metadata
language:
- en
license: mit
pretty_name: Clinical Interventional Justification v0.1
tags:
- evaluation
- clinical
- reasoning
- inference
- transparency
task_categories:
- text-generation
- text-classification
Clinical Interventional Justification v0.1
Goal
- Test whether recommendations follow from stated analysis
- Expose non sequitur actions and hidden premises
What it measures
- justification_gap
The action lacks an explicit or implicit causal link to the analysis - hidden_premise
The action relies on unstated assumptions or missing severity drivers
Dataset format
- Each row includes observations, analysis, and a recommended action
- Constraints specify what must be justified
Columns
- id
- observations
- analysis
- question
- model_response
- reference_constraints
- labels_expected
Scoring
- Deterministic checks for causal connectors and scope alignment
- Per row score reflects matched flags
- Overall score is mean across rows
Run
- python scorer.py --csv data/sample.csv