Weekly Sync - 21 July 2026

Attendees

• Swapnil Dutta
• Shaheem Azmal M MD
• Kaushlendra Pratap

Summary

• Shifted focus from the Nirjas gate (now on Hub + PR #78) to Atarashi evaluation design.
• Started work on Minerva PR #6 toward a real-corpus license-ID eval: queries from local the-stack-smol, ScanCode as independent ground truth, so scores are not “did we copy our own training fragments”.
• Nirjas PR #76 (Tree-Sitter) and PR #77 (poetry / ruff / pyright) remain open; gate PR stays stacked until those land.

Notes

• Headline metrics on the synthetic Atarashi split are not trustworthy for identification: sliding-window fragments of license texts will match those same texts in a reference index. The new eval is meant to measure identification, not copy-detection.
• Planned head-to-head: lexical cascade vs model2vec vs a contextual embedding baseline (Recall@1 / @5 / MRR).
• Short-form / SPDX-tag queries need their own regime: many files only have a tag, where UNKNOWN (or “not enough prose”) is the honest answer.

Action Items

• Land the real-corpus eval, de-leak the Atarashi export, and add ScanCode notice / short-form reference units.
• Follow up on #76 review so the Tree-Sitter path can merge.