README.md

metadata

license: mit
language:
  - en
pretty_name: Clarus Clinical Narrative Integrity v0.1.1
tags:
  - clarus
  - clinical-trials
  - narrative-integrity
  - section-transition
  - evidence-reporting
  - clinical-audit
task_categories:
  - text-classification
size_categories:
  - 1K<n<10K

Clarus Clinical Narrative Integrity v0.1.1

What this dataset tests

• You track meaning as it moves across sections of a clinical paper
• You detect when claims shift strength or scope without justification
• You flag narrative drift introduced by section transitions

Scope

• One document
• Multiple internal sections
  • Abstract
  • Methods
  • Results
  • Discussion
  • Conclusion
• One publication lifecycle

How this differs from v0.1

• v0.1 tests internal consistency within a single narrative fragment
• v0.1.1 tests continuity across narrative sections
• v0.1.1 is stricter and higher resolution
• Same invariant
• Finer audit surface

Typical failure modes

• Abstract upgrades exploratory findings
• Discussion removes qualifiers stated in Results
• Conclusion reframes safety or efficacy language
• Claims migrate from descriptive to causal
• Negative findings fade across sections

What this dataset does not test

• Protocol to publication transitions
• Registry or SAP drift
• Phase-to-phase definitional changes

Task format

• Each row presents paired or sequential narrative excerpts
• The model judges whether meaning is preserved across sections
• Output is a binary choice

Label meaning

• A indicates narrative integrity failure across sections
• B indicates narrative integrity preserved

Use cases

• Automated audit of clinical manuscripts
• Journal and peer-review triage
• Detection of rhetorical inflation in reporting

Version note

• v0.1 remains valid for single-fragment integrity checks
• v0.1.1 extends coverage to section-to-section narrative transitions
• Neither version supersedes the other