| <svg xmlns="http://www.w3.org/2000/svg" width="1280" height="1100" viewBox="0 0 1280 1100" role="img" aria-labelledby="accuracy-final-title accuracy-final-desc"><title id="accuracy-final-title">APUS Reflex and community models on the original 1,000-question panel</title><desc id="accuracy-final-desc">Accuracy 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.</desc><rect width="1280" height="900" rx="18" fill="#f6f8fc"/><g font-family="Arial, Helvetica, sans-serif"><text x="40" y="48" font-size="28" fill="#14243c" font-weight="700">Decision models: reproducible results on the same 1,000-question panel</text><text x="40" y="78" font-size="15" fill="#60708a">Expanded development · 1,000 questions · full-input variants separated from published defaults</text><line x1="430.0" y1="105" x2="430.0" y2="737" stroke="#dce3ee"/><text x="418.0" y="767" font-size="12" fill="#60708a">0%</text><line x1="592.5" y1="105" x2="592.5" y2="737" stroke="#dce3ee"/><text x="580.5" y="767" font-size="12" fill="#60708a">25%</text><line x1="755.0" y1="105" x2="755.0" y2="737" stroke="#dce3ee"/><text x="743.0" y="767" font-size="12" fill="#60708a">50%</text><line x1="917.5" y1="105" x2="917.5" y2="737" stroke="#dce3ee"/><text x="905.5" y="767" font-size="12" fill="#60708a">75%</text><line x1="1080.0" y1="105" x2="1080.0" y2="737" stroke="#dce3ee"/><text x="1068.0" y="767" font-size="12" fill="#60708a">100%</text><text x="40" y="150" font-size="16" fill="#14243c" font-weight="600">APUS Reflex 35B-A3B · final-5949</text><text x="40" y="168" font-size="12" fill="#60708a">Historical baseline</text><rect x="430" y="129" width="534.30" height="30" rx="6" fill="#7952b3"/><text x="976.30" y="150" font-size="16" fill="#14243c" font-weight="700">82.20% · 822/1000</text><text x="40" y="202" font-size="16" fill="#14243c" font-weight="600">APUS Reflex 9B · final-5949</text><text x="40" y="220" font-size="12" fill="#60708a">Historical baseline</text><rect x="430" y="181" width="527.15" height="30" rx="6" fill="#245cdb"/><text x="969.15" y="202" font-size="16" fill="#14243c" font-weight="700">81.10% · 811/1000</text><text x="40" y="254" font-size="16" fill="#14243c" font-weight="600">APUS Reflex 4B · final-5949</text><text x="40" y="272" font-size="12" fill="#60708a">Historical baseline</text><rect x="430" y="233" width="523.25" height="30" rx="6" fill="#007f73"/><text x="965.25" y="254" font-size="16" fill="#14243c" font-weight="700">80.50% · 805/1000</text><text x="40" y="306" font-size="16" fill="#14243c" font-weight="600">Jev API · 1.13.0</text><text x="40" y="324" font-size="12" fill="#60708a">Official baseline</text><rect x="430" y="285" width="500.50" height="30" rx="6" fill="#52677f"/><text x="942.50" y="306" font-size="16" fill="#14243c" font-weight="700">77.00% · 770/1000</text><text x="40" y="358" font-size="16" fill="#14243c" font-weight="600">Nimble 9B · full-input-8192</text><text x="40" y="376" font-size="12" fill="#60708a">1000/1000 valid</text><rect x="430" y="337" width="442.65" height="30" rx="6" fill="#0b8f83"/><text x="884.65" y="358" font-size="16" fill="#14243c" font-weight="700">68.10% · 681/1000</text><text x="40" y="410" font-size="16" fill="#14243c" font-weight="600">Kev 9B · full-input-8192</text><text x="40" y="428" font-size="12" fill="#60708a">1000/1000 valid</text><rect x="430" y="389" width="437.45" height="30" rx="6" fill="#3478d4"/><text x="879.45" y="410" font-size="16" fill="#14243c" font-weight="700">67.30% · 673/1000</text><text x="40" y="462" font-size="16" fill="#14243c" font-weight="600">Nimble 9B · default 2048</text><text x="40" y="480" font-size="12" fill="#60708a">621/1000 valid</text><rect x="430" y="441" width="300.95" height="30" rx="6" fill="#6ba69f"/><text x="742.95" y="462" font-size="16" fill="#14243c" font-weight="700">46.30% · 463/1000</text><text x="40" y="514" font-size="16" fill="#14243c" font-weight="600">Kev 4B · full-input-8192</text><text x="40" y="532" font-size="12" fill="#60708a">1000/1000 valid</text><rect x="430" y="493" width="407.55" height="30" rx="6" fill="#449b92"/><text x="849.55" y="514" font-size="16" fill="#14243c" font-weight="700">62.70% · 627/1000</text><text x="40" y="566" font-size="16" fill="#14243c" font-weight="600">Von Marker · full-input</text><text x="40" y="584" font-size="12" fill="#60708a">1000/1000 valid · round4</text><rect x="430" y="545" width="291.20" height="30" rx="6" fill="#b37429"/><text x="733.20" y="566" font-size="16" fill="#14243c" font-weight="700">44.80% · 448/1000</text><text x="40" y="618" font-size="16" fill="#14243c" font-weight="600">Laya typed · full-input</text><text x="40" y="636" font-size="12" fill="#60708a">1000/1000 valid</text><rect x="430" y="597" width="328.25" height="30" rx="6" fill="#956b8c"/><text x="770.25" y="618" font-size="16" fill="#14243c" font-weight="700">50.50% · 505/1000</text><text x="40" y="670" font-size="16" fill="#14243c" font-weight="600">Laya English · full-input</text><text x="40" y="688" font-size="12" fill="#60708a">1000/1000 valid</text><rect x="430" y="649" width="325.65" height="30" rx="6" fill="#a97991"/><text x="767.65" y="670" font-size="16" fill="#14243c" font-weight="700">50.10% · 501/1000</text><text x="40" y="722" font-size="16" fill="#14243c" font-weight="600">Verdict · full-input-8192</text><text x="40" y="740" font-size="12" fill="#60708a">1000/1000 valid · K+1</text><rect x="430" y="701" width="44.20" height="30" rx="6" fill="#c15a48"/><text x="486.20" y="722" font-size="16" fill="#14243c" font-weight="700">6.80% · 68/1000</text><text x="40" y="915" font-size="16" fill="#14243c" font-weight="600">Von · Berta · full-input-16384</text><text x="40" y="933" font-size="12" fill="#60708a">1000/1000 valid · FP32/RoPE extrapolation</text><rect x="430" y="919" width="271.70" height="30" rx="6" fill="#c47736"/><text x="714" y="940" font-size="16" fill="#14243c" font-weight="700">41.80% · 418/1000</text><text x="40" y="1010" font-size="14" fill="#60708a">Valid = 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.</text><text x="40" y="1036" font-size="14" fill="#60708a">Nimble default 2048 and Kev default state/branch gates reduce coverage; full-input variants only raise the input budget and do not change weights.</text><text x="40" y="1062" font-size="14" fill="#60708a">Panel SHA256: f7562356… · expanded-dev，not sealed final · Browser is candidate-action selection only, not end-to-end browser task success.</text></g></svg> |