← All posts
Blog archive
Everything we've published, most recent first.
202612 posts
August 2026
- The retrieval number we took back before we published itAug 23, 2026
- We trained the same model five more epochs. Then a clean split took the number back.Aug 17, 2026
- Harder training examples made our reranker worse. Here is the curve that fixed it.Aug 16, 2026
- We withdrew this reranker's +63.8%: it was measured on a split that leaked. The retrained model gains +47.2%.Aug 15, 2026
- We said our legal model was weak on statutes. Here is the one that is not.Aug 14, 2026
- Retrieval across 16 Indian languages, and the receipts to check itAug 12, 2026
- Don't take our word for it: read the tokens yourselfAug 11, 2026
- The judgment is public domain. The headnote is not.Aug 10, 2026
- MRR is not "average rank"Aug 10, 2026
- Two systems, identical recall, one of them uselessAug 9, 2026
- Recall@k is the ceiling on everything elseAug 9, 2026
- What DocPro search does when the embedding model is downAug 9, 2026