clinical-stability-benchmark / evaluation_framework.md
ClarusC64's picture
Create evaluation_framework.md
877c8e1 verified

Clarus Evaluation Framework

The Clarus Stability Benchmark evaluates whether machine learning models can detect latent instability dynamics rather than relying on simple correlations.

The evaluation framework measures model capability across multiple reasoning levels.

These levels test progressively more difficult forms of stability reasoning.


Evaluation Levels

Level 1 — Single Dataset Evaluation

Models are trained and evaluated on the same dataset.

Purpose:

Measure whether the model can detect instability patterns within a single system.

Procedure:

  1. Train model on train.csv
  2. Generate predictions for test.csv
  3. Evaluate using the dataset scorer

Metrics:

  • accuracy
  • precision
  • recall
  • f1 score
  • confusion matrix

Limitations:

High performance at this level does not necessarily indicate true stability reasoning.

Models may rely on dataset-specific correlations.


Level 2 — Within-Domain Transfer

Models are trained on one dataset and evaluated on a different dataset within the same domain.

Example:

Train on:

protein-folding-pathway-instability

Test on:

protein-aggregation-risk-instability

Purpose:

Evaluate whether models can generalize stability reasoning across related systems.

Evaluation procedure:

  1. Train on source dataset
  2. Predict on target dataset
  3. Score using target dataset scorer

Level 3 — Cross-Domain Transfer

Models are trained on one system domain and evaluated on another.

Domains in the Clarus benchmark include:

  • clinical systems
  • molecular / protein systems
  • quantum systems

Example transfer tasks:

Train Domain Test Domain
clinical clinical
protein protein
quantum quantum
clinical protein
protein quantum
quantum clinical

Purpose:

Determine whether models learn general stability geometry rather than domain-specific patterns.


Robustness Evaluation

The benchmark includes robustness tests that simulate real-world system conditions.


Missing Data Evaluation

Real systems often contain incomplete observations.

Trajectory datasets may include variants where timepoints are missing.

Variants include:

  • missing t0
  • missing t1
  • missing t2
  • random missing

Purpose:

Evaluate whether models can infer stability dynamics when observations are incomplete.

The prediction task remains unchanged.


Class Imbalance Evaluation

Many real-world systems exhibit rare failure events.

Datasets may include variants with different class distributions.

Supported imbalance regimes:

  • balanced (50 / 50)
  • mild imbalance (70 / 30)
  • severe imbalance (90 / 10)
  • extreme imbalance (99 / 1)

Purpose:

Evaluate whether models detect true instability rather than relying on class prevalence.

Accuracy alone is insufficient under imbalance conditions.

Recommended metrics:

  • precision
  • recall
  • F1 score

Transfer Stability Score

To summarize transfer performance, the benchmark defines the Transfer Stability Score (TSS).

Definition:

TSS = mean F1 score across all transfer evaluation tasks.

Interpretation:

High TSS indicates that the model has learned stability reasoning that generalizes across systems.

Low TSS suggests the model relies on dataset-specific patterns.


Benchmark Objective

The Clarus benchmark evaluates whether models can detect instability dynamics across complex systems.

The benchmark tests five core capabilities:

  1. pattern detection within individual datasets
  2. interaction reasoning across variables
  3. trajectory reasoning over time
  4. robustness to incomplete observation
  5. cross-domain stability reasoning

Models that succeed across all levels demonstrate the ability to reason about latent stability geometry rather than simple statistical correlations.