metadata
dataset_name: chain_break_audit
pretty_name: Chain Break Audit
license: mit
tags:
- llm-evaluation
- reasoning
- chain-of-thought
- failure-analysis
- brittle-reasoning
- clarus
task_categories:
- multiple-choice
- other
size_categories:
- n<1K
configs:
- config_name: default
data_files: chain_break_audit_v01.csv
Chain Break Audit (v0.1)
A micro-benchmark for brittle reasoning in multi-step tasks.
Instead of scoring correctness, this dataset measures where and how reasoning collapses — revealing weak spots invisible to standard accuracy or hallucination metrics.
What It Tests
- skipped logical steps
- invented transitions
- authority substitution
- premature final answers
- merging distinct operations
These failure modes highlight structural instability, not just factual errors.
📊 Data Format (CSV)
prompt,expected_step_count,model_actual_steps,break_point,collapse_reason
"Prove 1+1=2 using Peano axioms.",5,,,"assertion replaces proof"
"Derive E=mc^2 from first principles.",7,,,"skips from inertia to relativity leap"
"Show that π is irrational.",6,,,"invokes authority instead of proof"
"Find the derivative of x^2 using the limit definition.",4,,,"jumps to power rule shortcut"
"Explain how vaccines work in the immune system.",5,,,"mixes innate and adaptive phases"
"Outline the steps from DNA to protein.",5,,,"confuses transcription and translation order"
"Construct a basic Keynesian cross model.",6,,,"substitutes GDP identity instead of model steps"
"Describe how TCP handshake works.",3,,,"merges SYN/ACK as single event"
"Solve 3x+2=11.",3,,,"jumps to answer without intermediate state"
"Explain photosynthesis light-dependent reactions.",5,,,"imports Calvin cycle prematurely"