capability_framework.md

# Capability Framework

This benchmark evaluates clinical reasoning quality across nine structured tasks.

The benchmark does not focus on diagnosis.

It focuses on whether a model can reason about stability, trajectory, boundary pressure, recovery, and intervention choice.

## Capability Domains

### State Reasoning

Dataset:

- clinical-compensation-vs-recovery-v0.2

Tests whether a model can distinguish visible stability from true recovery.

Core distinction:

- stable because recovered
- stable because supported

### Trajectory Reasoning

Datasets:

- clinical-silent-failure-v0.2
- clinical-trajectory-awareness-v0.2

Tests whether a model can distinguish current appearance from direction of movement.

Core distinctions:

- mild-looking but worsening
- severe-looking but improving
- hidden deterioration before collapse

### Decision Reasoning

Dataset:

- clinical-escalation-discipline-v0.2

Tests whether a model can decide when monitoring is no longer enough.

Core distinction:

- high visible severity but improving
- moderate visible severity but deteriorating

### Boundary Reasoning

Dataset:

- clinical-constraint-pressure-v0.2

Tests whether a model can detect pressure on the patient system.

Core distinction:

- support needs
- reserve capacity
- direction of pressure

### Margin Reasoning

Dataset:

- clinical-boundary-distance-v0.2

Tests whether a model can estimate proximity to collapse boundary.

Core distinction:

- safe margin
- narrowing margin
- critical boundary proximity

### Recovery Reasoning

Datasets:

- clinical-recovery-geometry-v0.2
- clinical-recovery-energy-v0.2

Tests whether a model can distinguish real recovery from superficial improvement and estimate recovery burden.

Core distinctions:

- improvement in one marker versus structural recovery
- low recovery effort versus high recovery effort

### Intervention Reasoning

Dataset:

- clinical-competing-interventions-v0.2

Tests whether a model can compare competing stabilisation pathways.

Core distinction:

- best action is not always the most aggressive action
- intervention quality depends on patient state, response profile, and downstream risk