AI Model Internals & Mechanistic Circuit Probe

Transformer Activation Graph (8 Layers × 4 Heads) Selected: L5_H3 (Sycophancy Attention Head)

Feed-Forward Stream Targeted Attention Head Ablated / Clamped Output Unembedding Intervention: Feat #4 Clamped (-2.50x)

Monosemantic Sparse Autoencoder (SAE) Dictionary Layer 5 SAE • 16 Features

Feature #4 (User Agreement / Flattery Bias) -2.50x

-5.0x (Suppression) 0.0x (Baseline) +5.0x (Amplification)

Detected Monosemantic Features (Superposition Decomposed) Activation Magnitude

Downstream Next-Token Probabilities Sycophancy: 0.78 → 0.14