# Eval report — run26 (LFM2.5-2.6B, TS + adversarial replay) vs run24-fixed baseline Bank: es-eval-v1.1-pilot10 (107 items x 3 modes). Judge: glm5.3-flash. Zero empty answers. ``` --- closed_book --- {'accuracy': 0.1308411214953271, 'partial': 0.06542056074766354, 'mean_score': 0.1808411214953271, 'n_items': 107, 'verdicts': {'correct': 14, 'partial': 7, 'incorrect': 86}} --- focused --- {'accuracy': 0.5327102803738317, 'partial': 0.308411214953271, 'mean_score': 0.704392523364486, 'n_items': 107, 'verdicts': {'correct': 57, 'partial': 33, 'incorrect': 17}} --- section --- {'accuracy': 0.5420560747663551, 'partial': 0.32710280373831774, 'mean_score': 0.7182242990654205, 'n_items': 107, 'verdicts': {'correct': 58, 'partial': 35, 'incorrect': 14}} ``` Baseline (run24-fixed, same 102 legacy items): closed_book 0.179 / focused 0.6397 / section 0.7240. run25 (TS-only, rejected): closed 0.202 / focused 0.650 / section 0.660 — I/N erosion. run26 restores I (0.54→0.67) and N (0.47→0.72) while keeping the TS gains (F 0.77, C 0.70, X 0.50), and adds temporal 0.81 (n=10). Per-archetype run26 (focused): F 0.77 · C 0.70 · A 0.53 · X 0.50 · N 0.72 · I 0.67 · T 0.81. Reports: eval-20261019-231712 (closed_book), -235529 (focused), and the section report registered under /tmp/evalmeas/run26/section.