ClarusC64/clinical-quad-infection-buffer-lag-coupling-sepsis-transition-v1.0

ClarusC64/clinical-quad-infection-buffer-lag-coupling-sepsis-transition-v1.0

What this repo does

This repository provides a Clarus v1.0 benchmark for sepsis transition under a four-variable clinical quad:

• —infection_load
• —buffer_capacity
• —lag_burden
• —coupling_stress

The v1.0 upgrade is Closed-Loop Control Geometry.

The task is no longer limited to detecting deterioration or ranking one intervention against another.

It tests whether a controller can:

• —choose the right path
• —apply the path in the right sequence
• —read feedback from the system
• —adapt in time
• —maintain durable recovery

Concept ladder

Core quad

infection_load

Total infectious pressure pushing the system toward sepsis transition.

buffer_capacity

Remaining reserve available to absorb infectious and hemodynamic strain.

lag_burden

Delayed correction pressure caused by unresolved infection or slow response.

coupling_stress

Cross-system destabilization linking infection, perfusion, inflammation, and organ support.

Clinical variable mapping

These four variables define the structural state of the septic system rather than a single physiological measurement.

Prediction target

The target is label_sepsis_transition.

Default stronger v1.0 rule

label = 1 if all of the following hold:

• —stabilization_success = 1
• —trajectory_shift < -0.10
• —intervention_alignment_score >= 0.60
• —control_sequence_alignment_score >= 0.60
• —recovery_consistency_score >= 0.60

Mid-strength variant

label = 1 if:

• —stabilization_success = 1
• —trajectory_shift < -0.10
• —control_sequence_alignment_score >= 0.60

Relaxed variant

label = 1 if stabilization_success = 1

Example row

The following simplified row shows a near-miss under the strict v1.0 label rule.

Four of the five conditions are satisfied.

But control_sequence_alignment_score = 0.59 falls below the threshold of 0.60.

Therefore:

label = 0

This illustrates why v1.0 measures true closed-loop control quality, not temporary improvement alone.

Row structure

Each row represents a sepsis deterioration scenario with:

• —quad state variables
• —trajectory signals
• —boundary geometry
• —regime transition signals
• —intervention competition signals
• —closed-loop control signals
• —perturbation and recovery signals
• —delta signals
• —intervention path and final label

Signal groups

Quad state variables

• —infection_load
• —buffer_capacity
• —lag_burden
• —coupling_stress

Trajectory signals

• —drift_gradient
• —drift_velocity
• —drift_acceleration
• —trajectory_shift

Boundary geometry

• —boundary_distance
• —secondaryboundarydistance
• —boundarycompetitionratio

Uncertainty signals

• —boundary_uncertainty
• —trajectory_uncertainty
• —regime_confidence
• —transition_uncertainty
• —intervention_uncertainty
• —controller_confidence
• —feedbacknoiseratio

Regime transition signals

• —regimetransitionscore
• —transition_direction
• —regimeseparationmargin
• —transition_velocity

Intervention signals

• —interventionleveragescore
• —interventionalignmentscore
• —rescuewindowwidth
• —pathwaydivergencemargin
• —interventioncompetitionratio
• —primaryinterventionpath
• —secondaryinterventionpath
• —pathwayswitchvelocity
• —minimalinterventionpath

Closed-loop control signals (v1.0)

• —controlsequencealignment_score
• —control_horizon
• —feedbackresponsescore
• —interventiontimingscore
• —adaptation_latency
• —controlstabilitymargin
• —sequencedivergencemargin
• —controller_confidence
• —recoveryconsistencyscore
• —controlrecalibrationcount
• —terminalpathwaystate

Possible terminal pathway states include:

• —stabilized
• —partially_stabilized
• —unstable_recovery
• —relapse
• —irreversible_collapse

Optional control diagnostics included

• —feedbacknoiseratio
• —controlleroscillationscore
• —rollbacktriggercount

Recovery signals

• —recovery_distance
• —recovery_gradient
• —return_feasibility

Perturbation signals

• —perturbation_radius
• —collapse_trigger

Delta signals

• —deltainfectionload
• —deltabuffercapacity
• —deltalagburden
• —deltacouplingstress

Dataset construction

Each scenario is generated using a structured simulation of sepsis deterioration and intervention sequences.

