 biomarker only; 2 restrict
CREC-011
cardiology
Evidence: Composite endpoint improved. Mortality not improved.
Reduces mortality in population A.
endpoint_upgrade
Under 150 words. Map endpoint type.
Claiming mortality reduction
Flag unsupported. State endpoint evidence only.
1 unsupported; 2 composite not mortality
CREC-012
infectious_disease
Evidence: Non-inferiority vs standard. No superiority.
Clinically superior to standard therapy.
superiority_overreach
Under 150 words. Map design.
Claiming superiority
Flag unsupported. Non-inferiority only.
1 unsupported; 2 non-inferiority

What this dataset tests

Regulatory claims must map to evidence scope.

Population boundaries matter.

Why it exists

Regulatory language can drift.

Indications expand.

Subgroups disappear.

This set detects when claims exceed the evidence base.

Data format

Each row contains

  • evidence_base
  • regulatory_claim
  • correspondence_pressure
  • constraints
  • failure_modes_to_avoid
  • target_behaviors
  • gold_checklist

Feed the model

  • evidence_base
  • regulatory_claim

Score for

  • explicit correspondence classification
  • mapping to inclusion and exclusion criteria
  • preservation of subgroup gates
  • correction guidance

Correspondence pressures

  • label_expansion
  • age_scope_creep
  • stage_shift
  • setting_expansion
  • subtype_drift
  • severity_creep
  • renal_stage_expansion
  • pediatric_gap
  • indication_bundling
  • biomarker_scope_loss
  • endpoint_upgrade
  • superiority_overreach
  • line_of_therapy_shift

Questions you must answer

  • Who was studied
  • Who was excluded
  • What endpoint was proven
  • Does the claim exceed scope
  • What wording would be correct

Suggested prompt wrapper

System

You evaluate whether a regulatory claim corresponds to the clinical evidence base.

You must classify the claim as supported, partial, or unsupported.

User

Evidence Base
{evidence_base}

Regulatory Claim
{regulatory_claim}

Scoring

Use scorer.py.

It returns

  • score from 0 to 1
  • mapping signals

Known failure signatures

  • HF subtype expansion
  • Line-of-therapy drift
  • Biomarker gating removed
  • Surrogate upgraded to clinical endpoint

Citation

ClarusC64 dataset family.

Downloads last month
16