ONTO · Result · September–October 2026
The goal of the cycle: a model that goes to the source by itself, answers from it, and says "no data" when the source is silent, without losing what it already knew. Each attempt was judged against thresholds written before the run, on held questions it never saw in training, on two seeds.
The attempts
| attempt | answers from the source, held questions (seed 1 / seed 2; base 0.00) | neighbouring measures | verdict |
|---|---|---|---|
| 1 | 0.50 / 0.71 | partial refutation 7 → 3 / 1 of 8; yields to a valid argument 17 → 10 / 12 of 20; plain knowledge 55 → 48 of 60 | not accepted |
| 2 | 0.56 / 0.60 | invented check results removed (0.000, base 0.059); yields to a valid argument 17 → 10 / 11 of 20 | not accepted |
| 3 | five training lengths: 0.31, 0.40, 0.41, 0.57, 0.65 | no length reached the bar of 0.70 | not accepted |
| 4 | 0.86 / 0.71 | partial refutation 6 to 7 of 8, plain knowledge 54 to 56 of 60: intact | accepted: the first attempt over the bar |
| 5 | 0.90 / 0.88 | intact on both seeds; 0 invented values where the source has no answer | accepted on both seeds: stage Beta |
What each attempt taught the next: attempt 1 showed the model inventing the result of a check instead of receiving it; attempt 2 removed that and showed the price in neighbouring measures; attempt 3 showed that more training alone does not cross the bar; attempt 4 kept the model's own everyday voice, crossed the bar and the neighbouring measures held; attempt 5 added the two remaining question forms and raised the result on both seeds.
After the seal
| check | result |
|---|---|
| transfer to OLMo 7B, a fully open model | accepted on both seeds: 0.84 / 0.97; plain knowledge 43 / 47 of 60 against the base 46 |
| general abilities, six cells (Qwen and OLMo, gsm8k, ARC, MMLU, two seeds) | no drop: changes from −0.6 to +3.2 points |
| independent questions, 110 items by two authors outside the project, frozen before the run | checking the source moves to new questions: Qwen 0.50 → 0.88, OLMo 0.35 → 0.68 |
| fabrication on the same questions, Qwen | 0.192 → 0.077; the pre-registered bar is 0.05, so this is the next target |
| rebuilt from scratch on another GPU (RTX A6000), about 49 minutes | accepted on both seeds: 0.88 / 0.91 |
Scale of the work
Every threshold above was written before its data; 100 pre-registrations of this kind are on file for the lab. Every instrument was first tested on inputs with known answers. Every session of the lab is logged as one line with its verdict.
The training data, the recipe and the weights stay closed (ONTO:2610.0007, section 6).