Create stability_manifold.md

stability_manifold.md

@@ -0,0 +1,64 @@
+# Stability Manifold View
+
+The Clarus Clinical Stability Benchmark can be read as a set of local views into a shared stability manifold.
+
+Each dataset exposes a different clinical regime.
+
+The observable variables change across regimes, but the benchmark asks whether models can detect the same deeper question:
+
+Is the system moving toward stability or instability?
+
+---
+
+# Core Idea
+
+A model should not only learn:
+
+- lactate patterns
+- glucose patterns
+- respiratory patterns
+- staffing patterns
+
+It should learn the broader structure:
+
+- pressure rising
+- buffer weakening
+- coupling increasing
+- response delay widening
+- recovery margin closing
+
+This is the hidden stability geometry the benchmark probes.
+
+---
+
+# Regime Map
+
+| Stability Axis | Example Dataset |
+|---|---|
+| Pressure load | clinical-hemodynamic-collapse-v0.1 |
+| Buffer exhaustion | clinical-fluid-balance-instability-v0.1 |
+| Coupling cascade | clinical-organ-coupling-cascade-v0.1 |
+| Delayed response | clinical-intervention-delay-failure-v0.1 |
+| Recovery window | clinical-recovery-window-detection-v0.1 |
+| Compensation failure | clinical-hemorrhage-compensation-collapse-v0.1 |
+| Control-loop failure | clinical-autonomic-instability-v0.1 |
+| Cellular energy failure | clinical-cellular-energy-instability-v0.1 |
+| Operational collapse | clinical-hospital-operational-collapse-v0.1 |
+
+---
+
+# Why This Matters
+
+Most models can fit one dataset.
+
+The harder question is whether they can recognize instability across regimes.
+
+A strong model should detect similar stability geometry even when the surface variables change.
+
+That is the purpose of the cross-regime transfer tests.
+
+---
+
+# Benchmark Claim
+
+The Clarus benchmark evaluates whether models can move from local pattern recognition toward general stability reasoning.