Update README.md

README.md

@@ -14,51 +14,45 @@ tags:
 
 # Clarus Clinical Stability Benchmark
 
-The Clarus Clinical Stability Benchmark evaluates whether machine learning models can infer latent system instability from interacting proxy signals across multiple physiological and operational regimes.
-
-Unlike traditional tabular benchmarks that reward threshold detection or single-feature correlations, Clarus datasets are designed so that instability emerges from interactions between variables.
-
-The benchmark therefore measures stability reasoning rather than pattern memorization.
-
 The Clarus Clinical Stability Benchmark evaluates whether machine learning models can detect **latent instability in complex clinical systems**.
 
-Most tabular benchmarks test pattern recognition from static variables.
-
-This benchmark instead evaluates **reasoning about interacting system signals** that determine whether a system remains stable or moves toward collapse.
+Most tabular benchmarks reward models for learning correlations within a single dataset.  
+The Clarus benchmark instead evaluates whether models can infer instability from **interacting proxy signals across multiple physiological and operational regimes**.
 
-The datasets represent different physiological and operational regimes where instability emerges from **multi-variable interaction dynamics** rather than single-variable thresholds.
+Each dataset represents a simplified regime in which instability emerges from multi-variable interaction rather than single-variable thresholds.
 
 ---
 
 # Benchmark Concept
 
-The core idea of the benchmark is **latent stability geometry**.
+In real clinical systems, deterioration rarely occurs because one measurement crosses a threshold.
 
-In real clinical systems, instability arises when multiple interacting components drift simultaneously.
+Instead, instability emerges when several components drift simultaneously.
 
 Examples include:
 
-- circulatory compensation failure
-- microvascular perfusion loss
-- metabolic energy collapse
-- respiratory control failure
-- endocrine dysregulation
-- thermoregulatory breakdown
-- coagulation instability
-- hospital operational overload
+- circulatory compensation failure  
+- microvascular perfusion loss  
+- metabolic energy collapse  
+- respiratory control failure  
+- endocrine dysregulation  
+- thermoregulatory breakdown  
+- coagulation instability  
+- hospital operational overload  
 
-Each dataset captures one regime.
+Each dataset exposes a different regime while keeping the underlying structure similar:  
+**instability arises from interacting system signals.**
 
-The true generative logic is not published.
+The generative rules that determine the labels are intentionally not published.
 
-Models must infer instability from **interacting proxy signals**.
+Models must infer instability from observable proxies.
 
 ---
 
 # Included Datasets
 
-| Regime | Dataset Repo |
-|------|------|
+| Stability Regime | Dataset |
+|---|---|
 | Hemodynamic collapse | ClarusC64/clinical-hemodynamic-collapse-v0.1 |
 | Sepsis trajectory instability | ClarusC64/clinical-sepsis-trajectory-instability-v0.1 |
 | Intervention delay failure | ClarusC64/clinical-intervention-delay-failure-v0.1 |
@@ -75,18 +69,20 @@ Models must infer instability from **interacting proxy signals**.
 | Coagulation instability | ClarusC64/clinical-coagulation-instability-v0.1 |
 | Hospital operational collapse | ClarusC64/clinical-hospital-operational-collapse-v0.1 |
 
-Each dataset includes:
-
-train.csv  
-test.csv  
-scorer.py  
+Each dataset repository contains:
+data/train.csv
+data/test.csv
+scorer.py
 README.md
 
+
 ---
 
 # Evaluation Protocol
 
-All datasets use the same prediction format:
+Predictions must follow the format:
+
+
 scenario_id,prediction
 
 
@@ -97,20 +93,23 @@ MC101,0
 MC102,1
 
 
-Predictions are evaluated with the official Clarus scorer:
+Evaluation is performed using the **scorer located in the dataset repository**.
+
+Example:
 
 
 python scorer.py --predictions predictions.csv --truth data/test.csv --output metrics.json
 
 
-Metrics:
+The `--truth` path refers to the dataset's local `data/test.csv` file.
+
+Metrics reported include:
 
 - accuracy
 - precision
 - recall
 - f1
 - confusion matrix
-- dataset integrity diagnostics
 
 ---
 
@@ -122,7 +121,9 @@ The benchmark supports three evaluation settings.
 
 Train and test on the same dataset.
 
-This measures performance on a single stability regime.
+Purpose:
+
+Measure baseline performance within a single stability regime.
 
 ---
 
@@ -132,15 +133,12 @@ Train on one regime and test on another.
 
 Example:
 
-Train:
-
-clinical-hemodynamic-collapse
+Train → clinical-hemodynamic-collapse-v0.1  
+Test → clinical-microcirculation-instability-v0.1
 
-Test:
+Purpose:
 
-clinical-microcirculation-instability
-
-Large performance drops indicate the model learned **surface patterns rather than stability reasoning**.
+Determine whether models learn **general stability reasoning** rather than dataset-specific correlations.
 
 ---
 
@@ -148,17 +146,21 @@ Large performance drops indicate the model learned **surface patterns rather tha
 
 Train on multiple datasets simultaneously.
 
-Evaluate whether models can learn **general stability reasoning across physiological systems**.
+Evaluate performance across all regimes.
+
+Purpose:
+
+Test whether models can learn shared stability representations across physiological systems.
 
 ---
 
 # Dataset Design Principles
 
-The Clarus datasets follow several design rules.
+The Clarus datasets follow several explicit design rules.
 
 ### No Single-Feature Dominance
 
-No observable variable strongly predicts the label alone.
+No observable variable strongly predicts the label independently.
 
 Target:
 
@@ -168,7 +170,7 @@ Target:
 
 ### Interaction-Based Labels
 
-Instability emerges from interactions between multiple variables.
+Instability emerges from interactions between multiple variables rather than isolated thresholds.
 
 ---
 
@@ -182,26 +184,40 @@ This prevents trivial heuristics.
 
 ### Decoy Variables
 
-Some variables appear meaningful but are not part of the label rule.
+Some variables appear meaningful but do not determine the label independently.
 
 ---
 
 ### Hidden Generative Logic
 
-The dataset generator and latent rule equations are not published.
+The dataset generator and rule equations are intentionally not published.
 
 Models must infer instability from proxy signals.
 
 ---
 
-# Structural Note
+# Baseline Results
+
+Reference baseline experiments are provided in:
 
-These datasets represent simplified proxies for real clinical system dynamics.
+baseline_results.md
 
-They are designed for research into **latent stability reasoning**, not clinical decision support.
+
+These establish approximate difficulty levels for common tabular models.
 
 ---
-- stability_manifold.md — explains the shared stability geometry across regimes
+
+# Benchmark Architecture
+
+The benchmark can be interpreted as observing a **shared stability manifold** through different clinical regimes.
+
+Each dataset exposes a different control system while preserving the underlying concept of instability emerging from interacting signals.
+
+Additional details are provided in:
+
+
+stability_manifold.md
+
 
 ---
 
@@ -209,13 +225,11 @@ They are designed for research into **latent stability reasoning**, not clinical
 
 The benchmark supports research into:
 
-- stability reasoning in machine learning
-- interaction-based tabular reasoning
-- cross-domain system modeling
-- clinical early warning systems
-- multi-variable instability detection
-
-The Clarus benchmark evaluates whether models can detect latent instability across interacting system regimes rather than learning surface patterns within a single dataset.
+- system stability reasoning
+- interaction-based tabular learning
+- cross-domain generalization
+- clinical early warning modeling
+- infrastructure and system risk detection
 
 ---