The Self-Referential Limit of Affective Alignment: Simulation Dataset

Simulation dataset for the paper "The Self-Referential Limit of Affective Alignment: Why Self-Modifying Emotion AI Cannot Internally Guarantee Its Own Evaluative Stability."

This dataset contains results from a minimal sufficient model demonstrating two structural vulnerabilities of self-modifying affective AI systems: evaluative drift and affective category collapse.

Three simulation blocks are included:

(1) Evaluative Drift — tests four conditions (fixed evaluator, noisy oracle, self-referential update, self-referential + periodic external anchor) across three update rules (EMA, Bayesian MAP, Selection-based), 50 repetitions, T=1000 timesteps.

(2) Affective Category Collapse — models an agent with K=8 emotion categories under uniform vs. biased self-referential scenario generation, with and without external anchors, 50 repetitions, T=1000.

(3) Combined Model — integrates drift and collapse dynamics, sweeps anchor frequency from f=0 to f=1.0, identifies critical stabilization threshold (f* ≈ 0.005), T=2000, 50 repetitions.

A sensitivity analysis tests robustness across four parameter dimensions: vector dimensionality (N), candidate count (M), anchor strength, and softmax temperature, 30 repetitions each.

Key findings: (a) self-referential evaluative drift emerges even under exogenous-noise-free conditions; (b) biased self-referential generation reduces category entropy from 3.0 to approximately 1.5 bits; (c) external anchors at f ≥ 0.005 reduce drift by over 90% and restore entropy to near-maximum; (d) results are robust across all parameter sweeps.

All Python source code is included for full reproducibility. Simulations use fixed random seeds (base seed = 42). No external APIs or datasets are required. Dependencies: Python 3.10+, numpy, scipy, matplotlib, seaborn.