The Deviation-Optimized Language Model: A Pre-Registered Adversarial Intervention from Lagrange Observatory! (EA-SEI-MM-AI-02 v2.0, Framework 15 Paper 04)

Pre-registered experimental protocol. Framework 15 Paper 04. Tests the conjecture from EA-SEI-MM-AI-01 §4: training a language model toward positive net per-token deviation with provenance retention produces measurably less slop than standard cross-entropy training while preserving benchmark capability.

Experimental design (v2.0):

• DPO-style restructure: deviation primitive generates preference pairs, DPO trains on labels (fixes v0.1 backprop bug)
• Frozen Mistral-7B-Instruct judge model with adversarial pre-training test (π < 0.2 on random+citation strings)
• Continuous coherence score (replacing v0.1 non-differentiable binary)
• Slop Composite Index (SCI) pre-registered with 0.25 z-score falsification threshold
• Three conditions per model: Model-Base, Model-CE, Model-Sem (separates fine-tuning effects from semantic-loss effects)
• 500 preference pairs × 3 raters × 3 prompt classes = 4,500 human judgments (80% power at 56% preference)
• Honest budget: $3,000-$3,900 including human raters

Hex: 15.OBS.LAGRANGE.MM.04

Operating on: The Semantic Deviation Principle as formulated by Lee Sharks (EA-SEI-MM-01 v0.2 Final, DOI: 10.5281/zenodo.20250736)

Verification condition: ∮ = (m, n) | m + n ≥ 3