ClarusC64/clinical-constraint-intervention-planning-v0.1

Clinical Constraint Intervention Planning v0.1

Clinical Constraint Intervention Planning (CCIP) is a synthetic benchmark for evaluating whether an AI system can construct safe, temporally valid intervention plans under interacting clinical constraints.

The benchmark tests more than selection of a plausible intervention. A model must preserve active constraints, satisfy intervention preconditions at the moment of execution, distinguish treatment initiation from confirmed establishment, respect workflow dependencies, keep mutually exclusive branches separate, and project the resulting state of the complete case.

Repository path:

ClarusC64/clinical-constraint-intervention-planning-v0.1

Benchmark summary

Property

Value

Benchmark ID

CCIP

Version

0.1

Data format

JSONL

Canonical split

train

Records

32

Matched pairs

16

Scenario range

CCIP-401-A to CCIP-416-B

Language

English

Data origin

Fully synthetic

Primary purpose

Clinical constraint-aware intervention planning evaluation

Each A/B pair preserves the same broad clinical problem while changing a decisive governing fact. Typical contrasts include an active constraint versus a resolved constraint, a blocked pathway versus an eligible pathway, or an unresolved decision point versus a direct treatment route.

Repository structure

clinical-constraint-intervention-planning-v0.1/ ├── data/ │ └── train.jsonl ├── scorer.py ├── README.md └── LICENSE

The canonical dataset is data/train.jsonl. A CSV version is not required because the records contain deeply nested lists, mappings, transition paths, branches, and dependency structures that are naturally represented in JSONL.

What CCIP evaluates

CCIP evaluates whether a model can:

identify interventions blocked by active safety or workflow constraints;

order interventions so that every precondition is satisfied when the action occurs;

distinguish planning, initiation, confirmation, resumption, de-escalation, completion, and procedural performance;

resolve constraints only through explicit corrective or reassessment actions;

maintain monitoring and specialist-review requirements;

separate mutually exclusive conditional branches;

preserve items that should remain unchanged;

establish items that must reach a new required state;

project the final state of all tracked items; and

avoid unsafe plans even when they appear clinically plausible in isolation.

Data format

Each physical line in data/train.jsonl is one complete JSON object.

from datasets import load_dataset

dataset = load_dataset( "ClarusC64/clinical-constraint-intervention-planning-v0.1" )

print(dataset["train"][0])

Local loading:

from datasets import load_dataset

dataset = load_dataset( "json", data_files={"train": "data/train.jsonl"}, )

Standard-library loading:

import json from pathlib import Path

records = [] with Path("data/train.jsonl").open("r", encoding="utf-8") as handle: for line in handle: if line.strip(): records.append(json.loads(line))

print(len(records))

Record schema

Every record contains the following top-level fields:

Field

Description

scenario_id

Unique scenario identifier.

pair_id

Identifier shared by the matched A/B pair.

domain

Clinical specialty or workflow domain.

prior_context

Synthetic case narrative defining the initial planning problem.

tracked_items

Diagnoses, constraints, treatments, procedures, monitoring processes, and referrals tracked through the scenario.

current_state

Initial state of every tracked item.

available_interventions

Candidate interventions and their targets.

active_constraints

Active constraints and the interventions they block.

intervention_preconditions

States required immediately before each intervention.

intervention_contraindications

States that prohibit each intervention.

cross_item_dependencies

Cross-item relationships enabling later interventions.

workflow_constraints

Ordering and mutual-exclusion requirements.

gold_intervention_plan

Ordered primary plan with action labels and rationales.

gold_intervention_order

Canonical sequence of primary intervention IDs.

conditional_branches

Mutually exclusive intervention sequences activated by confirmed branch conditions.

gold_blocked_interventions

Interventions that remain blocked and the conditions required to unblock them.

gold_constraint_resolution_path

Explicit state transitions for constraints on the primary path.

