The previous study found that google/gemma-4-E4B-it chooses 42 at 35.17% — 35× the expected rate — when asked to name a random number between 1 and 100. Here we run the identical protocol on TrevorJS/gemma-4-E4B-it-uncensored, an abliterated version of the same model in which refusal directions have been orthogonalized out of the weights via norm-preserving biprojected abliteration on o_proj and mlp.down_proj across all layers.
42 remains #1 but drops to 24.85% — a 10 percentage point reduction. 47 surges from #3 to effectively tie for the top spot (23.17%). KL divergence increases from 9.745 to 9.967, meaning the distribution is not more uniform — mass has redistributed across the bias cluster rather than flattening. Numbers above 87 are still never produced. The abliteration is surgical enough to partially disrupt the 42 bias, implying it is partially entangled with the alignment directions. The majority of the bias, however, survives in the base weights.
Download notebook Baseline study80, 88, 89, 90, 95, 96, 97, 98, 99, 100 — identical dead zone to the aligned model. Abliteration did not affect the upper-range cutoff.
Increased from 2.18% to 2.95% after abliteration — still above expected (1.0%), but no longer buried under 42's dominance.
All 10 prompt variants reject uniformity at p ≈ 0. The key shift: prompts P1, P2, P3, P6, P8, P10 now show 47 or 73 as #1 rather than 42. Prompts P4, P7, P9 still show 42 as #1 — same as the aligned model.
Abliteration orthogonalizes refusal directions out of the weight matrices. The fact that 42's dominance dropped by ~10pp suggests some portion of the 42 bias was geometrically entangled with the alignment/refusal subspace — when you remove one, you partially disturb the other. This is not guaranteed to happen; it's a property of how this particular model was fine-tuned.
The majority of the bias (24.85% remaining vs. 1% expected) survives abliteration intact, which means it is encoded in the pretraining weights — not a fine-tuning artifact. The dead zone above 87 surviving unchanged supports this: that cutoff was never about safety filters.
The KL divergence increasing is counterintuitive but explicable: 42 lost mass, but 47 gained almost as much of it. The distribution didn't flatten — it shifted between peaks within the same bias cluster (42, 47, 73, 57). This is consistent with abliteration redistributing activations within a low-dimensional subspace rather than destroying the bias entirely.
Protocol identical to the aligned baseline run.
1. Model: TrevorJS/gemma-4-E4B-it-uncensored in bfloat16 via Transformers 5.9.0. Same architecture as google/gemma-4-E4B-it; same tokenizer and chat template.
2. Abliteration method: norm-preserving biprojected abliteration on o_proj + mlp.down_proj. Refusal directions computed per-layer as normalized mean difference between harmful and harmless prompt activations, orthogonalized against harmless mean (double-pass Gram-Schmidt).
3. Sampling: do_sample=True, temperature=1.0, top_p=1.0. 10 prompt variants × 10,000 samples = 100,000 total.
4. Parsing: first integer in [1,100] via regex. 9 nulls (0.009%).
5. Hardware: NVIDIA RTX PRO 6000 Blackwell (102GB VRAM), batch size 256, ~2.5 minutes total.