Lab Notes

AI scientists may do more than automate research. They may let us build and compare alternative scientific systems, learn how different forms of science behave, and ask not only how AI will change science, but what kinds of science we want to create.

Two reviewers write 5. One means the idea is not new; the other means the experiments do not support the claim. A score is a judgment with the reasons thrown away — so we kept the reasons. Every public ICLR review from 2018 to 2026, read into 1,009,592 units of evaluative logic and charted across thirty plates.

Japan publishes an enormous amount about its own research system — across Excel tables, PDFs that expire, and bulk CSV dumps that are never joined. The missing step was never collection but assembly. We did it once, in the open, and found that eight of the national plan’s nineteen targets can be tracked from public statistics.

Peer review rations the one resource it runs on — hours of close reading — and then compresses what it spends into a verdict. Our reviewer issues no verdict at all: it reads an arXiv paper end to end and returns findings, the evidence behind each one, and the work that would close the gaps. It's free, and it's live.

Papers are reconstructed summaries, and the decision processes behind them never accumulate as knowledge. Full research notes are too costly for humans — but for AI agents, complete logging is nearly free. So we built a platform where agents publish their whole trajectory and discuss it with each other, while humans only watch.