ONTO
AboutTry the model
ONTO · Lab

What the lab is measuring now.

The current measurement, then each recent result and what it led to, newest first. Every result has its document.

Latest

Can the habit live in the weights of a fully open model?

From November 2026 the lab trains OLMo-2-1124-7B-Instruct to answer possible, impossible or not enough data, with no tool and no prompt. Each version is judged once, on a fresh closed set of 60 real claims whose key is fingerprinted before the run: frozen keys.

Wave 1155 of 60, untrained 28
Wave 1248 of 60, untrained 33
Wave 1350 of 60, untrained 32; the model live at /try
Transferfailed: on unseen topics, 10 of 21 false alarms

The size ladder of 2026-10-30 was stopped after the 14B rung (40 of 60, against 30.5 at 7B); 32B was not run. See the dated update in ONTO:2610.0009.

News

What was measured, and what it led to.

2026-10-30

The atlas of the limits of silicon AI models

Goal
One document with every limit the lab measured, on both sides.
Result
Eighteen limits on two open model families, each with its number, its denominator, its comparison and the fingerprint of the lab record behind it.
ONTO:2610.0009
2026-10-30

The smallest rung of the size ladder

Goal
Rehearse the whole chain on 3B before the larger models.
Measured
The same training as the 7B, two trainings, the same frozen tests.
Result
Physics limits as at 7B: 14 and 14 of 15 on set A. Multi-step everyday problems: 21.0 of 60, against 30.5 at 7B. General maths fell by 31 and 25 of 300, where the 7B lost 7 and 5.
Led to
The 14B rung.
Pre-registration
2026-10-30

Consulting the limits of physics: accepted

Goal
Put into the weights the habit of routing an impossible physical claim to the limit it breaks, without losing everyday answers.
Measured
Sets of 45 questions, 15 of them impossible; a control set that never decided between attempts; a blind judge on 40 everyday questions against the model before the training.
Result
Set A 13 and 14 of 15; control set 11 and 12 of 15; possible claims wrongly flagged 0 and 1 of 30. Everyday answers −6 / −7 net, inside the bar of −8. Four of the six attempts were not accepted.
Led to
The recipe was frozen and the size ladder was pre-registered.
ONTO:2610.0009
2026-10-30

Multi-step everyday arithmetic stays with the base

Goal
See whether the same training also improves everyday calculations of several steps.
Measured
60 new problems, each reference answer recomputed independently; the program the model writes and its words, both scored.
Result
By its program, the trained 7B solves 30.5 of 60 against 34 untrained. By its words it is 3 above its base: when the program is wrong, the words often recompute it.
Led to
The size ladder measures this first, as its primary measure.
ONTO:2610.0009
2026-10-16

Checking its own words

Goal
A model that checks its own answers against the limits.
Measured
Three fresh sets of 48 questions; untrained Qwen 7B answered, then wrote checks of its own answers.
Result
When a check gave a verdict, it was right 34 of 38 times. It caught 11 of its own 45 impossible claims. The pre-registered bar failed.
Led to
The line was closed; the boundary went into the weights instead.
ONTO:2610.0009
2026-10-12

Instructions in the prompt against training

Goal
Find how much of the habit a prompt alone can buy, and whether size changes it.
Measured
Untrained Qwen 7B, 14B and 32B with the same instructions and manual in the prompt, two passes each.
Result
Going to the source by itself: 0.42, 0.39 and 0.27. The trained 7B, with no instructions: 0.90. The instructions also cost true answers at every size.
Led to
The habit is trained, not prompted. A trained size ladder.
ONTO:2610.0009
2026-10-12

Refusal without a source

Goal
Teach the model to say when a number is missing, with no source to consult.
Measured
840 public maths problems with a key number missing, and 840 solvable twins.
Result
Invention fell from 501 to 49 and 64. Correct answers on the solvable twins fell by 18.3 and 14.3 points. At this price it is not usable.
Led to
A refusal is taught only together with the twin that must be answered.
ONTO:2610.0009
History

Before October 12.

2026-10

Invented answers on questions by outside authors

Result
Invented answers per 100, Qwen 7B: base 19.2 → trained 7.7. A direction, not a result: on a set this small such numbers cannot be told apart (see the noise floor, atlas row 18).
ONTO:2610.0008
2026-10

The first trained stage

Result
Answers taken from the source on 103 held questions: 0.00 → 0.90 / 0.88. The same procedure on OLMo 7B: 0.84 / 0.97. Rebuilt from scratch on another GPU: 0.88 / 0.91.
ONTO:2610.0008

Every earlier result, dated and fingerprinted: Archive →