Method

An entry goes on the ledger only after we have read the claim on the page where it was first made.

What counts as a source

The lab's own post, the paper, the code repository, or the person's own post. Press coverage and posts on X are leads: we use them to find claims, never as the source of one. When a page shows no date, we say so and mark any date we took from press with a †.

Quotes

Quotes are copied from the source page. When we check a claim we keep a text copy of the page, and the site will not build if a quote does not appear in that copy word for word.

Evidence levels

Paper or code

A paper or public code lets others check the result.

Official post with figures

The lab gives numbers but no way to reproduce them.

Statement only

A position or plan, with no result attached.

A claim that exists only in someone else's retelling does not go on the ledger.

Corrections

Many claims circulate in a changed form. Where the circulating version differs from the source, the entry says how: ▢ contradicted the source says otherwise; ○ no primary source we could not find it anywhere first-hand; ◩ left out the source says something the retellings drop.

What we don't do

We don't rank or score models. Vals, Scale and Hillclimb run RSI benchmarks; see Benchmarks for what each measures.

Who runs this

Zentor, the team behind Personal Agent Bench.