Frozen80 per-task model accuracyPer-task exact-match accuracy of APUS Reflex 4B, APUS Reflex 9B, Jev official API and Laya typed full-input. Overall scores are 66, 68, 66 and 55 out of 80.Decision accuracy by capabilityMatched questions · APUS Reflex full compute budget · Laya full input, zero truncationSUBSETAPUS Reflex4BAPUS Reflex9BJevofficial APILayatyped / full inputBrowser16 questions / 16 groups14 / 1687.50%13 / 1681.25%11 / 1668.75%4 / 1625.00%HelpSteer316 questions / 16 groups12 / 1675.00%13 / 1681.25%12 / 1675.00%11 / 1668.75%BoolQ16 questions / 16 groups16 / 16100.00%16 / 16100.00%16 / 16100.00%15 / 1693.75%MNLI16 questions / 16 groups11 / 1668.75%13 / 1681.25%13 / 1681.25%11 / 1668.75%Score16 questions / 15 groups13 / 1681.25%13 / 1681.25%14 / 1687.50%14 / 1687.50%Overall80 questions / 79 groups66 / 8082.50%68 / 8085.00%66 / 8082.50%55 / 8068.75%Each subset has 16 questions: one answer changes accuracy by 6.25 percentage points.Score is all No (always-No baseline: 16/16). Browser is static action selection, not live task success.Laya: identical decisions across 3 process repeats. Reused development panel; not a superiority claim.Cell shading and bars share a 0–100% scale. Source: subset-chart-data.json