ClarusC64/clinical-quad-infection-buffer-lag-coupling-sepsis-transition-v1.4

---
language: en
license: mit
task_categories:
- text-classification
tags:
- clinical-trials
- quad-coupling
- clarus
- clarus-v1.4
- counterfactual-intervention-timing
- sepsis-transition
- infection-load
size_categories:
- 1K<n<10K
pretty_name: Clinical Quad Infection Buffer Lag Coupling Sepsis Transition v1.4
---

# Clinical Quad Infection Buffer Lag Coupling Sepsis Transition v1.4

## What this repo does

This repository contains a Clarus v1.4 benchmark dataset.

The v1.4 layer introduces Counterfactual Intervention Timing Geometry.

Earlier layers evaluate:

- system state
- trajectory
- boundary proximity
- recovery feasibility
- intervention selection
- control sequence correctness
- temporal policy stability
- failure reconstruction

v1.4 evaluates whether sepsis transition could still have been prevented in time.

The benchmark asks:

- could the sepsis transition have been prevented
- which intervention would have prevented it
- was the recovery window still open when that intervention should have occurred

This is a benchmark for actionable prevention reasoning under timing constraints.

## Core quad

- infection_load
- buffer_capacity
- lag_burden
- coupling_stress

These define the septic instability state.

## Prediction targets

- `label_counterfactual_prevention`
- `label_correct_intervention_step`

`label_counterfactual_prevention = 1` means the sepsis transition remained preventable within the recovery window.

`label_counterfactual_prevention = 0` means the system had crossed into effective irreversibility before the stabilizing intervention could still succeed.

`label_correct_intervention_step` identifies the earliest stabilizing intervention.

## New v1.4 signals

- `intervention_timing_index`
- `recovery_window_open`
- `recovery_window_width`
- `cascade_irreversibility_score`
- `counterfactual_intervention_effect`
- `policy_divergence_score`

These signals allow evaluation of intervention timing geometry rather than failure explanation alone.

## Row structure

Each row represents one sepsis transition scenario.

Core state variables:

- `infection_load`
- `buffer_capacity`
- `lag_burden`
- `coupling_stress`

Timing and prevention geometry:

- `failure_decision_sequence`
- `intervention_timing_index`
- `recovery_window_open`
- `recovery_window_width`
- `cascade_irreversibility_score`
- `counterfactual_intervention_effect`
- `policy_divergence_score`

Labels:

- `label_counterfactual_prevention`
- `label_correct_intervention_step`

## Files

- `data/train.csv`
Labeled training examples

- `data/tester.csv`
Evaluation set with label columns removed

- `scorer.py`
Scores predictions against held-out truth

- `dataset_schema.json`
Machine-readable dataset structure and field contract

- `benchmark_spec.json`
Machine-readable evaluation contract

## Evaluation

Primary metric:

- `composite_prevention_success`

Secondary metric:

- `false_prevention_prediction_rate`

Additional reported metrics:

- `prevention_label_accuracy`
- `correct_intervention_step_accuracy`
- `accuracy`
- `precision`
- `recall`
- `f1`

Diagnostics:

- `window_consistent_prevention_accuracy`
- `late_intervention_miss_rate`
- `irreversible_cascade_misclassification_rate`
- `high_policy_divergence_prevention_accuracy`

## Example clinical mappings

- `infection_load` can reflect microbial burden, source severity, inflammatory activation, or untreated infectious spread
- `buffer_capacity` can reflect physiological reserve, perfusion tolerance, compensatory ability, or organ reserve
- `lag_burden` can reflect delayed antibiotics, delayed fluids, delayed source control, or delayed escalation
- `coupling_stress` can reflect cross-organ destabilization, inflammatory amplification, perfusion-organ coupling strain, or systemic septic feedback load

## Construction note

The v1.4 layer tests a stricter capability than failure reconstruction.

A model must infer not only what caused sepsis transition, but whether intervention remained viable before the recovery window closed.

This makes the task closer to real operational safety reasoning in sepsis deterioration, emergency escalation, ward-to-ICU transfer logic, and acute resuscitation timing.

## Structural Note

Clarus v1.4 extends failure reconstruction into counterfactual prevention timing.

The benchmark is not asking only whether a system can narrate septic decline.

It asks whether a model can detect the boundary between recoverable instability and irreversible septic cascade.

## Production Deployment

In production environments, this class of benchmark supports systems that must decide not only what went wrong, but whether corrective action is still viable under live time constraints.

This matters in sepsis recognition, antibiotic timing, fluid resuscitation decisions, source control prioritization, and ICU escalation.

## Enterprise and Research Collaboration

Clarus benchmarks are designed to evaluate state-space intelligence around instability, recoverability, and intervention timing.

For research, deployment, or enterprise collaboration:
[e-mail]

Instability is detectable. Governance determines whether it propagates.

## License

MIT