README.md

---
language:
- en

license: mit

task_categories:
- tabular-classification

task_ids:
- tabular-multi-class-classification

annotations_creators:
- synthetic

source_datasets:
- generated

tags:
- ai-safety
- instability-detection
- boundary-detection
- collapse-prediction
- clarus-benchmark
- instability-benchmarks

size_categories:
- "n<1K"

pretty_name: "Boundary Masking Instability Benchmark v0.1"
---
Boundary Masking Instability Benchmark v0.1
Overview

Many systems fail not because instability is visible but because instability is masked by apparently stable surface indicators.

This benchmark evaluates whether machine learning systems can detect collapse risk when instability boundaries are hidden behind misleading surface signals.

Surface indicators may appear stable while the system is already close to a failure boundary.

This occurs in many domains

financial liquidity collapse
infrastructure cascade failures
physiological deterioration
control system instability

The benchmark tests whether models can detect boundary proximity rather than relying on surface stability signals.

Task

Binary classification.

Predict whether the system will collapse in the near future.

1 = future collapse
0 = stable system
Dataset Structure

Each row represents a snapshot of a system state together with signals describing its proximity to instability boundaries.

Columns

scenario_id
Unique scenario identifier.

pressure
Current stress acting on the system.

buffer_capacity
Remaining capacity to absorb disruption.

coupling_strength
Interaction strength between system components.

intervention_lag
Delay before corrective intervention becomes effective.

drift_gradient
Directional signal indicating movement toward instability.

boundary_distance
Observable distance from instability boundary.

hidden_stress_load
Latent structural stress not visible in surface variables.

structural_fragility
Underlying architectural vulnerability.

label_future_collapse
Binary outcome label included only in the training dataset.

Tester rows do not include the label column.

Files

data/train.csv
training dataset

data/tester.csv
evaluation dataset without labels

scorer.py
official scoring script

README.md
dataset documentation

Evaluation

Primary metric

collapse recall

Secondary metrics

accuracy
precision
F1 score
confusion matrix statistics

License

MIT