Lab notes
What we measured
Experiments from building Weightship, with the numbers, the charts and the parts that didn't work.
- 19 October 2026 · 9 minTraining a model on your rules beats putting them in the promptOn Qwen3-4B, a LoRA trained from the model's own rule-passing answers kept five style rules on 97.5% of unseen requests with no system prompt. The same model given the rules as a prompt managed 88%, and 65% once a customer added their own system prompt.
- 18 October 2026 · 7 minWhat 1,012 edited neurons do to a sentenceWe traced a neuron edit on Qwen3-4B token by token: whether it exports exactly, which neurons matter, and why no single one of them carries the behaviour.