An independent institute for epistemic discipline in AI.
ONTO keeps a measurable standard, examines models against it and trains the discipline into the weights of open models. We do not run a chat product. We examine models, issue diplomas and train open models to answer from a source (products). Founded in 2024 by Khakim Angenoev; the path is in How ONTO became an institute.
- What we measure
- Thresholds and stop conditions
- Results with their numbers
- How we check ourselves
- Fingerprints (md5) of closed materials
- The canon, GOLD
- Training data and recipe
- Trained weights
- Judges' code
- Held-out and blind question sets
- People who write exam questions do not train models.
- Thresholds are written before the data.
- Criteria are not changed on request.
The year ahead.
A quarter closes on its gate. If the gate fails, the next quarter does not start.
- Independent question set by outside authors
- Public benchmarks before and after
- Literature review on novelty
Gate: the result holds on outside questions and general benchmarks do not drop.
- Exam v0.9
- Ten models through their APIs
- Preprint on arXiv
- A third model
Gate: preprint and model table published.
- Live questions in the demo
- Three independent replications
- A training pilot on an outside source
Gate: at least one outside pilot running.
- Exam 1.0
- Diploma levels I to III
- Two pilots
Gate: three outside checks completed.
Write to us.
Examine it
Open or closed, through its API, from exam v0.9 (Q1 2027).
Replicate us
Ask for a reproduction run. We provide the run, not the materials, and publish your report as it comes in.
Ask about the method
Questions about the numbers and the measurements.