ONTO · Products
Three services, by application.
Each is provided under a written agreement. The terms depend on the model and the scope, not on the result.
Measurement of a model's limits
- What
- Where a language model stops, measured with the instruments of the atlas: both sides of each behaviour as one measure, against the model's own untrained base, on question sets that stay closed so that they keep measuring.
- You get
- A report in which each number has its denominator, its comparison and the fingerprint of the record that states it.
- For
- Providers of models, research labs, and teams that fine-tune a model for one field and need to know what it will and will not do.
Shifting the boundary into an open model's weights
- What
- On an open model, we train the habit of consulting a boundary before answering, and measure the result against the same model untrained, on both sides of each behaviour.
- What it moved
- On Qwen 7B, answers taken from the source rose from 0.00 to 0.90 / 0.88 of 103 held questions; impossible physical claims routed to their limit rose from 0 to 11 / 12 of 15 on the control set.
- What it cost
- Everyday answers, judged blind against the same model before the training, lost 6 / 7 of 40 net in the accepted training. Every result is reported with its price.
- What stays where
- The weights stay with you. The recipe and the closed canon GOLD stay with the institute.
Institute conclusion on technical claims
- What
- A written conclusion of the institute on whether a technical claim breaks a known limit, and by how much.
- When
- After a test on real documents. Until then it is not offered.
Write to us.
Write to council@ontostandard.org with the service in the subject, the model, and what you need to know about it. No prices are published; terms are agreed per application. Before you write, you can check how we measure: Method · Verify.