Confirmed the Atarashi production path: lexical cascade + learned candidate ranker, not a 3,015-class classifier. rycerzes/atarashi-dataset stays useful as training material, with the circularity caveat on the card.
Started a non-circular eval/train corpus from Debian DEP-5debian/copyright records: source-file headers labelled by maintainers, with no matcher in the labelling loop. Hub publish of rycerzes/atarashi-dep5 follows this week (2,050 files / 544 packages; 671 notice vs 1,379 no-signal).
Nirjas stack unchanged pending review: PR #82 still open (review comments addressed 11 August); PR #78 still stacked on it. Minerva PR #6 still open with the packaged evals.
DEP-5 is the right shape for ranking: notice rows have license prose and an answerable SPDX-style label; no-signal rows should yield UNKNOWN. Only-vs-or-later is kept distinct (GPL-2.0-only vs GPL-2.0-or-later).
Label noise is real (~7% on a dual-detector disagreement sample): maintainers sometimes write GPL-2 for “or later”, and Files: globs can stamp a package license on a file whose header differs. Good enough to compare engines; not an absolute accuracy ceiling.
No Atarashi application PR yet — this stretch is dataset + eval contract so a ranker can be trained without scoring against its own sliding-window fragments.