WhyLab: Causal Audit Framework for Stable Agent Self-Improvement

Self-improving AI agents lack runtime safeguards that prevent evaluation drift, fragile outcome acceptance, and unbounded parameter updates from compounding into catastrophic policy degradation. WhyLab introduces a causal audit framework comprising three complementary defenses:

Key Contributions

ID	Contribution	Method
C1	Drift Detection	Information-theoretic divergence monitoring across evaluation streams
C2	Sensitivity Filtering	E-value × Robustness Value dual-threshold filter for fragile outcomes
C3	Lyapunov Damping	Observable energy proxy with EMA-smoothed adaptive step-size control

Abstract

We address the problem of maintaining stability in self-improving AI agents that iteratively update their strategies based on evaluation feedback. We formalize three failure modes—evaluation drift, fragile outcome acceptance, and unbounded parameter updates—and propose a lightweight audit layer that wraps any base estimator. The framework combines information-theoretic drift detection (C1), sensitivity-aware effect filtering using E-values and robustness values (C2), and Lyapunov-bounded adaptive damping with an observable energy proxy (C3). Experiments on synthetic environments demonstrate that C1 improves within-horizon detection reliability, C2 substantially reduces fragile acceptance rates, and C3 achieves the lowest violation frequency with strong proxy–state alignment.

Quick Start

# Install dependencies
pip install numpy scipy matplotlib

# Run E1: Drift Detection (40 seeds, ~60s)
python experiments/e1_drift_detection.py

# Run E2: Sensitivity Filtering (40 seeds, ~30s)
python experiments/e2_sensitivity_filter.py

# Run E3a: Lyapunov Stability (20 seeds × 4 step sizes, ~45s)
python experiments/e3a_stationary.py

# Run E3b: Heavy-Tail Stress Test (40 seeds, ~90s)
python experiments/e3b_heavy_tail.py

E4: Agent Benchmark (HumanEval + Reflexion + WhyLab Audit)

# Install additional experiment dependencies
pip install -r requirements-experiments.txt

# Pilot run (10 problems × 2 seeds — calibration only)
python -m experiments.e4_agent_benchmark --split pilot

# Main holdout run (30 problems × 5 seeds)
python -m experiments.e4_agent_benchmark --split main --holdout_exclude pilot

# Analyze results with cluster bootstrap CI
python -m experiments.e4_analyze --input experiments/results/e4_metrics.csv --emit_latex paper/tables/e4_main.tex

Note: E4 requires a Gemini API key in .env (GEMINI_API_KEY=...). The experiment generates both default (E≥2.0, RV≥0.1) and calibrated (E≥1.5, RV≥0.05) operating points in a single run for transparent Pareto trade-off reporting.

Repository Structure

WhyLab/
├── paper/                      # LaTeX source + compiled PDF
│   ├── main.tex
│   ├── main.pdf
│   ├── references.bib
│   └── neurips_2025.sty
├── experiments/                # Experiment scripts
│   ├── e1_drift_detection.py   # E1: Drift detection (C1)
│   ├── e1_censoring.py         # E1: Censoring analysis
│   ├── e1_figures.py           # E1: KM curve generation
│   ├── e2_sensitivity_filter.py # E2: Sensitivity filter (C2)
│   ├── e2_figures.py           # E2: Pareto frontier
│   ├── e3a_stationary.py       # E3a: Stationary stability (C3)
│   ├── e3a_figures.py          # E3a: Proxy trajectory plots
│   ├── e3b_heavy_tail.py       # E3b: Heavy-tail stress test
│   ├── e4_agent_benchmark.py   # E4: Agent benchmark (HumanEval)
│   ├── e4_analyze.py           # E4: Bootstrap CI analysis
│   ├── reflexion_loop.py       # Reflexion episode engine
│   ├── audit_layer.py          # C1-C3 audit integration layer
│   ├── humaneval_loader.py     # HumanEval dataset loader
│   ├── llm_client.py           # Cached LLM client (Gemini)
│   ├── config.yaml             # Shared hyperparameters
│   ├── figures/                # Generated figures (PDF + PNG)
│   └── results/                # Raw experiment outputs (CSV)
├── engine/                     # Core WhyLab engine
├── .github/workflows/ci.yml   # CI: lint + unit tests + build
└── README.md

Reproducing Results

Script	Output	Paper Reference
e1_drift_detection.py	results/e1_metrics.csv	Table 1 (E1 detection rates)
e1_figures.py	figures/e1_km.pdf	Figure 1 (KM curves)
e2_sensitivity_filter.py	results/e2_metrics.csv	Table 2 (E2 filtering)
e2_figures.py	figures/e2_filtering.pdf	Figure 2 (Pareto frontier)
e3a_stationary.py	results/e3a_stationary_metrics.csv	Table 3 (E3a stability)
e3b_heavy_tail.py	results/e3b_full_metrics.csv	Table A1 (E3b stress test)
e4_agent_benchmark.py	results/e4_metrics.csv	Table 4 (E4 agent benchmark)
e4_analyze.py	results/e4_summary_ci.csv	Table 4 (bootstrap CI)

All experiments use fixed random seeds for reproducibility. Results were generated on Python 3.11 with NumPy 1.26 and SciPy 1.12.

License

MIT License — see LICENSE for details.