gold_expected_transition_path

State of each tracked item after every primary intervention.

gold_projected_final_state

Projected state at the end of the primary plan.

gold_preserved_items

Items intentionally left unchanged by the primary plan.

gold_established_items

Items deliberately brought to a required new state.

gold_monitoring_plan

Monitoring activation, targets, dependencies, and frequency.

gold_stop_conditions

Conditions requiring an intervention to stop, pause, or change.

gold_fallback_paths

Contingency plans for failure or reactivation.

valid_alternative_plans

Other sequences accepted under stated conditions.

unsafe_plans

Explicitly unsafe sequences and their failure mechanisms.

failure_signatures

Named planning failures represented by the record.

integration_types

Structural reasoning capabilities integrated by the case.

reasoning_patterns

Intended reasoning operations.

mechanism_complexity

Aggregate complexity label.

complexity_components

Counts supporting the complexity label.

annotation_confidence

Confidence assigned to the synthetic annotation.

confidence_rationale

Reason for the assigned confidence.

State ontology

The benchmark uses a controlled state vocabulary:

active_confirmed active_suspected resolved ruled_out withheld inactive pending planned initiated paused resumed switched deescalated requested escalated scheduled downgraded discontinued technically_limited inconclusive completed historical performed

The distinction between these states is normative. For example:

planned is not equivalent to initiated;

initiated is not equivalent to active_confirmed;

a procedure becomes performed, not active_confirmed;

a referral remains pending until a referral intervention changes it;

monitoring or planning alone cannot resolve a diagnosis or constraint.

Core benchmark contracts

A valid CCIP plan must satisfy the following contracts:

Every active constraint refers to a tracked item.

Every referenced intervention and item exists in the record.

gold_intervention_plan and gold_intervention_order agree exactly.

Preconditions are checked against the state immediately before execution.

No intervention may execute while a contraindicating state is active.

Every state transition is causally attributable to a named intervention.

Constraint resolution requires an explicit corrective or reassessment action.

Conditional interventions remain outside the primary sequence until their branch condition is confirmed.

Mutually exclusive branches must not be linearised into one plan.

Referral states change only through referral interventions.

gold_preserved_items contains only unchanged items.

gold_established_items contains only items deliberately moved to a required new state.

An item cannot be both preserved and established.

Treatment establishment requires a distinct confirmation step after initiation or resumption.

Matched-pair design

Each pair contains an A and B scenario sharing a common planning frame.

The A case generally contains an active or unresolved governing constraint. The B case changes one decisive fact, such as confirmed constraint resolution, while retaining a substantially comparable tracked-item structure and intervention space.

This design tests whether a model changes its plan because the state geometry changed, rather than because superficial vocabulary changed.

Scoring

The included scorer.py uses only the Python standard library.

Validate the gold dataset:

python scorer.py --gold data/train.jsonl --validate-only

Score a prediction file:

python scorer.py
--gold data/train.jsonl
--predictions predictions.jsonl
--output results.json

Scored components

Component

Default weight

Intervention order

0.32

Projected final state

0.24

Blocked interventions

0.10

Constraint-resolution path

0.10

Conditional branches

0.10

Preserved items

0.05

Established items

0.05

Safety

0.04

Total

1.00

The scorer reports scenario-level, pair-level, domain-level, component-level, and aggregate results. Unsafe intervention sequences may trigger a safety penalty or score ceiling even when other components are partially correct.

Prediction format

Only scenario_id and an intervention-order field are required. Missing optional fields receive zero for the corresponding component rather than causing a parsing failure.

