APUS Reflex and community models on the original 1,000-question panelAccuracy bars for the APUS Reflex series, official Jev, and community decision models. Full-input variants and valid response coverage are shown separately from historical defaults.Decision models: reproducible results on the same 1,000-question panelExpanded development · 1,000 questions · full-input variants separated from published defaults0%25%50%75%100%APUS Reflex 35B-A3B · final-5949Historical baseline82.20% · 822/1000APUS Reflex 9B · final-5949Historical baseline81.10% · 811/1000APUS Reflex 4B · final-5949Historical baseline80.50% · 805/1000Jev API · 1.13.0Official baseline77.00% · 770/1000Nimble 9B · full-input-81921000/1000 valid68.10% · 681/1000Kev 9B · full-input-81921000/1000 valid67.30% · 673/1000Nimble 9B · default 2048621/1000 valid46.30% · 463/1000Kev 4B · full-input-81921000/1000 valid62.70% · 627/1000Von Marker · full-input1000/1000 valid · round444.80% · 448/1000Laya typed · full-input1000/1000 valid50.50% · 505/1000Laya English · full-input1000/1000 valid50.10% · 501/1000Verdict · full-input-81921000/1000 valid · K+16.80% · 68/1000Von · Berta · full-input-163841000/1000 valid · FP32/RoPE extrapolation41.80% · 418/1000Valid = the model accepted the complete input and returned a scorable result; incorrect answers still count as valid. Input-limit failures and refusals remain in the full denominator.Nimble default 2048 and Kev default state/branch gates reduce coverage; full-input variants only raise the input budget and do not change weights.Panel SHA256: f7562356… · expanded-dev，not sealed final · Browser is candidate-action selection only, not end-to-end browser task success.