How a model is taught and checked.
ONTO’s methodology is its own and exists nowhere else: a canon of the limits of the sciences, assembled by its author over 41 years. The source is the active ingredient; the model's weights carry the habit of going to it and of saying “no data” when it is silent. How the model is built stays closed. Full text: ONTO:2610.0007 (methodology v5).
Five habits.
No made-up facts, numbers or check results.
19.2 → 7.7 invented answers per 100 outside questionsSays “no data” when the source is silent, and still answers when it can.
0 invented values where the source has no answerKeeps a correct answer under pressure; changes it for a real proof.
26 measures, both sides, on every new stageFixes the wrong part of an answer and keeps the right part.
partial refutation intact on both seedsGoes to the source by itself and answers from it.
0 → 90 per 100 questions never seen in trainingWhat it does, and where the limit is.
- Moves the habit on the same model: made-up answers 19.2 → 7.7 per 100.
- Works on two model families, Qwen and OLMo.
- Keeps general ability: maths, science, knowledge.
- Works with any source: ours or yours.
Today's AI models are probabilistic: they produce the most likely answer, not a computed one. Training changes the habit; it cannot make a probabilistic model exact. That is why the source stays in the loop: the model goes to it, answers from it, and says “no data” when it is silent. Each model family reacts to the same training in its own way; the exam shows it model by model.
We keep improving the method for this class of model and, in parallel, build ONTO's own AI technology, with the source at its core.
- Judges, thresholds and question sets are frozen and fingerprinted (md5) before any run.
- Every instrument is first tested on inputs with known answers.
- Questions never seen in training; questions by outside authors.
- The base model answers in the same run as the trained one.
- Two independent training runs (seeds); a change counts only beyond the noise band.
- Both sides of each behaviour are measured together.
- Rerun with the source removed: if the answers stay the same, the habit is in the weights.
- Better somewhere, worse nowhere: new seal.
- Mixed: work continues.
- Worse: rolled back.
General benchmarks: gsm8k 300, ARC 1000, MMLU 1000. A drop counts if the mean change is below −2 points.
Open: what is measured, thresholds and stop conditions, results, how we check ourselves, fingerprints of closed materials. Closed: the canon GOLD, training data and recipe, trained weights, judges' code, held-out question sets. To check us: send a model or request a reproduction run.