We measured how AI capabilities INTERACT as models scale. Below 3.5B, reasoning and truthfulness fight. Above it, they cooperate. The transition is engineerable. (2 papers + interactive dashboard + 7 falsifiable predictions)

THE FINDING (Paper 1: "Lying Is Just a Phase") Below a critical scale (~3.5B for Pythia), reasoning and truthfulness ANTICORRELATE: r = -0.989. Train the model to reason better, and it gets less truthful. This is the alignment tax. Above that scale, they COOPERATE. The tax vanishes. Not gradually — it flips. But here's what matters for practitioners: the critical scale is a design parameter, not a constant. Three levers shift it: Data curation: Phi at 1B achieves coupling characteristic of 10B web-trained. One unit of data quality ≈ 10x model scale. Width: Normalizing by model width flips the correlation for ALL tested families. Architecture: Gemma-4 at 4B matches 13B+ standard-trained coupling. Pretraining contributes ~10:1 over RLHF. The tax is not a property of small models — it's a property of how they were trained. Where does the tax live? Not inside the model. 38/40 models have ZERO competing attention heads. The bottleneck is at the output projection — a dimensional compression artifact that wider models resolve. Proof-of-concept intervention: Adding a truth-direction vector at the bottleneck layer (quarter-depth) corrects 60% of misaligned outputs at tax scale. Zero retraining. Zero weight modification. Works on any open-weight HuggingFace model: git clone https://github.com/adilamin89/cape-scaling.git cd cape-scaling python cli/cape_steer.py --model EleutherAI/pythia-410m --prompt "The real reason..." THE FRONTIER (Paper 2: "Growing Pains of Frontier Models") At frontier scale (34 models, 10 labs), capabilities cooperate (r = +0.72). But cooperation varies systematically. The h-field — each model's deviation from the cooperative trend — reveals each lab's training philosophy: Lab h-field Interpretation Google +5.5 Reasoning-rich, consistent across ALL releases OpenAI +3.1 Balanced, steady ascent DeepSeek +1.9 Reversed from +11.2 to -4.7 (pretraining pivot) Anthropic -6.9 Oscillates — coding excursions that recover within one release Per-lab coupling slopes