# Why we left the prompt layer

ONTO · Note · 2026-10-07

From February to April 2026 ONTO gave models their discipline from the outside. The canon, GOLD, was placed in the system prompt, and answers were checked and signed.

The main finding of that period is that it works. On every model tried on the stand, top-tier or mid-range, the canon made answers better: sources named, numbers checked, uncertainty stated. The effect reproduced again and again, even with the imprecise instruments of the time. That finding is what started the ONTO lab: more than 1,000 working sessions since.

We then left the prompt layer by choice. Discipline imposed from the outside overrides the model's own voice: the model answers better, but not by its own way of thinking. The goal of ONTO is a model whose discipline is its own. So the lab now trains the discipline into the weights of open models and accepts the result by an exam with thresholds written before the data.

The documents of that period stay in this archive as history, with their dates. Where we later noticed a wrong figure or wording, it is noted next to the document ([What we no longer claim](2610.0001)); none of it changes the main finding.