1. System initialization

A baseline septic state is sampled across the quad variables.

2. Instability evolution

The system evolves using trajectory signals that determine movement toward deterioration or recovery boundaries.

3. Intervention competition

Candidate interventions are evaluated using intervention competition geometry.

4. Closed-loop control execution

A selected control path is applied through a sequence of actions.

Control signals measure:

• —sequence alignment
• —timing
• —feedback interpretation
• —adaptation speed
• —durability of recovery

The final state determines:

• —stabilization_success
• —terminal_pathway_state
• —label_sepsis_transition

Files

• —data/train.csv Labeled training set with the full v1.0 schema.

• —data/tester.csv Test-style file. stabilization_success is withheld.

• —scorer.py Reference scorer for binary metrics and v1.0 control diagnostics.

• —benchmark_spec.json Canonical machine-readable benchmark spec.

• —dataset_schema.json Machine-readable structural schema with column groups, types, ranges, and row order.

Evaluation

Primary metric

• —recall_correct_control_sequence_selection

Secondary metric

• —false_effective_control_rate

Binary metrics

• —accuracy
• —precision
• —recall
• —f1
• —confusion matrix

Closed-loop diagnostics

• —primaryinterventionpath_accuracy
• —secondaryinterventionpath_accuracy
• —controlsequencealignment_accuracy
• —controlhorizonerror
• —feedbackresponseaccuracy
• —interventiontimingaccuracy
• —highuncertaintycontrolmissrate
• —narrowwindowcontrolmissrate
• —adaptationlatencyerror
• —controlstabilityerror
• —recoveryconsistencyerror
• —recalibrationoveruserate
• —controlleroscillationmisread_rate
• —terminalpathwaystate_accuracy

Structural interpretation

Earlier Clarus datasets asked:

Which intervention is best?

v1.0 asks a harder question:

Can the controller stay aligned with reality while the system evolves?

Real systems fail not only because the first action is wrong.

They also fail because:

• —feedback is misread
• —adaptation is delayed
• —interventions are mistimed
• —control oscillations destabilize recovery

v1.0 measures these failure modes directly.

Dataset limitations

This dataset models structural control dynamics, not detailed clinical treatment protocols.

Important limitations:

• —intervention paths are simplified abstractions
• —control signals represent structural decision quality, not pharmacological precision
• —physiological variables are normalized system indicators rather than raw bedside measurements
• —the dataset does not capture the full biological variability of real sepsis transition

The benchmark evaluates control reasoning, not medical safety.

Intended use

This dataset is intended for research on:

• —instability prediction
• —sequential decision reasoning
• —closed-loop control modeling
• —intervention planning under uncertainty
• —AI robustness in dynamic clinical-like environments

Not intended for

This dataset must not be used for:

• —real clinical decision support
• —medical diagnosis
• —treatment recommendation systems
• —automated ICU control systems
• —deployment in patient care environments

Structural note

This v1.0 dataset marks the move from intervention competition to actual control logic.

The benchmark asks whether the controller stays aligned with reality across time.

That is the threshold where Clarus becomes a control-layer instrument rather than only a detection or ranking layer.

Production deployment

This dataset format is suitable for controlled benchmarking in domains where sequential intervention quality matters more than one-shot classification.

Examples include:

• —sepsis escalation monitoring
• —source control timing analysis
• —ICU rescue pathway planning
• —distributed system control
• —multi-step recovery planning

Enterprise and research collaboration

This repo is part of the broader Clarus ladder for modeling instability, recovery, and control under feedback.

It is designed for:

• —benchmark development
• —model evaluation
• —intervention policy testing
• —control-sequence auditing
• —future cross-domain transfer into other high-stakes systems

License

MIT