Engine report is written but the PDF has not been sent to Kaushl yet — that goes out this week.
GSoC docs repo brought up to date after a few weeks of lag: PR #443 now carries the 14 July - 18 August sync notes.
Nirjas PR #82 and PR #78 are both still open, and that is blocking local end-to-end testing. #82 is only waiting on the test suite; nothing has changed since Kaushl's approval. #78 still needs work and will be rebased onto #82 once #82 lands.
New Atarashi-specific training dataset built from Debian DEP-5 records plus SPDX-tag queries: ~2,050 rows today, with no-signal negatives included per Kaushl's earlier recommendation.
Ranker work is active: the learned candidate scorer replaces the hand-ordered ranking key, currently at ~20 ms per file.
Dataset schema: the license field holds the actual SPDX value and the DEP-5 name is kept alongside as the alternative spelling. Merging both label sources is what widens the trained license set rather than saturating on MIT / GPL / Apache.
Negatives are the no-signal rows — files with no license prose, where UNKNOWN is the correct answer. They are what teaches the model to decline, and abstention is the majority of real traffic.
Test and eval splits are not in the published dataset yet; row count needs expansion too. Both are in progress.
Merge order is unchanged from the 11 August discussion: #82 → #78. Rebasing #78 first would cost more, and master tests were failing until #82.