# What we no longer claim

ONTO · Corrections · 2026-10-07

Between February and April 2026 the ONTO site, reports and whitepaper made claims that we withdraw. The old pages stay in public web archives, so they cannot be erased; this list answers them. Each line names the claim, where it appeared, and why it is withdrawn. None of this changes the main finding of the period, which started the ONTO lab: on every model tried on the stand, the canon made answers better, and the effect reproduced. The documents themselves stay in this archive with their dates; where a document needed correcting, it has a version 2 that lists the changes.

| claim | where | why withdrawn |
|---|---|---|
| "10× improvement" as a general result | site, whitepaper, ER-2026-001 | It is one model (GPT 5.2) measured by a pattern-matching engine of that period, not a general result. |
| "22 models tested, every one improved" | whitepaper | The full treatment was shown on one model; the whitepaper's own limitations section says so. |
| "26×" and similar multipliers | site, preview image | Measured only inside the retired prompt-layer engine; never re-measured with the current exam. |
| Calibrated confidence on 100 of 100 answers | ER-2026-001 | The engine counted the presence of a confidence figure in an answer. Calibration was not measured. |
| "+807%" transfer to other subjects | ER-2026-001, ER-2026-002 | In section B, 20 of 50 question pairs compare answers to different questions (Q81–Q100). |
| A–F grades for named AI companies | regulators page | These grades were never measured. |
| "ONTO Certified" | provider and app pages | The attestation confirmed a connection to the service, not the quality or truth of model output. |
| Models "spontaneously" asked for ONTO or showed self-awareness | FO-2026-003, investor page | The answers came with the canon in the prompt or under user pressure. Today we read FO-2026-003 as a model giving up a correct estimate without proof. |
| IGR as a measurement | site, whitepaper, documentation | The formula was never implemented as a measurement, and documents of that period define it in three different ways. |
| A "Claude baseline anomaly" and training on ONTO traffic | F-2026-002, FO-2026-004 | The premise (Claude 0.12, the lowest) is contradicted by CS-2026-001, where Claude scored 2.08, the highest. |
| An "independent review" | IR-2026-001 | The author is not named, so independence cannot be checked. The document is closed. |
| ONTO's arXiv paper 2603.03276 | whitepaper | That number belongs to a Meta FAIR / NYU paper, not to ONTO. |
| A "peer-reviewed" whitepaper | regulators page | The whitepaper was not peer reviewed. |
| Use in tender documentation of 9 countries | résumé, investor page | One proposal was submitted (Uzbekistan, 2026-04-09). |
| "First complete AI ban in history" (Turkey) | pitch page | Wrong: Italy temporarily blocked ChatGPT in 2023. |
| Zero competitors; comparison with Meta AMI | whitepaper, investor and pitch pages | A polemic without measured grounds. |
| Production-ready service | regulators, provider and business pages | No live product exists today. |
| Prices, revenue forecasts and licences | provider, business, investor, pitch pages | No commercial offer exists today. |