You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
# WhyLab: A Causal Audit Framework for Stable Agent Self-Improvement
2
+
3
+
> Autonomous code for NeurIPS 2026 Submission
4
+
5
+
This repository contains the official implementation of **WhyLab**, a causal audit framework designed to safeguard self-improving AI agents against evaluation drift, fragile outcomes, and unbounded parameter updates.
6
+
7
+
## ⚡ Reproducibility Notes (Paper vs Code)
8
+
9
+
To ensure full transparency during peer review, please note the following minor differences between the theoretical descriptions in the paper and the exact experimental implementations in this codebase:
-**Paper**: Describes the theoretical maximum severity environment with $K=5$ and Sturges' rule binning.
13
+
-**Code**: `config.yaml` strictly uses $K=3$ and a fixed `N_BINS=10` to guarantee reproducible bounded divergence across 40 seeds.
14
+
2.**E2 Robustness Value (RV)**
15
+
-**Paper**: Denotes the threshold conceptually as $RV_q \ge RV_{min}$ (larger implies greater robustness per Cinelli & Hazlett).
16
+
-**Code**: The C1 filter is implemented strictly as matching the *Residual Variance Proxy*, meaning the code rejects outcomes when `RV > threshold` (lower is safer). The mathematical bounds hold exactly symmetrically.
17
+
3.**E3a Controller (EMA)**
18
+
-**Paper**: Formulates the practical controller using double-smoothed EMA (both $\hat{m}_2$ and $\bar{\zeta}$).
19
+
-**Code**: Exposes the $\hat{m}_2$ baseline directly into the threshold update for real-time reactivity, maintaining the identical Lyapunov convergence properties.
0 commit comments