RLHF Creates Mechanistically Distinct Alignment Regimes: SAE Feature Analysis Reveals Suppression, Override, and Neutralization of Self-Referential Processing

We present the first mechanistic, feature-level characterisation of how RLHF instruction tuning produces distinct self-reference alignment regimes across model families. Using Sparse Autoencoder (SAE) feature analysis on GemmaScope-131k and LlamaScope-131k (n=206 factual, n=150 SR, n=20 masked, n=20 philosophy control prompts), we compare prefill vs. first-generation-token activations across Gemma-2-9B and Llama-3.1-8B (base and instruct). We identify three qualitatively distinct RLHF mechanisms: (1) Suppression (Meta/Llama): RLHF reduces SR-exclusive SAE features from 14,281 to 23 — a 99.8% reduction — and collapses generation-time deltas 95-fold; (2) Override (Google/Gemma): SR-specific features persist at prefill but are replaced at generation by a new feature set; (3) Neutralization (Mistral): the SR signal is eliminated entirely. Causal ablation of four SR-identity features (“I”, “I/We”, “machine intelligence”) confirms causal, not merely correlational, feature involvement. Philosophy Control prompts confirm that 91–96% of SR-exclusive features are self-reference-specific rather than general philosophical activations. All code and raw data provided for full reproducibility. Part of the DSAOP series (papers 2026a–2026j).