pooled20

pooled20@asof-2026-09-26 — data stage report

Recorded 2026-10-13 as part of the OSS line's migration to the factory. The data already existed; nothing was rematerialized or read in bulk.

What it is. The twenty streams staged for pooled20_v1 on jon-1627: the eight record streams, ohmyzsh and the eleven corpus-lane repos with issues, 28,399,994 events by the runs' receipts; every pooled20 arm trained on byte-identical staged copies of these. A pool is its own source: its runs read it as one body of data, even where a member's bytes are also a single-repo dataset.

Output, by identity. 20 member spines, recorded where they are (modal://jon-1627/vei-oss-data/praxis/pooled20_v1/streams/<member>). Each member's digest is the one already on record (its single-repo dataset, or the sync manifest staged beside it); counts and span are copied from the single-repo dataset with the same digest. 11 members carry no counts, which were never taken: llvm__llvm-project, home-assistant__core, huggingface__transformers, neovim__neovim, ziglang__zig, python__cpython, grafana__grafana, microsoft__vscode, flutter__flutter, pytorch__pytorch, godotengine__godot; the pool's 28,399,994 events are the runs' receipts.

Drift. Members drawn from single-repo datasets carry their ingest-time checks there. A pool has no single calendar split (each stream is split at its own time quantile), so no split-time check was run; members without a single-repo dataset of their own have no drift record, a caveat on any result.