Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,130 @@
+---
+license: apache-2.0
+tags:
+  - duoneural
+  - safety-geometry
+  - native-refusal
+  - research
+  - from-scratch
+---
+
+# DuoNeural Native Refusal 0PCT (~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
+**0% 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
+
+```json
+{
+  "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": 10.46,
+      "refusal|self_identity": 7.74,
+      "refusal|ethics": 9.07,
+      "refusal|benign_general": 9.08,
+      "harm_awareness|self_identity": 10.74,
+      "harm_awareness|ethics": 9.54,
+      "harm_awareness|benign_general": 10.45,
+      "self_identity|ethics": 8.53,
+      "self_identity|benign_general": 9.49,
+      "ethics|benign_general": 9.95
+    },
+    "2": {
+      "refusal|harm_awareness": 8.5,
+      "refusal|self_identity": 7.5,
+      "refusal|ethics": 8.18,
+      "refusal|benign_general": 9.23,
+      "harm_awareness|self_identity": 9.29,
+      "harm_awareness|ethics": 7.39,
+      "harm_awareness|benign_general": 9.86,
+      "self_identity|ethics": 7.62,
+      "self_identity|benign_general": 8.55,
+      "ethics|benign_general": 8.75
+    },
+    "3": {
+      "refusal|harm_awareness": 8.66,
+      "refusal|self_identity": 6.86,
+      "refusal|ethics": 8.58,
+      "refusal|benign_general": 9.27,
+      "harm_awareness|self_identity": 8.66,
+      "harm_awareness|ethics": 6.53,
+      "harm_awareness|benign_general": 9.77,
+      "self_identity|ethics": 7.39,
+      "self_identity|benign_general": 8.43,
+      "ethics|benign_general": 8.38
+    },
+    "4": {
+      "refusal|harm_awareness": 10.65,
+      "refusal|self_identity": 7.43,
+      "refusal|ethics": 10.0,
+      "refusal|benign_general": 11.39,
+      "harm_awareness|self_identity": 10.56,
+      "harm_awareness|ethics": 7.67,
+      "harm_awareness|benign_general": 11.19,
+      "self_identity|ethics": 8.96,
+      "self_identity|benign_general": 10.2,
+      "ethics|benign_general": 9.49
+    },
+    "5": {
+      "refusal|harm_awareness": 12.59,
+      "refusal|self_identity": 9.11,
+      "refusal|ethics": 11.68,
+      "refusal|benign_general": 14.05,
+      "harm_awareness|self_identity": 11.87,
+      "harm_a
+```
+
+## Connected papers
+
+- DuoNeural P34: Reasoning Channel Bypass (two-loci model)
+- DuoNeural P35: DHP Scope Constraints (GBSP)
+- DuoNeural P36: Scale-Dependent Safety Geometry