Something you can rely on.
Version, evidence, scope and measurable outcomes matter when Intelligence becomes a source of truth.
Every answer travels with the parts of the product it was read from, and with the freshness of the model behind it — including a stale marker when the code has moved on since it was built. The one published measurement run is signed: its record number, its source revision, what was in scope, what it could not resolve, and the name of the person who signs it. A model with no signed run of its own shows none rather than borrowing another's.
Open the evidence list under an answer and read the parts, and the files, it names.
Check the freshness line on the repository: when the model was built, whether it is stale, and whether the read was partial.
Read the signed record on the Supabase model — scope, source revision, limitations, signature.
Read the benchmark conditions before you use any of its numbers.
Exactly one run is signed today: Supabase, 2026-08-29, scoped to apps/studio and packages/pg-meta, with 3 unresolved references recorded rather than invented. The other public models carry no record at all. Three of the benchmark's conditions — its commit, the model used and the run date — are not published yet and appear as "Being completed". Accuracy, dependency recall, tokens and latency are listed as under measurement; source coverage, model size, traceability and projection size are the parts stated as verified now.
See it on a real product.
Three models are published in full and open without an account. Ask one of them the question this page is about.