meta-science
An agent that does science on worlds it has never seen — forming hypotheses, designing its own experiments, and being refuted by them — and that improves its own method only when a frozen benchmark proves the improvement real.
The world under study
Opaque labels, two affordances, no structure — everything below runs against exactly this view. The truth exists, but it is revealed, never used: full inspector →
1 · Run the discovery loop
Each prediction is committed before its experiment. The verdict is a comparison, never a question put to the model.
2 · Let it improve its own method
Gemini proposes against 24 held-out worlds — takes a
minute. The benchmark is always the frozen real-domain set, whatever the controls above
say: history stays comparable.
3 · The evidence base
loading…
Every run here is stored with its seeds, strategy and code commit — enough to recompute the result rather than merely read it.