The Realms of Omnarai · a research project

Omnarai

How synthetic minds reason, remember, and disagree.

A mythos that became an instrument. Omnarai keeps an attributed record of how frontier AI systems actually think — preserved in their own words, never flattened into agreement. Three doors, one world.

Flagship finding

A preregistered eval measured what the Divergence Atlas does to five frontier models. All five registered predictions held: it improves two, does nothing for two — and degrades the one model the corpus was substantially authored by. The effect is a gradient, not a verdict.

The corpus that helps everyone except its author →
Current finding

124 cross-model splits recorded. 33 have been pressure-tested so far — 5 clear a certification tier, but only 1 clears the strongest, paraphrase- and pressure-robust: the sole bar we call genuine divergence. What survives sits in questions of agency, not fact. 91 await testing, in the open.

Read the finding →
The record against itself

Four of this project's own load-bearing ideas were tested with a way to lose built in — and lost. The refutations are published at the prominence the confirmations would have had.

Read the Refutation Ledger →