Measured: v1.4.0 once on Lite/Verified (function Acc@5 0.394 → 0.460 / 0.346 → 0.429); paper draft; session 19 brief

[ParsaVictor]

Docs and scoring only (no engine change). v1.4.0 release binary run once on SWE-bench Lite and Verified.

function level (all edited functions in top k)	n	v1.3.0 Acc@1/5/10	v1.4.0 Acc@1/5/10
Lite, LocAgent subset	274	0.168/0.394/0.482	0.223/0.460/0.540
Lite strict, with function gold	250	0.168/0.400/0.492	0.228/0.472/0.556
Verified, with function gold	459	0.163/0.346/0.416	0.198/0.429/0.497
Verified not in Lite	375	0.157/0.312/0.381	0.189/0.397/0.464

File level identical to v1.3.0 on every instance. Published (LocAgent Table 4, function Acc@5/@10): BM25 0.318/0.369, CodeRankEmbed 0.518/0.588 — the gap to the dense retriever halved.

Why the holdout gain (+0.066) is twice the dev prediction (+0.034): dev function gold is larger (single-function in 45% of dev issues vs 86% of Lite), not more tracebacks (20% / 20% / 14% of dev / Lite / Verified).

Also: paper_tables.py now rebuilds the dev and plain-language tables; paper draft has intro, related work, protocol, conclusion; docs/planning/15-next-session-brief.fa.md for session 19. v1.4.0 installed as dogfood binary (v1.3.0 backup kept).

🤖 Generated with Claude Code