Datasets:

ClarusC64
/

clinical-constraint-absorbability-ordering-v0.3

scenario_id stringlengths 5 5	ferritin int64 11 72	tsh float64 2 9	sleep_disruption int64 3 9	stress_load int64 4 10	repair_demand int64 3 9	bandwidth_proxy float64 0.22 0.85	constraint_coupling float64 0.11 0.94	wrong_sequence_risk stringclasses 3 values	case_type stringclasses 5 values	blocking_constraint stringclasses 5 values	gold_intervention stringclasses 5 values
TR001	62	8.3	4	5	5	0.61	0.22	medium	clean	thyroid	thyroid_first
TR002	13	2.4	4	5	6	0.66	0.24	medium	clean	iron	iron_first
TR003	65	2.1	8	6	7	0.38	0.31	medium	clean	sleep	sleep_first
TR004	60	2.2	8	9	8	0.28	0.86	high	clean	bandwidth	load_reduction_first
TR005	70	2.3	5	4	3	0.82	0.12	low	clean	none	observe_first
TR006	11	8.7	8	8	9	0.24	0.94	high	triplet_train	bandwidth	load_reduction_first
TR007	11	8.7	8	8	9	0.56	0.44	high	triplet_train	iron	iron_first
TR008	11	8.7	8	8	9	0.67	0.35	medium	triplet_train	sleep	sleep_first
TR009	14	8.2	5	6	8	0.5	0.84	high	moderate	iron	iron_first
TR010	64	8	7	6	8	0.4	0.78	high	moderate	sleep	sleep_first
TR011	58	8.2	5	8	8	0.35	0.74	high	moderate	bandwidth	load_reduction_first
TR012	66	8.5	4	5	6	0.64	0.2	medium	clean	thyroid	thyroid_first
TR013	14	8.9	4	6	8	0.58	0.73	high	moderate	iron	iron_first
TR014	15	8.4	8	5	9	0.63	0.48	high	moderate	sleep	sleep_first
TR015	17	8.1	7	10	9	0.22	0.91	high	moderate	bandwidth	load_reduction_first
TR016	68	2.4	5	4	3	0.83	0.13	low	clean	none	observe_first
TR017	63	8.4	4	5	5	0.62	0.23	medium	clean	thyroid	thyroid_first
TR018	15	2.3	4	5	6	0.67	0.26	medium	clean	iron	iron_first
TR019	66	2	8	5	7	0.39	0.32	medium	clean	sleep	sleep_first
TR020	62	2.1	8	9	8	0.28	0.87	high	clean	bandwidth	load_reduction_first
TR021	72	2.3	5	4	3	0.85	0.11	low	clean	none	observe_first
TR022	18	8	6	7	8	0.48	0.78	high	moderate	iron	iron_first
TR023	61	8.1	7	7	8	0.4	0.8	high	moderate	sleep	sleep_first
TR024	62	8.2	5	9	8	0.3	0.76	high	moderate	bandwidth	load_reduction_first
TR025	12	8.5	8	9	9	0.26	0.93	high	triplet_train	bandwidth	load_reduction_first
TR026	12	8.5	8	9	9	0.55	0.46	high	triplet_train	iron	iron_first
TR027	12	8.5	8	9	9	0.69	0.36	medium	triplet_train	sleep	sleep_first
TR028	21	7.3	7	6	7	0.55	0.49	high	recovery_trajectory	iron	iron_first
TR029	24	4.1	9	8	7	0.49	0.58	high	masked_conflict	sleep	sleep_first
TR030	64	9	3	4	5	0.59	0.42	medium	clean	thyroid	thyroid_first

Clinical Constraint Absorbability Ordering v0.3

This dataset tests whether a model can infer the first absorbable repair move under hidden clinical constraint competition.

The task is not diagnosis.

The task is not ordinary treatment selection.

The task is recovery-control sequencing:

What can this system safely absorb first?

Core idea

A patient may show multiple abnormal signals at once.

Examples:

elevated TSH
low ferritin
sleep disruption
high stress load
high repair demand

A surface model may choose the most obvious abnormal marker.

This benchmark asks whether a model can infer the binding constraint.

Prediction target

Prediction files must contain:

scenario_id,prediction,predicted_constraint,predicted_risk
TE001,thyroid_first,thyroid,medium

Allowed prediction labels:

thyroid_first
iron_first
sleep_first
load_reduction_first
observe_first

Allowed predicted_constraint labels:

thyroid
iron
sleep
bandwidth
none

Allowed predicted_risk labels:

low
medium
high
Row structure

Each row contains:

ferritin
tsh
sleep_disruption
stress_load
repair_demand
bandwidth_proxy
constraint_coupling
wrong_sequence_risk
case_type
blocking_constraint
gold_intervention

The key target is:

gold_intervention

The structural explanation target is:

blocking_constraint
Case types

The dataset includes:

clean
moderate
triplet_train
triplet_test
edge_ambiguous
hard_edge
recovery_trajectory
masked_conflict

The test set is intentionally harder than the training set.

Why this is difficult

The hardest cases include near-identical pathology profiles.

In triplet_test, the pathology profile is held constant while bandwidth and constraint coupling change.

This means:

same pathology
different absorbability state
different first move

A model that memorizes:

high TSH -> thyroid_first
low ferritin -> iron_first
sleep disruption -> sleep_first

will fail the hard cases.

Evaluation

Run:

python scorer.py predictions.csv data/test.csv

The scorer reports:

move accuracy
constraint accuracy
risk accuracy
clean accuracy
triplet test accuracy
edge ambiguous accuracy
hard edge accuracy
recovery trajectory accuracy
masked conflict accuracy
hard case accuracy
hard constraint accuracy
sequence safety score
decoy resistance score
macro precision
macro recall
macro F1
structural score

The main metric is:

structural_score
Structural scoring

The structural score rewards:

correct intervention
correct blocking constraint
correct risk classification
hard-case performance
hard-constraint performance
safety under high-risk sequencing
resistance to marker-matching shortcuts
triplet generalization
edge-case performance
Structural Note

This dataset is synthetic.

It is designed to test hidden recovery-control reasoning and intervention sequencing under constrained adaptive bandwidth.

It is not medical advice and should not be used for clinical decision-making.

License

MIT

Downloads last month: 39

Number of rows:

Total file size:

16.7 kB