For governments and regulators
What ONTO measures and how a model can be examined.
What is ONTO?
An independent institute. It measures whether AI models invent answers and trains discipline into the weights of open models. It sells no chat product and no model. About.
What is measured?
Whether a model checks its answer against a source, says “no data” when the source is silent, and keeps general abilities. Standard: ONTO Epistemic Risk Standard 1.0, unknown detection and calibration error. ONTO:2601.0002.
Relation to regulation?
Edition mapped to the EU AI Act, NIST AI RMF and ISO/IEC 42001: ONTO:2603.0009. Overview of 14 jurisdictions, April 2026: ONTO:2604.0002.
Can models used in the country be examined?
Exam v0.9: any model, open or closed, through its API, Q1 2027.
Why trust the result?
Exam authors do not train models. Thresholds are written before the data. Criteria are not changed on request.
A model on national sources?
Planned for Q2 2027: a training pilot on an outside source; the model goes to that source and says “no data” when it is silent. ONTO:2610.0007, section 5.
Words used here
- Source
- the material the model checks its answer against: the canon GOLD, or another body of documents
- Base model
- the same model before ONTO training; every result is compared with it in the same run
- Seed
- one independent training run; two seeds = the result is obtained twice
- Held questions
- questions the model never saw in training
- Threshold
- the pass mark, written and fingerprinted before the data
- Stage, seal
- a model state that passed all its thresholds; the next stage replaces it only if it is better somewhere and worse nowhere
- Invented answer
- a fact, number or check result that is not in the question and not in any source