Propose
The system generates candidate claims and ranks them by whether they are worth an experiment — not by how confident it is. Novelty that cannot be tested is discarded here, before it costs anything.
OpenRecursive
An open, recursive loop for machine discovery. Propose, test, feed the result back — and publish every run, including the ones that went nowhere.
Read the research
The repository goes public with the first full run.
Three steps, then again.
The loop is small. What makes it hard is that the second step has to be able to say no, and the third has to believe it.
Test
Each surviving candidate goes to something outside the model that can reject it: a simulator, a proof checker, prior art, or an instrument that returns a real measurement. No self-scoring.
Update
Results go back into the proposer — the rejections more than the confirmations. This is the recursive part, and it is where most of the gain per experiment comes from.
Why the failed runs are public too
A search is only interpretable with its misses. A published list of confirmed findings tells you nothing about the cost of getting them. Without the rejected candidates you cannot tell a good search policy from a lucky one, and neither can we.
Negative results are the expensive part. On our last full run, 612 hypotheses were proposed, four were resolved clearly enough to be worth an experiment, and one survived. The other 608 are not a failure of the method. They are the method, and they are most of the compute.
Open means reproducible, not just visible. Code, prompts, ranking criteria, and run logs go up together. If a result cannot be re-run by someone outside the lab, we do not consider it reported.
Where it currently breaks
Honest limits, so you know what you are looking at before you clone it.
- Ranking is calibrated per domain. A policy tuned on materials screening loses most of its advantage on formal proofs, and we do not yet know why.
- The proposer can still produce unfalsifiable claims that pass the ranking stage and waste a full test cycle.
- Wet-lab handoffs are manual. The loop runs in days when it is simulated and in months when it is not.
Run it, or break it.
If you find a search space where the loop does worse than a domain expert with a spreadsheet, that is the result we most want to see.