Weekly Sync - 14 July 2026

Attendees

• Swapnil Dutta
• Shaheem Azmal M MD
• Kaushlendra Pratap

Summary

• Midterm checkpoint week: Nirjas gate training and Hub deployment are now public, not just a local experiment.
• Published the trained gate to Hugging Face: rycerzes/nirjas-gate (potion-base-32M / model2vec, ~32M params, torch-free inference via skops).
• Opened Nirjas PR #78: train-and-release pipeline, nirjas[gate] extra, CI workflow, and Hub download at inference time. Stacks on #76 (Tree-Sitter) and #77 (poetry / ruff / pyright).
• Training data is already on the Hub as rycerzes/nirjas-dataset (published 6 July): binary gate labels, not a broad “license-related” detector.

Notes

• Operating point is recall-first at threshold 0.20 (not argmax): missing a license is the costly error; a false positive only costs one extra Atarashi call.
• Held-out test: license recall 0.9952, FPR 0.0262. Real-corpus check (the-stack-smol + ScanCode): recall 1.0000, FPR 0.0133.
• Gate is not yet wired into the Nirjas CLI / extraction flow — that stays a follow-up after #76/#77 land.

Action Items

• Keep #78 in review while #76 (Tree-Sitter) and #77 (tooling) move first.
• Start Atarashi Phase 3: a non-circular license-ID eval on real source, not only sliding-window fragments of the same reference texts.