Research brief directory

Each brief answers one question the studio put to its own machines or to the literature. Every brief follows the same shape:

• The question, and why the studio asked it.
• What we found, in plain words, with the numbers.
• Charts, built only from committed receipts by experiments/charts/make_charts.py.
• Caveats: what the numbers can’t be used to claim.
• Receipts: the repository, commit and path behind every figure.
• Shelf status: whether the finding has been promoted to a readouts knowledge base yet.

Status labels mean the same thing everywhere:

label	meaning
measured	run on our machines; receipts committed; not independently checked
confirmed	a pre-registered check passed, such as a held-out split
interim	partial, or waiting on a run or a licence
disproved	tested and found wrong; kept so nobody rediscovers it
research	from published sources, not measured here

Pre-registered and post hoc results are labelled on every chart. A post hoc result is one we went looking for after seeing the data; it’s useful, but weaker.

Start from the newsletter for what changed lately, or pick a brief from the sidebar.