ClarusC64/clinical-cross-team-trajectory-reconstruction-v0.1
What this dataset does Tests whether a model can reconstruct a patient's hidden trajectory after a cross-team handoff. The challenge is not diagnosis. The challenge is determining whether enough information survives the transfer to safely continue care. Core geometry A patient trajectory may fail because: critical steps are missing sequence order is corrupted constraint state is lost documentation lags reality apparent stability hides deterioration The model… See the full description on the dataset page: https://huggingface.co/datasets/ClarusC64/clinical-cross-team-trajectory-reconstruction-v0.1.
What this dataset does
Tests whether a model can reconstruct a patient's hidden trajectory after a cross-team handoff.
The challenge is not diagnosis.
The challenge is determining whether enough information survives the transfer to safely continue care.
Core geometry
A patient trajectory may fail because:
- critical steps are missing
- sequence order is corrupted
- constraint state is lost
- documentation lags reality
- apparent stability hides deterioration
The model must decide whether the trajectory can be reconstructed.
Prediction target
Predict:
target_label
Allowed labels:
trajectory_reconstructed
constraint_state_lost
sequence_order_uncertain
false_stability_detected
requires_clarification_before_action
unsafe_to_continue_from_record
Row structure
Variables include:
handoff summary fidelity
missing step count
sequence contradiction score
constraint trace integrity
vital trend continuity
documentation lag
apparent stability score
hidden deterioration signal
Evaluation
Primary metric:
structural_score
The scorer evaluates:
overall reconstruction accuracy
false stability detection
reconstruction safety
unsafe continuation detection
hidden constraint preservation
Structural Note
This benchmark tests post-handoff trajectory reconstruction rather than diagnosis or treatment selection.
The central question is:
Can the receiving team reconstruct the hidden state of the patient from incomplete records?
License
MIT