NeurIPS 2026 Submission

WhyLab is a causal safety monitoring framework that prevents cognitive policy oscillation in self-improving AI agents. It provides three contributions:

• C1: Information-theoretic drift detection
• C2: Sensitivity-aware effect filtering (E-values + partial R^2)
• C3: Lyapunov-bounded adaptive damping

flowchart LR
    Agent[Agent Policy $\theta_t$] --> Evaluate[Test & Reward $\Delta R$]
    Evaluate --> Audit{Audit Layer: C1, C2, C3}
    Audit --> Safe[Safe Update $\theta_{t+1}$]
    Safe --> Agent

paper/          # LaTeX source (main.tex, references.bib)
experiments/    # All experiment code and results
  results/      # CSV/Parquet output files
  prompts/      # LLM prompt templates for E5/E7
  cache/        # Cached API responses (gitignored)
  data/         # Downloaded datasets (gitignored)
submission/     # Packaged ZIP files for OpenReview

pip install -r requirements.txt

These experiments are fully reproducible with no external dependencies:

# E1 & E2: Core components
python -m experiments.e1_drift_detection
python -m experiments.e2_sensitivity_filter

# E3a: Stability Validation (PID, SGD, Adam, Lyapunov)
python -m experiments.e3a_stationary

# Proxy Correlation Analysis (Theorem 1 validation)
python -m experiments.proxy_correlation_analysis

# E6: Non-stationary Agent Environment
python -m experiments.e6_nonstationary_agent

These experiments require a .env file containing API keys (e.g., GEMINI_API_KEY or OPENAI_API_KEY depending on the model):

cp .env.example .env
# Edit .env and add your respective API Keys

# E5: SWE-bench Lite (300 problems x 5 seeds)
python -m experiments.e5_swebench_benchmark

# E7: Dynamic ReAct Benchmark with GPT-5.4
python -m experiments.e7_react_dynamic_benchmark

Note: E5/E7 results are non-deterministic due to LLM sampling. Pre-computed results are available in experiments/results/.

# E5 subset analysis (oscillating vs non-oscillating)
python -m experiments.e5_subset_analysis

# Safety baseline comparison (Best-of-N, Rollback, etc.)
python -m experiments.e5_safety_baselines

Statistical Reproducibility Note: All non-deterministic differences reported in the paper (e.g., Pass@1 improvements in E5) are validated using Bootstrap 95% Confidence Intervals ($N=10,000$). For reviewers wishing to recalculate the CI, you can adapt the following snippet:

import numpy as np
# Example: Bootstrap CI for Pass@1 differences
diffs = []
for _ in range(10000):
    idx = np.random.randint(0, len(pass_reflexion), len(pass_reflexion))
    diff = pass_whylab[idx].mean() - pass_reflexion[idx].mean()
    diffs.append(diff)
print(f"Pass@1 diff 95% CI: {np.percentile(diffs, [2.5, 97.5])}")

Requires a TeX distribution (MiKTeX or TeX Live):

cd paper
pdflatex main.tex
bibtex main
pdflatex main.tex
pdflatex main.tex

• Best-of-N baselines in Table 6 are estimated via 1-(1-p_1)^N under an independence assumption (not actual parallel runs). This is stated in the paper.
• SWE-bench evaluation (E5) uses lightweight (string-match) test execution.
• Dynamic ReAct Framework (E7) utilizes OpenAI's GPT-5.4 model to demonstrate that the framework generalizes to flagship out-of-family models.

This code is provided for academic review purposes.