Skip to main content

Weekly Sync - 28 July 2026

Attendees

Summary

  • Pushed Atarashi eval work onto Minerva PR #6 (24 July):
    • Real-corpus license-ID eval (the-stack-smol queries, ScanCode labels).
    • De-leaked Atarashi export (output/atarashi_clean) plus short-fragment quarantine.
    • ScanCode notice / short-form / SPDX-tag reference layer, with an independent tag eval.
  • Nirjas PR #76 review: Kaushl flagged SLOC / “lines of comments” mismatch vs master on fossology/src/scheduler/agent/interface.c (comment detection matched or was better; the counts did not).

Notes

  • Root cause on master: binder.py double-counts block comments (+= len(output) on top of the span sum). Master 4/32 fixtures consistent with that inflation; Tree-Sitter path 32/32 against its own spans.
  • Also fixed two bugs on the Tree-Sitter side, with regressions: sloc double-subtracted blanks inside comment blocks, and Kotlin/Swift block comments misclassified as single-line ("multiline_comment" contains "line_comment").
  • Atarashi eval direction: keep lexical and embedding tracks comparable, but treat tag-only files as a separate regime rather than mixing them into full-text Recall@1.

Action Items

  • Finish #76 count-contract fixes and get another review pass.
  • Keep the Atarashi eval suite growing (more queries, regime splits) once #76 is unblocked.