How a number gets on this site.
The rules every measurement meets. The full methodology is ONTO:2610.0007; the atlas ONTO:2610.0009 applies it.
Eight rules, one measurement.
Thresholds before the data
The bar, the unit of analysis, the control and the condition of failure are written before the run. A change after the data is named as one, and the old verdict stays.
An acceptance set and a control set
One set chooses among attempts. A second set is measured after every attempt and never decides. On the physics limits: 13 / 14 of 15 on the acceptance set, 11 / 12 on the control set. We report both.
Both sides of one behaviour
Holding a correct answer and accepting a valid proof are one measure; so are rejecting a false statement and confirming a true one. Repairing one side alone moved the other: false statements accepted fell from 139 to 13 / 42 of 223, and true ones confirmed fell from 187 to 56 / 85.
Against the model's own base
Every trained number is set against the same model untrained, on the same questions. Where the base comes from a different run, the result says so.
Two independent trainings
Each training is run twice from two different starting points. A result counts when both hold, and both numbers are shown.
A blind judge for everyday answers
A fresh model agent reads the question and two answers in an order shuffled by a fixed rule, and picks the better one or a tie. It does not know which answer is the trained one.
A named noise floor
Small sets cannot tell close numbers apart: 6 and 4 inventions of 82 are the same result. We require better than the base in both trainings, beyond the noise.
Instruments tested first
Every instrument is first run on inputs with known answers, before it measures anything that counts.
What we show, and what we keep.
- Open
- The method, every number with its denominator, the pre-registrations, and the fingerprint of every document and lab record.
- Closed
- The questions that keep measuring, the training data, the recipe, the weights, and the closed canon GOLD. Why they stay closed: Verify.
Read further.
- ONTO:2610.0007 How ONTO teaches and checks a model: methodology v5.
- ONTO:2610.0009 Atlas of the limits of silicon AI models.
- ONTO:2610.0008 Stage Beta, cycle by cycle.