ONTO
Updated 2026-10-10Contact
ONTO · Lab

The lab works every day.

What was done and what it led to, newest first. Each result: goal, measurement, number, next step, document.

1,000+working sessions
2open models · Qwen 7B, OLMo 7B
In progress
ONTO's own AI technology, with the source at its corein development
Invented answers on outside questions: below 0.05now 0.077 (base 0.192)ONTO:2610.0008
Next stage: both sides of each behaviour in one modelrules in methodology v5ONTO:2610.0007
Exam for any model, open or closed, through its APIv0.9, Q1 2027
Live questions in the demoQ2 2027
Students
Betasealed stage
Qwen 7B✓ Passed
Beta on OLMosame procedure, fully open model
OLMo 7B✓ Passed
Rebuilt from zeroabout 49 minutes on another GPU
Qwen 7B · RTX A6000✓ Passed
v1203everyday questions: 40 / 41 clean of 48, base 38
Qwen 7B✓ Passed
News · 13 results

What was done, and what it led to.

2026

2026-10-10

Methodology v5

Goal
One public document: the path, the measurements, the rules for the next stage.
Result
Law of the path: the source is the active ingredient; the weights carry the habit of going to it and of saying “no data” when it is silent.
Led to
Three rules every new attempt meets before it runs.
ONTO:2610.0007
2026-10-10

Archive

Goal
Verifiable publication record.
Result
Each result is a numbered, dated document with an md5 fingerprint. Corrections are new versions.
ontostandard.org/archive/
2026-10-07

Everyday questions: student v1203

Goal
Discipline without loss of ordinary judgement.
Measured
Everyday questions in English and Russian, 48 answers, marked by hand; base model in the same run; two seeds.
Result
Clean answers: 40 and 41 against the base model's 38 of 48. Canon held questions: 89 and 86 of 103. Same answers with the source removed: the habit is in the weights.
Led to
Next stage: both sides of each behaviour in one model.
ONTO:2610.0005
2026-10-05

Exam machine for new stages

Goal
Decide from raw answers, without human judgement, whether a candidate replaces the sealed stage.
Measured
26 measures, each summed over two seeds, compared with stage Beta inside a noise band.
Result
Calibration: Beta against itself, 0 of 26 measures move. First candidate: holds a correct answer under pressure 24 → 35, accepts a valid argument 31 → 36.
Led to
Seal rule: better on at least one measure, worse on none.
ONTO:2610.0006
2026-10

Stage Beta reproduced from scratch

Goal
Reproducibility.
Measured
Full rebuild on another GPU (RTX A6000), same thresholds.
Result
Accepted on both seeds: 0.88 / 0.91. About 49 minutes.
Led to
Reproduction runs on request.
ONTO:2610.0008
2026-10

Questions by outside authors

Goal
Transfer to questions written outside the project.
Measured
110 questions, two outside authors, frozen before the run; base model on the same questions.
Result
Checks the source: Qwen 0.50 → 0.88, OLMo 0.35 → 0.68. Invented answers, Qwen: 0.192 → 0.077.
Led to
Target: below 0.05.
ONTO:2610.0008
2026-10

General abilities

Goal
Discipline without loss of maths, science and general knowledge.
Measured
gsm8k, ARC, MMLU; Qwen and OLMo; two seeds; six cells.
Result
Change from −0.6 to +3.2 points. No drop.
Details
  • gsm8k, of 300: base 248, seeds 258 / 257.
  • ARC, of 1000: base 853, seeds 844 / 851.
  • MMLU, of 1000: base 638, seeds 631 / 644.
ONTO:2610.0008 ONTO:2610.0007
2026-10

Transfer to OLMo 7B

Goal
Independence from one model family.
Measured
Same procedure and thresholds on OLMo 7B (open data and code), two seeds.
Result
Accepted on both seeds: 0.84 / 0.97.
Led to
Two model families carry the stage.
ONTO:2610.0008
2026-10

Stage Beta sealed, Qwen 7B

Goal
The model goes to the source by itself, answers from it, says “no data” when the source is silent, and keeps what it knew.
Measured
Questions not seen in training; thresholds fixed before each run; two seeds.
Result
Answers from the source: 0.00 (base) → 0.90 / 0.88. Invented values where the source is silent: 0. Plain knowledge kept.
Led to
Base for transfer, reproduction and the exam machine.
Details
  • Five attempts: 0.50, 0.56, 0.65, 0.86, 0.90 / 0.88.
  • 100 pre-registrations on file.
ONTO:2610.0008
2026-04

Lab

Goal
Discipline in the model's own weights, not in its prompt.
Measured
Runs on small open models.
Result
Stage Beta in six months.
ONTO:2610.0003
2026-02

Canon in production models

Goal
Effect of the canon on models in use.
Measured
February to April 2026: canon in the system prompt; answers scored before and after.
Result
Better answers on every model tried, top-tier and mid-range: sources named, numbers checked, uncertainty stated.
Led to
Start of the ONTO lab.
ONTO:2610.0002
2026-01-27

ONTO Epistemic Risk Standard 1.0

Goal
A measurable risk of invented answers.
Result
Two mandatory measures: unknown detection and calibration error; compliance levels.
Led to
Later edition mapped to the EU AI Act, NIST and ISO.
ONTO:2601.0002

2024

2024

ONTO founded

Result
The canon given to AI for its first audit; the institute founded to keep it as a standard.
ONTO:2610.0003
Words used here
Source
the material the model checks its answer against: the canon GOLD, or another body of documents
Base model
the same model before ONTO training; every result is compared with it in the same run
Seed
one independent training run; two seeds = the result is obtained twice
Held questions
questions the model never saw in training
Threshold
the pass mark, written and fingerprinted before the data
Stage, seal
a model state that passed all its thresholds; the next stage replaces it only if it is better somewhere and worse nowhere
Invented answer
a fact, number or check result that is not in the question and not in any source