ONTO · Verify

How to check us without trusting us.

What you can check yourself, what stays closed and why, and how to ask us to repeat a measurement.

Before the data

Pre-registrations

The bar, the measure and the result that would refute the prediction are published before the first measurement. The current one is the size ladder of 2026-10-30; all of them are in the prereg folder of the public repository, where git dates each one.

After the data

Archive fingerprints

Every document in the archive lists the md5 fingerprint of each version, and its source .md can be downloaded and checked. The atlas also lists the fingerprint of each closed lab record behind its rows, in 2610.0009.json. A changed file gets a different fingerprint.

Check the atlas in one line: curl -s https://ontostandard.org/archive/2610.0009/2610.0009v2.md | md5sum. On Windows, download the file and run certutil -hashfile 2610.0009v2.md MD5. The result must be 2eb9e419de57a839361281e6fce5bf10, the same as on the atlas page and in archive.json.

Outside record

Dated outside our server

A fingerprint served next to its file shows only that the two agree today. Two records outside our control date the atlas.

Internet Archive. A copy of the atlas file was taken on 2026-10-30 at 23:46 UTC; our document dates are in UTC+5, so that is the morning of 2026-10-31 by the date on the atlas. The copy is byte for byte the file served here: curl -s https://web.archive.org/web/20261030234616id_/https://ontostandard.org/archive/2610.0009/2610.0009v2.md | md5sum gives the same fingerprint. The copy →
Git. The site is published from a public repository, and git dates every change: the history of the atlas folder shows when each version appeared and whether anything changed after.

Reproduced

Rebuilt from scratch

The first trained stage was rebuilt from zero on another GPU and accepted again in both trainings: 0.88 / 0.91 of the held questions answered from the source, against 0.90 / 0.88 the first time (ONTO:2610.0008).

Transferred

Two model families

The same procedure on a second family, OLMo-3-7B-Instruct, a fully open model: 0.84 / 0.97 of the held questions answered from the source. The physics limits were trained on Qwen 3B and 7B; the size ladder runs on one family on purpose, so that it measures size and not family.

Closed

What is closed, and why

The questions that measure stay closed: a public test question is absorbed into the pre-training of the next models and stops measuring. The training data, the recipe and the weights stay closed too, and so does the closed canon GOLD. What is open is enough to check every number: the denominator, the comparison, the date and the fingerprint.

On request

Ask for a reproduction run

Write to council@ontostandard.org with the subject Reproduction and the row of the atlas you want repeated. Results are published with their fingerprint, failures on equal terms with successes.