GemmaForge · heldout benchmark

spanmax · F1-max τ

loading repo index…

pick a repo on the left

136 SVEN-derived microrepos, each pre-scanned with the spanmax probe (layer 8, paired with the layer-8 per-CWE head).

ground-truth vuln span (SVEN diff oracle) + probe lead (windows scored ≥ F1-max τ). Sub-τ leads listed in the bottom table (greyed out) but not painted on the code.