Corpora

Corpora designed over the archive

Training corpora designed as SQL over the Scry archive — measured in place, time-stamped, source-attributed, and reproducible by query. Nothing is copied to describe it.

Designs

Pretraining Corpus 1 · 20T tokens · designed · 2026.08.21 A 20-trillion-token pretraining mixture designed as SQL over public Scry surfaces: web text, public discourse, scholarly full text, the public record, code, and reference — every document time-stamped and source-attributed. Designed and measured in place; nothing is copied.

How a corpus is stated

A corpus page names each slice by role, the public surfaces it reads, its raw mass as measured from engine metadata, and its designed retained size and weight. Every slice is a predicate over public SQL surfaces, so the design is reproducible by query. Measured numbers and design targets are labeled apart, and the page carries the date its denominators were read.