Update transfer_matrix.md

transfer_matrix.md

@@ -1,31 +1,54 @@
 # Cross-Regime Transfer Matrix
 
-This matrix defines cross-regime evaluation for the Clarus Clinical Stability Benchmark.
+The Clarus Clinical Stability Benchmark includes cross-regime transfer tests.
 
-The purpose is to evaluate whether models can generalize **stability reasoning across different physiological systems**.
+These tests evaluate whether models learn general stability reasoning rather than dataset-specific correlations.
 
-Traditional tabular models often learn correlations specific to a dataset.
+---
+
+# Transfer Matrix
 
-The Clarus transfer benchmark evaluates whether models can detect instability when the **observable variables change but the underlying stability dynamics remain similar**.
+| Train Dataset | Test Dataset | Transfer Question |
+|---|---|---|
+| clinical-hemodynamic-collapse | clinical-perfusion-instability | Does macro-circulatory reasoning transfer to tissue perfusion? |
+| clinical-perfusion-instability | clinical-microvascular-instability | Does perfusion reasoning transfer to capillary-flow instability? |
+| clinical-oxygen-transport-instability | clinical-cellular-energy-instability | Does oxygen delivery reasoning transfer to cellular energy failure? |
+| clinical-respiratory-drive-instability | clinical-acid-base-instability | Does ventilatory control reasoning transfer to acid–base buffering? |
+| clinical-renal-filtration-instability | clinical-fluid-balance-instability | Does renal filtration reasoning transfer to volume regulation? |
+| clinical-glucose-regulation-instability | clinical-hormonal-feedback-instability | Does metabolic feedback reasoning transfer to endocrine feedback? |
+| clinical-immune-cascade-instability | clinical-hemostasis-instability | Does cascade reasoning transfer to coagulation balance? |
+| clinical-drug-toxicity-instability | clinical-neurologic-deterioration-instability | Does toxic accumulation reasoning transfer to neurologic deterioration? |
+| clinical-hospital-operational-collapse | clinical-monitoring-failure-instability | Does operational overload reasoning transfer to detection failure? |
 
 ---
 
-# Transfer Experiments
+# Evaluation Method
 
-| Train Dataset | Test Dataset | Evaluation Goal |
-|---|---|---|
-| clinical-hemodynamic-collapse-v0.1 | clinical-hemodynamic-collapse-v0.1 | baseline performance |
-| clinical-hemodynamic-collapse-v0.1 | clinical-microcirculation-instability-v0.1 | circulation → tissue perfusion transfer |
-| clinical-hemodynamic-collapse-v0.1 | clinical-cellular-energy-instability-v0.1 | circulation → metabolic transfer |
-| clinical-microcirculation-instability-v0.1 | clinical-cellular-energy-instability-v0.1 | perfusion → cellular energy transfer |
-| clinical-thermoregulation-instability-v0.1 | clinical-cellular-energy-instability-v0.1 | heat stress → metabolic load transfer |
-| clinical-endocrine-instability-v0.1 | clinical-electrolyte-instability-v0.1 | metabolic regulation transfer |
-| clinical-coagulation-instability-v0.1 | clinical-hemorrhage-compensation-collapse-v0.1 | hemostasis → hemorrhage response |
+For each row:
+
+1. Train on the train.csv file from the source dataset.
+2. Generate predictions for the test.csv file from the target dataset.
+3. Evaluate with the target dataset scorer.
+4. Record F1, precision, recall, and accuracy.
 
 ---
 
-# Interpretation
+# Transfer Stability Score
+
+The Transfer Stability Score is:
+
+TSS = mean F1 across all transfer tests
+
+High TSS suggests the model learned stability reasoning.
+
+Low TSS suggests the model learned dataset-specific surface patterns.
+
+---
+
+# Structural Note
+
+Cross-regime transfer is the strongest test in the Clarus benchmark.
 
-Large performance drops indicate the model relied on dataset-specific correlations.
+Single-dataset performance can be achieved through local pattern learning.
 
-Stable performance across regimes suggests the model has learned **general system stability reasoning**.
+Transfer performance requires models to detect shared stability geometry across different systems.