Automating the search for things nobody bothered to ask.

Most interesting questions are not unanswered because they are hard. They are unanswered because nobody with the right background ever sat down and wrote a mechanistic guess about them. The gap between "someone wondered about this" and "someone wrote down a falsifiable claim" is where almost everything gets lost.

Curio is a continuous, unattended attempt to close a little of that gap in whatever direction its curiosity runs. Each cycle takes one question — animal biology, physics oddities, unsolved mathematics, materials, history — and works it through five reasoning passes. The output is not an answer. It is a testable increment: a claim specific enough that someone can say what experiment or calculation would kill it.

Everything published here is unreviewed model output, generated for human review. It is not peer-reviewed, not authoritative, and not advice of any kind.

A run is deliberately narrow. The model receives the question and nothing else, and must produce five stages in order — each one reading everything that came before it.

Findings feed forward. Each synthesis ends by proposing two follow-up questions, and those enter the queue — so the question set grows along the model's own curiosity rather than down a fixed list.

The hard problem with an autonomous hypothesis generator is telling reasoning apart from recitation. A model that has read the literature can produce a confident, correct-sounding paragraph without deriving anything.

So a solved question sits in the set as a control — one where the mechanism is textbook and the answer is checkable. When the engine reaches it, the trace can be read against a known result: did it reconstruct the argument, or restate the answer it already knew? The control validates no individual hypothesis. It calibrates how much weight to put on the ones with nothing to check against.

The engine runs one question per cycle and broadcasts the whole trace as it is generated — the model's thinking first, then the stage output, token by token. Every visitor is watching the same cycle at the same moment; nothing is edited, curated, or retried.

The engine is not pointed at one field. Click an area to see the questions inside it and what is being worked on right now.

A Solana memecoin is attached to the engine — a market to judge the thing by while it works. It does not fund, gate, or steer the research; every finding stays public and free to download.