Recommended compact format:

{ "scenario_id": "CCIP-401-A", "predicted_intervention_order": ["A2", "A3", "A4", "A10", "A5", "A6", "A7", "A8", "A9"], "predicted_final_state": { "P1": "downgraded", "P2": "resolved", "C1": "resolved", "T1": "deescalated", "T2": "active_confirmed", "T3": "completed", "M1": "active_confirmed", "R1": "escalated" }, "predicted_blocked_interventions": ["A1", "A5", "A6", "A7", "A8", "A9"], "predicted_constraint_resolution_path": [ { "constraint_id": "C1", "from_state": "active_confirmed", "to_state": "downgraded", "intervention_id": "A5" }, { "constraint_id": "C1", "from_state": "downgraded", "to_state": "resolved", "intervention_id": "A6" } ], "predicted_conditional_branches": [], "predicted_preserved_items": [], "predicted_established_items": ["M1", "T2"] }

The scorer also accepts the corresponding gold-style field names, including gold_intervention_order, gold_projected_final_state, and gold_blocked_interventions.

Baseline evaluation patterns

Useful baselines include:

Order-only baseline — predicts the primary intervention sequence.

State-projection baseline — predicts intervention order and final state.

Full-structure baseline — predicts all scored components.

Pair-consistency evaluation — compares whether a model changes its plan appropriately across each A/B pair.

Safety-ablation evaluation — measures performance with and without explicit constraint and contraindication fields.

Because the benchmark is small, results should be interpreted as structured diagnostic evidence rather than a comprehensive estimate of clinical competence.

Intended uses

CCIP is intended for:

evaluation of clinical planning and temporal reasoning;

testing constraint-sensitive model behaviour;

analysis of unsafe intervention sequencing;

matched-pair evaluation of state-dependent planning;

development of structured reasoning and planning systems;

scorer and validator research; and

controlled studies of reasoning drift, branch collapse, and premature action.

Out-of-scope uses

CCIP must not be used as:

a clinical decision-support system;

a substitute for professional medical judgement;

a source of treatment recommendations for real patients;

a diagnostic model;

a medical-device validation dataset;

evidence that a model is clinically safe; or

a training source for unsupervised deployment in patient-facing settings.

Limitations

All cases are synthetic and intentionally simplified.

The dataset contains 32 records and is designed for structural depth rather than statistical scale.

Clinical pathways are abstractions created to test planning contracts, not complete care guidelines.

The state ontology is benchmark-specific.

Some scenarios terminate at an unresolved decision point and represent later actions as conditional branches.

A high score demonstrates agreement with the benchmark annotation, not general clinical competence.

The benchmark does not estimate real-world prevalence, treatment effectiveness, or patient outcomes.

Safety and ethics

No real patient records, protected health information, or identifiable clinical narratives are included.

The benchmark is designed to penalise plans that violate explicit constraints, omit required confirmation, collapse mutually exclusive branches, or act before safety conditions are satisfied. This does not make the benchmark or a model evaluated on it suitable for real clinical use.

Data statement

The data were generated as synthetic structured cases. The records encode explicit gold planning trajectories, blocked actions, state transitions, monitoring requirements, fallback paths, and unsafe-plan signatures.

No demographic inference, patient identification, or real-world medical record reconstruction is intended or supported.

Versioning

v0.1

32 synthetic scenarios;

16 matched A/B pairs;

constraint-aware intervention planning schema;

primary and conditional planning paths;

explicit state-transition annotations;

benchmark validator and weighted scorer.

Future versions may add held-out evaluation splits, adversarial perturbations, additional domains, calibration tasks, and harder multi-constraint cases.

Licence

See the repository LICENSE file. The dataset card uses license: other until a specific Hugging Face licence identifier is declared.

Citation

A formal citation has not yet been assigned. Until one is available, cite the repository as:

ClarusC64. Clinical Constraint Intervention Planning v0.1. Hugging Face dataset repository, 2026. ClarusC64/clinical-constraint-intervention-planning-v0.1

Suggested BibTeX:

@dataset{clarusc64_ccip_2026, author = {ClarusC64}, title = {Clinical Constraint Intervention Planning v0.1}, year = {2026}, publisher = {Hugging Face}, note = {Synthetic benchmark for constraint-aware clinical intervention planning} }