nDCG@10
+7.74%
0.2476 -> 0.2668
Held-out public evidence
Bounded Learned Reranker
A 41-feature linear model trained from pinned public traces, embedded in the binary with a deterministic fallback.
MRR@10
+10.12%
0.2009 -> 0.2212
Acceptance
PASS
Quality, per-task loss, and latency gates.
Per-task quality
| Task | Deterministic nDCG | Learned nDCG | nDCG delta | MRR delta |
|---|---|---|---|---|
| codetrans-dl | 0.2380 | 0.2785 | +0.0404 | +0.0359 |
| codetrans-contest | 0.4008 | 0.3937 | -0.0070 | -0.0033 |
| cosqa | 0.1521 | 0.1637 | +0.0116 | +0.0137 |
| codefeedback-st | 0.4077 | 0.4769 | +0.0692 | +0.0725 |
Latency and portability
Warm p50 delta: +15.04 ms. Warm p95 delta: -54.42 ms. Set IVYGREP_RERANKER=deterministic to retain the zero-model fallback.
Identity
Commit 76a404ec051cc9009433fe0ed021f51774b104ee
Binary 92f91629fa9a5dbad4f9f540e35703e27cc188776a763c7818cc06ee7b619a61
Model public-linear-reranker-v2