`}
function openModel(n,changeHash=true){const m=models.find(x=>x.number===n);if(!m)return;current=n;const a=m.architecture||{},p=m.parts||{},profile=profiles[m.profile]||{};
let out=`
#${n} · ${h(title(m))}
${h(m.id)}
${tags(m).map(t=>`${h(t)}`).join('')}
${h(tx('固定IDは構造・設定・seedを識別します。最高点は採点した時点の中の最大値です。','The permanent ID identifies a configuration and seed. Peaks are maxima over the evaluated ages.'))}
`;
if(m.condition_conflicts?.length)out+=`
${tx('このモデルには設定記録の食い違いがあります。下の出典と記録を確認してください。','This record has configuration conflicts. Inspect the evidence and metadata below.')}
`;
if(!m.present)out+=`
${tx('停止・削除された実験の登録情報だけが残っています。得点・構成の未確認部分は補完していません。','Only the registry identity of a cancelled/deleted experiment remains. Missing scores and configuration fields have not been inferred.')}
`;
out+=`
${tx('部品と構成','Components and architecture')}
`+props([
[tx('seed','Seed'),m.seed],[tx('状態幅 / FFN中間幅 / 層数','State width / FFN intermediate width / layers'),`${m.width??'?'} / ${m.ffn??'?'} / ${m.depth??'?'}`],['Attention',p.attention],['Router',p.router],[tx('候補の扱い','Candidate handling'),p.candidate],['Candidate Reader',p.reader],[tx('Attention受信ゲート','Attention receive gate'),p.received],['QK normalization',p.qk],[tx('出力部品','Output head'),p.output],[tx('各出力枝が使う層','Layers used by output branches'),p.reference_layers],[tx('FFN加算','FFN addition'),p.ffn_add],[tx('加算条件（記録値）','Addition condition (recorded)'),p.bypass_condition],[tx('F / D / C 記憶','F / D / C memory'),memory(m)||tx('該当する明示記録なし','No explicit matching memory record')],[tx('RoPE基数を学習','Learn RoPE base'),p.rope_learned],[tx('候補分散scaleを学習','Learn candidate variance scale'),p.candidate_var_learned],[tx('距離補正','Distance correction'),p.distance],[tx('学習パス数（記録値）','Training passes (recorded)'),a.passes],[tx('幾何（記録値）','Geometry (recorded)'),a.geometry],[tx('評価条件群','Evaluation cohort'),cohortNames[m.cohort]?.[lang()==='ja'?0:1]??m.cohort]]);
out+=`
${tx('各学習時点の旧評価','Legacy scores at each training age')}
${tx('Mは累積学習語数の百万単位。PIQA二集合の平均を1軸にしたものが6軸です。','M denotes millions of words of cumulative training exposure. Six axes treat the mean of the two PIQA sets as one axis.')}
`;
out+=`
M
${tx('5軸','Five')}
${tx('6軸','Six')}
${Object.values(taskNames).map(x=>`
${x}
`).join('')}
${m.points.map(pt=>`
${pt.age}
${fmt(pt.five)}
${fmt(pt.six)}
${Object.keys(taskNames).map(k=>`
${fmt(pt.tasks[k],2)}
`).join('')}
`).join('')||`
${tx('未採点','Not evaluated')}
`}
`;
out+=`
${tx('明確回答：総正答率','Clear-answer: pooled accuracy')}
`;
out+=m.clear_points.length?`
M
${tx('正答率','Accuracy')}
${tx('正答数 / 問題数','Correct / total')}
${tx('選択肢の得点同点数','Answer-score ties')}
${m.clear_points.map(pt=>`
${pt.age}
${fmt(pt.accuracy)}%
${pt.correct} / ${pt.total}
${pt.ties??'—'}
`).join('')}
`:`
${tx('この問題集では未採点です。','Not evaluated on this bank.')}
`;
out+=`
${tx('学習率・重み減衰・更新条件','Learning rates, weight decay, and updates')}
${tx('実LRを名前付きparameter群で示します。3値はscheduleに沿った各段階です。WDの「—」は未記録で、0ではありません。','Named parameter groups show effective LRs. Triplets follow the recorded schedule phases. A missing WD is unknown, not zero.')}