Make every parameter
earn its place.

The first phase built the foundation: a complete 125M-parameter model trained for 300,000 optimizer steps, with every layer of the stack owned end to end. It proved the training system; its low reasoning scores also proved that pretraining alone was not enough.

The current product architecture gives one Qwen3.5-9B family three roles: generate a complete internal draft, revise it with a trained later owner, and commit one coherent trajectory. It solves 383/538 protected product problems at 75.815% macro—67 more than the matched original second pass—without external proposals, tools, task routing, or correctness signals at inference. In parallel, the controlled DIVERGE line has qualified source-deleted plasticity, identity, command and value reading, law synthesis, execution, and operation binding. On Qwen3.6-35B-A3B, a compact temporal-causal gate now improves a source-disjoint screen from 111/256 to 143/256 while leaving native router and expert weights frozen. The original 125M scratch model has not demonstrated the product architecture's capability.