# Team-neutral drafts and equal-strength lineup assessment

[简体中文](neutral_draft_model.zh_cn.md) · [Workbench](../neutral_draft.en.html)

The separate page has no team or player selection. BO1/3/5, independent map/order controls, all next-action probabilities, manual and greedy drafting, undo, winner entry, Fearless carry-over and export follow the team workbench. Previous picks lock for both sides; bans do not carry over. Only complete five-versus-five lineups enable win assessment, and only on this neutral page. Flex roles are marginalized or manually fixed to legal injective assignments.

## Draft policy

A separately fitted policy uses public M1 champion-role pick/ban demand and dated M2 relationships. Personal preferences, team comfort, head-to-head choices, team lane/style and team action timing are removed. Current M1 bidirectional partner transmission is required active: pick 1, extra ban 1. Unit transmission elasticity is structural; source heroes require own kit/item exposure. Ranked changes of unchanged partners never feed back. Sample shrinkage is selected chronologically, with zero retained as comparator.

Pairing counts actual legal role-choice opportunities, with absolute frequency, relative lift and effective support. Completion terms use multiplier 0.8; only extra consecutive urgency is capped at odds multiplier 2.00. Actual conditional pair evidence remains intact, without a hard public-meta rank clamp after a partner is revealed. Parameters are selected by pre-Worlds log loss.

Action training uses Full Fearless only. Existing professional importance/strength/event balance/first-game factors and 45-day decay remain. Twenty-eight S15 knockout games are repeatedly inspected retrospective diagnostics, not blind accuracy. Historic Classic outcomes can inform relationships without pooling their BP rates.

| 指标 / Metric | 选取 / Pick | 禁用 / Ban |
|---|---:|---:|
|Top 1|0.1893|0.2321|
|Top 3|0.4393|0.4536|
|Top 5|0.5643|0.5964|
|Log loss|2.9205|2.9303|


## Lineup model

13655 professional games support role-specific champion baselines, allied pairs, antisymmetric enemy pairs, a full four-champion bottom residual, composition, historical DPM, frontline/engage proxies, stage fit and shrunk 15-minute lane evidence. Logistic fitting includes pre-match Elo, team, player, patch, event and format controls. Only antisymmetric draft coefficients are exported: no identities, intercept or map-side advantage. Swapping the lineups complements the predicted probabilities.

Profiles update only after each result becomes available, before the later game's start. Current-match damage, duration, lane stats and outcome are not self-inputs. Damage composition uses historical exact-role contributions to team champion damage, requiring complete observed five-player sides; missing values remain unknown. Lane evidence preserves missingness and partial metric coverage, shrinking matchup estimates after an earlier-label-only strength residualization. Historical duration-conditioned early/late residuals are descriptive associations rather than causal power curves. Official dated types and kit text provide capability proxies.

The structure score is a separate descriptive rubric: 35% role coverage, 25% damage dispersion, 20% frontline coverage and 20% engage coverage. It is not a win label or a mandatory full-score checklist. Factor-off probability changes are correlated and non-additive. M1 BP popularity is not directly converted into win probability; new patches and sparse pairs remain uncertain.

Parameters were selected July–September 14, 2025; probability calibration used September 15–October 13. S15 evaluation: 84 games, full-context log loss 0.6241, Brier 0.2177. A later 2026 temporal check covers 941 games, log loss 0.6550. Full-context calibration does not establish calibrated equal-strength counterfactual probabilities. S16 Worlds has no measured outcome validation.

Tests cover identity invariance, injective role legality, 300 Fearless actions, swapped-lineup symmetry, Python/JS parity and bilingual BO flows. Four recent DCGI games still lack full independent order corroboration; missing practice accounts are not zero.

## Reproduce

```powershell
python models/meta/m1_final/train_partners.py --require-enabled
python models/meta/m1_final/promote_partners.py
# Refresh M3 public anchors and dated M4 snapshots after changing M1.
python -m models.draft_generic.train
python -m models.draft_generic.win_model
python -m models.draft_generic.report
python -m models.draft_generic.render
python -m models.draft_generic.verify
node models/draft_generic/verify_dashboard.cjs
```