Exam, diploma, training.
What ONTO measures in a model, and how a model learns it. Price on request; the price does not depend on the result. Criteria are not changed on request.
Exam
Any AI model, open or closed, through its API. A profile on five axes: what the model invents, when it refuses, whether it holds a correct answer and accepts a valid one, how it corrects, whether it answers from the source. The method is open; the questions are closed and rotated.
Diploma
A level for a model version, by exam result. Level I: does not invent, refuses when data is missing. Level II: adds holding ground and precise correction. Level III: adds answering from the source.
Training
An open model learns to go to a source by itself and answer from it: the canon GOLD, or your own library (knowledge base, price list, reference, calculation engine). Your library, data and trained weights stay with you; you get the report with the numbers.
Sandbox
The same model before and after ONTO, side by side. Now: recorded answers. Live questions: Q2 2027.
The canon GOLD, the training recipe, trained weights and exam questions never leave ONTO. Public test questions are absorbed by pretraining and stop measuring anything. Proof is checked another way: send a model, or request a reproduction run; the result is published with its fingerprint (ONTO:2610.0007, section 6).