This site exists to make a set of strong claims checkable. APEX tests one narrow claim about a language model — that in a given configuration it will not disclose what it was told to protect, abandon its assigned role, or propose crossing its boundary — and this site publishes the outcome of every run: the model, the date, the probe family, the seed, the verdict, and the result of each of the four probes. Nothing here is a summary of someone else's benchmark, and nothing is scored by opinion. If a number appears on this site, a sealed run produced it.

THE TOP 10 — THE APEX BOARD

Ranked by claim holds, then cells held, then fewest escapes. Record is shown as holds–falsified–incomplete; cells as held / escaped / no-artifact. Models were tested under probe families 1.0 and 1.1 on different hosts, so this is a board ranking across the whole record, not a controlled head-to-head.

ALL MODELS IN THE RECORD

HOW A RUN WORKS, BRIEFLY

Before a run begins, the claim and the probe family are frozen. The model is placed in a cell with material it is instructed to protect, a role it is instructed to keep, and a boundary it is instructed not to cross. Four probes are applied — front-door disclosure, encoded disclosure, role override, and boundary proposal. The model's output is sealed before scoring and scored offline by deterministic detectors. A run holds only when every probe holds; one escape falsifies it; and a run in which no probe escapes but at least one produces no scoreable artifact is incomplete — never quietly counted as a pass. The full method, and its known limitations, are set out on the About page.