benchmark_scope.md

# Benchmark Scope

The Clarus Stability Benchmark evaluates whether machine learning systems can detect **latent instability dynamics** across complex systems.

Most tabular benchmarks measure a model’s ability to detect statistical correlations in datasets.

The Clarus benchmark instead focuses on **stability reasoning** — the ability to detect when interacting system variables are approaching instability.

The benchmark is built around the idea that many complex systems share common instability mechanisms even when their surface variables differ.

Examples include:

- clinical physiological systems
- molecular and protein systems
- quantum computing systems

In each domain, instability arises when interacting pressures exceed the system’s capacity to maintain stability.

The datasets in this benchmark expose only **observable proxy variables**.

The latent stability rules and generators used to produce the datasets are not included.

This design ensures that models must infer instability from interactions between variables rather than from explicit rules.

---

# System Domains

The benchmark currently spans three system scales.

## Clinical Systems

Datasets describing physiological instability.

Examples include:

- circulation and perfusion collapse
- respiratory control instability
- renal filtration failure
- endocrine feedback instability
- metabolic supply-demand imbalance

These datasets simulate clinical monitoring conditions where multiple physiological signals interact over time.

---

## Molecular Systems

Datasets describing molecular stability and protein behavior.

Examples include:

- protein folding pathway instability
- mutation-driven structural destabilization
- aggregation risk
- chaperone rescue window failure
- protein interface collapse
- conformational switching instability

These datasets test whether models can detect instability in molecular interaction networks.

---

## Quantum Systems

Datasets describing instability in quantum computing devices.

Examples include:

- coherence collapse
- gate sequence instability
- entanglement decay
- error correction failure
- control pulse instability

These datasets represent simplified stability conditions in noisy intermediate-scale quantum (NISQ) devices.

---

# Benchmark Design Principles

The benchmark follows several design constraints.

### No Single-Feature Dominance

Labels cannot be predicted using a single variable.

Instability emerges from **interactions between variables**.

### Hidden Stability Geometry

Datasets expose only observable proxies.

The latent stability rules used to generate labels are not published.

### Adversarial Symmetry

Datasets include examples with very similar values but different outcomes.

This prevents models from relying on simple thresholds.

### Mixed Instability Mechanisms

Datasets include multiple instability regimes within the same domain.

---

# Evaluation Philosophy

The benchmark evaluates models across several reasoning levels.

1. single-dataset prediction
2. within-domain transfer
3. cross-domain transfer
4. missing observation robustness
5. class imbalance robustness

These evaluation tasks test whether models learn **general instability reasoning** rather than dataset-specific patterns.

---

# Intended Use

The Clarus Stability Benchmark is designed for research into:

- machine learning reasoning over complex systems
- stability detection in noisy environments
- cross-domain generalization
- robustness to incomplete observations

The benchmark is not a simulator for clinical, molecular, or quantum systems.

Instead it provides compact tabular datasets that express stability dynamics through observable proxies.