---
base_model: unsloth/gemma-4-E4B-it
tags: [virgil, meeting-synthesis, peft-lora]
---
# virgil-meeting-e4b-a_v49
Shipping pair: `adapters/primary` = A_v43-extract (unchanged from a_v43 — ExtractBench
F1 0.54 vs base 0.49, P(beats base)=0.954, live register 0.04 -> 0.54). `adapters/fallback`
= A_v49-reconcile (minimax-m3 weak-dim distillation). Held-out reconcile bench (merged
serve, production guided decode) vs A_v40-reconcile: op_F1 0.778 -> 0.852, mutation
precision 0.216 -> 0.545, mutation recall 0.321 -> 0.859 (resolve precision 0.949).
Serve at MLX int8, unfused.