# M4 最新选取策略 / Current pick policy

默认队伍 BP 入口与最终赛前冻结版均使用已完成检验的 `joint_role_pair` 策略，修正已经合并封存。M1–M3、M4 禁用模型与赛前输入不变。补丁仍为 26.20，输入截止仍为 2026-10-15 15:18:24 UTC。原始发布包保留在历史 Release。

候选英雄的每一种分路，都会重新计算该位置仍空着时已选英雄的联合分配，再估计组合搭配、共同选择与联动。正向搭配总贡献受分路可用概率约束，并对重复的搭配证据总加分设置上限。这样保留真实组合偏好，同时降低为补组合而把已有英雄挤去低优先级分路的倾向。

共享参数使用 S15 赛前选择段确定。28 局淘汰赛、280 次选取的重复回顾性检验：Top 3 为 50.71%（不变），Top 5 为 63.57%（不变），Log loss 为 2.80130（原 2.79685）。这次改动修正局面逻辑，不能据此声称整体准确率提高。原上游与淘汰赛样本此前已用于研究，不属于独立盲测。

当前入口为 `draft_predictions.html` / `draft_predictions.en.html`；最终冻结入口为 `draft_predictions.frozen.html` / `draft_predictions.frozen.en.html`。这两组页面加载同一套已封存参数与输入，冻结入口另从带封存标识的目录加载资源。无队伍版维持原先的冻结策略。

## English

The canonical team-draft and final pre-event frozen entries both use the tested and sealed `joint_role_pair` pick policy. M1–M3, the ban policy and input cutoff are unchanged. The previous release is preserved as an archive. The patch remains 26.20; inputs stop at 2026-10-15 15:18:24 UTC.

Ally-role relationships condition on a candidate's role being free. Total positive pairing evidence is gated by role availability and capped across correlated features. Supported combinations remain preferred without granting a full pairing reward for an unlikely reassignment of an existing pick.

Shared settings were selected on pre-Worlds S15 games. The repeatedly inspected 28-game knockout audit retains pick Top 3 accuracy of 50.71% and Top 5 accuracy of 63.57%; log loss changes from 2.79685 to 2.80130. This is a behavioral correction, not evidence of improved overall accuracy or blind end-to-end validation. The neutral policy is unchanged.