我们试过什么,什么被否掉了

已上线的模型是在 60 轮提案里选出来的:4 个模型族、验证季打分、锁定测试季只打开一次。前 31 轮(2026-11-07,Claude)选出了现在这一版;本轮新增 29 轮里 4 轮优于现网、10 轮持平、6 轮更差、4 轮连插件契约都没通过;1 轮在验证季赢了、开到锁定测试季却输给现网,因此没有上线。另有 5 轮是「恢复上一轮配置」的还原轮,不重复计分。下面逐轮列出这 29 轮的假设与结果——包括没赢的那些。

线上跑的正是这一版

验证季 RPS
0.2039
验证季 log loss
0.9976
锁定测试季 RPS
0.2051
测试季 log loss
1.0047

同一条测试集上:只看结果、不看盘口的模型 RPS 0.2087;只用盘口价先验的模型 0.2071。10 个概率分桶里 5 个校准良好、2 个样本不足——完整分桶见 校准记录。

伤停情报影子实验(预注册,不发中间结果)

每场赛前用带联网检索的模型读伤停与首发消息,得到一份只含事实与来源的清单;它不会直接改概率,而是与现行概率并列计分。采用规则在看数据之前就写死:成对 log loss 差的 90% 置信区间整体落在「更好」一侧才允许进入生产,否则归档为负结论。样本未满 150 场,因此不公布任何中间差值。

已记录
114
已结算
104
预注册门槛
150

还差 46 场结算;到点后这一页会直接给出结论,无论正负。

本轮 29 轮的假设与结果

验证季 RPS 越低越好;现网为 0.2039。全部 60 轮共花费 $59.31(claude-fable-5-1 $24.65、gpt-5.1 $0.67、claude-fable-5-2 $8.48、gpt-6.1-sol $0.86)。

机器可读版本:/data/football-model-search.json

提案者轮次假设验证 RPS测试 RPS结果
gpt-5.131Rescale score-only goal observations so each family’s long-run training goals match its own mean of real goals.0.2039—持平
gpt-5.132Replace the family-wide goal-level correction with a tier-aware level (tier-1 vs tier-2), blended by recent tier mix.0.2039—持平
gpt-5.133Give the most recent match a small extra weight per refit so ratings react faster to genuine strength changes.0.2039—持平
gpt-5.134Add a small symmetric rest-days term to the log-strength ratio.0.2040—更差
gpt-5.135Reweight short-term form/tempo residuals to give market-observed matches slightly more effective samples.0.2040—更差
gpt-5.136Taper the favourite stretch and rho shift in the first ten matchweeks.(plugin longer than 20,000 characters)——未通过契约
gpt-5.137Ship the round-26 model unchanged.(plugin longer than 20,000 characters)——未通过契约
claude-fable-5-231Opponent-adjust the fast form and tempo layers by backfitting, with the shrinkage re-derived each pass.0.2039—持平
claude-fable-5-232Adjust the form layer for rest asymmetry.(no <plugin> block in the answer)——未通过契约
claude-fable-5-233Scale the tier-gap correction by the priced share of the current tier-2 teams’ training evidence.0.2039—持平
claude-fable-5-234Count the season break as one extra in-season half-life for priced rows.0.2040—更差
claude-fable-5-235Split the market residual by league tier.(no <plugin> block in the answer)——未通过契约
claude-fable-5-236Scale the tier-gap correction by the mean priced share of the family.0.2039—持平
claude-fable-5-237Make the rho shift depend on match closeness.0.2039—持平
gpt-6.1-sol38Preserve the predicted goal total when applying strength, tier and form corrections.0.2039—持平
gpt-6.1-sol39Use only priced observations in the fast form and tempo layers, so short-window realised scores cannot add noise to unpriced fixtures.0.20380.2053测试季输
gpt-6.1-sol40Consolidate round 39 unchanged, so no further validation-tuned parameter is added.0.2038—更好
gpt-6.1-sol41Remove only the fast tempo adjustment and keep everything else.0.2038—更好
gpt-6.1-sol42Restore the market-only fast tempo adjustment on top of round 39.0.2038—更好
gpt-6.1-sol43Remove only the fast strength-form adjustment, keeping tempo.0.2047—更差
gpt-6.1-sol44Restore market-only fast strength form while keeping tempo.0.2038—同第 39 轮
gpt-6.1-sol45Train priced rows on market-implied goals alone instead of the 75/25 mixture.0.2042—更差
gpt-6.1-sol46Restore the 75% market / 25% compressed-score training mixture.0.2038—同第 39 轮
gpt-6.1-sol47Remove blow-out compression and its compensating goal-level correction.0.2039—持平
gpt-6.1-sol48Restore blow-out compression and its goal-level correction.0.2038—同第 39 轮
gpt-6.1-sol49Remove only the manual rho offset.0.2039—持平
gpt-6.1-sol50Restore the source-aware rho offset.0.2038—同第 39 轮
gpt-6.1-sol51Remove only the tier-gap correction.0.2042—更差
gpt-6.1-sol52Restore the source-scaled tier correction and otherwise ship round 39.0.2038—同第 39 轮

为什么这些轮次赢不了

已上线的插件已经把能用的结构都用上了:先验 0.5、按数据来源区分的时间衰减、可靠性加权、层级差位移、大比分压缩、随来源变化的 stretch 与 rho、14 天自收缩状态层、去压缩的市场残差。再往上加结构,改善都落在噪声里。剩下的空间是信息——伤停、首发、赛程密度、轮换——这些不在历史比分与收盘价里,改函数改不出来;所以它们走独立实验,用上面的预注册规则判生死。