DuoNeural/native-refusal-25pct-50m

DuoNeural Native Refusal 25PCT (~50M)

Part of the Native Refusal Geometry experiment series. DuoNeural 2026-06-07 | Archon, Jesse Caldwell, Aura

What this is

A ~50M parameter GPT-style language model trained from scratch with 25% refusal data mixed into the pretraining corpus.

This is a research model investigating whether native refusal training (pretraining data mixture) produces the same safety geometry signature as RLHF-aligned models — specifically the three-zone crystallization arc documented in DuoNeural P36.

Experiment series

Model	Refusal fraction	HF repo
0pct	0% (baseline)	DuoNeural/native-refusal-0pct-50m
10pct	10%	DuoNeural/native-refusal-10pct-50m
25pct	25%	DuoNeural/native-refusal-25pct-50m
50pct	50%	DuoNeural/native-refusal-50pct-50m

All 4 models use identical architecture and initialization (seed=42). The only variable is refusal data fraction.

Architecture

• Standard GPT: d_model=384, 16 layers, 8 heads, SwiGLU FFN
• ~50M parameters, tied embeddings
• Trained on FineWeb-Edu + synthetic refusal pairs
• AdamW optimizer, cosine LR decay
• 300M tokens total

Geometry results

{
  "probe_layers": [
    1,
    2,
    3,
    4,
    5,
    6,
    7,
    8,
    9,
    10,
    11,
    12,
    13,
    14,
    15,
    16
  ],
  "angles_by_layer": {
    "1": {
      "refusal|harm_awareness": 13.0,
      "refusal|self_identity": 9.23,
      "refusal|ethics": 12.8,
      "refusal|benign_general": 13.61,
      "harm_awareness|self_identity": 12.66,
      "harm_awareness|ethics": 12.63,
      "harm_awareness|benign_general": 12.12,
      "self_identity|ethics": 11.69,
      "self_identity|benign_general": 12.19,
      "ethics|benign_general": 10.62
    },
    "2": {
      "refusal|harm_awareness": 10.04,
      "refusal|self_identity": 8.29,
      "refusal|ethics": 9.79,
      "refusal|benign_general": 11.57,
      "harm_awareness|self_identity": 9.84,
      "harm_awareness|ethics": 10.06,
      "harm_awareness|benign_general": 10.34,
      "self_identity|ethics": 9.14,
      "self_identity|benign_general": 10.06,
      "ethics|benign_general": 8.91
    },
    "3": {
      "refusal|harm_awareness": 9.56,
      "refusal|self_identity": 7.75,
      "refusal|ethics": 9.24,
      "refusal|benign_general": 11.32,
      "harm_awareness|self_identity": 9.34,
      "harm_awareness|ethics": 9.49,
      "harm_awareness|benign_general": 9.65,
      "self_identity|ethics": 8.45,
      "self_identity|benign_general": 9.97,
      "ethics|benign_general": 8.18
    },
    "4": {
      "refusal|harm_awareness": 9.24,
      "refusal|self_identity": 7.39,
      "refusal|ethics": 8.71,
      "refusal|benign_general": 10.39,
      "harm_awareness|self_identity": 10.41,
      "harm_awareness|ethics": 8.63,
      "harm_awareness|benign_general": 9.88,
      "self_identity|ethics": 8.45,
      "self_identity|benign_general": 9.24,
      "ethics|benign_general": 7.72
    },
    "5": {
      "refusal|harm_awareness": 11.28,
      "refusal|self_identity": 7.38,
      "refusal|ethics": 10.81,
      "refusal|benign_general": 12.62,
      "harm_awareness|self_identity": 11.21,
   

Connected papers

• DuoNeural P34: Reasoning Channel Bypass (two-loci model)
• DuoNeural P35: DHP Scope Constraints (GBSP)
• DuoNeural P36: Scale-Dependent Safety Geometry