Preserve four task readouts and eight matched action heads

.gitattributes

@@ -37,3 +37,4 @@ snapshots/2026-10-25-onpolicy-credit-v1/model.json filter=lfs diff=lfs merge=lfs
 snapshots/2026-10-25-onpolicy-credit-3000-v1/model.json filter=lfs diff=lfs merge=lfs -text
 snapshots/2026-10-27-policy-scaling-128-stable-ties-v2/model.json filter=lfs diff=lfs merge=lfs -text
 snapshots/2026-10-27-crossfit-current-policy-targets-v1/model.json filter=lfs diff=lfs merge=lfs -text
+snapshots/2026-10-27-matched-readout-action-v1/model.json filter=lfs diff=lfs merge=lfs -text

snapshots/2026-10-27-matched-readout-action-v1/README.md

@@ -0,0 +1,76 @@
+# KiloGram 完整读出与匹配选题头
+
+Git https://github.com/luojiaxuan/cvpr-multimodal-evidence @557746bf16d40fa8a1e638e35a2de7342de3bc52。
+入口code/train_setground_kilogram_joint_readout.py；配置code/configs/setground_kilogram_joint_readout.json。
+固定原三折及全128训练到原9development/25heldout；四个完整17维读出、各两个56维df32动作头全部冻结后再评价。每折只由其训练组重选固定顺序问数，不按评价结果选模型或超参。学习门槛要求超过同读出prior/两固定顺序、保留的历史强控制及自身STOP/random/public_argmax，并有反馈见证；续策与即时单列配对差。
+缓存candidate-argmax反馈上的监督学习。重训完整终点与匹配选题头，旧选题先验/回答预测/题池/费用/预算均不变；continuation是固定随机续策Q回归，不是新on-policy RL。三折共享32锚组且训练重叠，原评价集均已查看，bootstrap仅探索性；不使用oracle上界充当策略。
+
+四个训练范围：原32锚组加64新组到另32组，三折；全128训练到原9development和25heldout。
+每范围重新拟合全部17维任务读出，原中心化/scale=5、目标正确比例减0.2、ridge=0.01、initial_parameters先验不变。
+随后在相同训练组重建全部当前反馈状态的56维分支特征与即时正确率增量减0.005标签；新读出不改变旧动作先验或回答预测。
+两个动作头共享训练数据、核基底、df32及STOP=0，一个拟合即时标签，另一个拟合原动作先验temperature=0.2固定随机续策的Q。
+后者经有序尾路径枚举核对，不对未问答案做GT最大化，但它仍是监督回归，不是on-policy RL。
+全部四读出及八头在评价前冻结；固定index/score问数只按各范围训练前缀选。
+
+| cohort | 策略 | 实测正确 | 问数 | U |
+| --- | --- | ---: | ---: | ---: |
+| crossfit | continuation | 285/480 | 226 | 0.581979 |
+| crossfit | immediate | 286/480 | 187 | 0.586094 |
+| crossfit | prior | 285/480 | 184 | 0.584167 |
+| crossfit | fixed:index | 277/480 | 257 | 0.563698 |
+| crossfit | fixed:score | 286/480 | 268 | 0.581875 |
+| crossfit | history:initial4:fixed:index:mbr | 301/480 | 268 | 0.613125 |
+| crossfit | history:initial4:fixed:score:mbr | 301/480 | 248 | 0.614167 |
+| crossfit | history:old56 | 279/480 | 224 | 0.569583 |
+| crossfit | history:old:fixed:index:mbr | 283/480 | 268 | 0.575625 |
+| crossfit | history:old:fixed:index:original | 266/480 | 268 | 0.540208 |
+| crossfit | history:old:fixed:score:mbr | 282/480 | 248 | 0.574583 |
+| crossfit | history:old:fixed:score:original | 265/480 | 248 | 0.539167 |
+| crossfit | history:old:prior:original | 270/480 | 184 | 0.552917 |
+| crossfit | history:pool59 | 283/480 | 229 | 0.577656 |
+| development | continuation | 27/45 | 23 | 0.587222 |
+| development | immediate | 26/45 | 20 | 0.566667 |
+| development | prior | 24/45 | 19 | 0.522778 |
+| development | fixed:index | 29/45 | 30 | 0.627778 |
+| development | fixed:score | 31/45 | 27 | 0.673889 |
+| development | history:immediate-43 | 29/45 | 29 | 0.628333 |
+| development | history:initial | 26/45 | 23 | 0.565000 |
+| development | history:initial4:fixed:index:mbr | 27/45 | 30 | 0.583333 |
+| development | history:initial4:fixed:score:mbr | 25/45 | 27 | 0.540556 |
+| development | history:old:fixed:index:mbr | 27/45 | 30 | 0.583333 |
+| development | history:old:fixed:index:original | 32/45 | 30 | 0.694444 |
+| development | history:old:fixed:score:mbr | 27/45 | 27 | 0.585000 |
+| development | history:old:fixed:score:original | 34/45 | 27 | 0.740556 |
+| development | history:old:prior:original | 25/45 | 19 | 0.545000 |
+| development | history:return_to_go-43 | 25/45 | 28 | 0.540000 |
+| heldout | continuation | 53/125 | 53 | 0.413400 |
+| heldout | immediate | 52/125 | 45 | 0.407000 |
+| heldout | prior | 51/125 | 48 | 0.398400 |
+| heldout | fixed:index | 57/125 | 65 | 0.443000 |
+| heldout | fixed:score | 55/125 | 59 | 0.428200 |
+| heldout | history:immediate-43 | 58/125 | 60 | 0.452000 |
+| heldout | history:initial | 56/125 | 57 | 0.436600 |
+| heldout | history:initial4:fixed:index:mbr | 61/125 | 65 | 0.475000 |
+| heldout | history:initial4:fixed:score:mbr | 60/125 | 59 | 0.468200 |
+| heldout | history:old:fixed:index:mbr | 68/125 | 65 | 0.531000 |
+| heldout | history:old:fixed:index:original | 64/125 | 65 | 0.499000 |
+| heldout | history:old:fixed:score:mbr | 65/125 | 59 | 0.508200 |
+| heldout | history:old:fixed:score:original | 60/125 | 59 | 0.468200 |
+| heldout | history:old:prior:original | 55/125 | 48 | 0.430400 |
+| heldout | history:return_to_go-43 | 55/125 | 60 | 0.428000 |
+
+预定点学习门槛：{"crossfit": {"continuation": false, "immediate": false}, "development": {"continuation": false, "immediate": false}, "heldout": {"continuation": false, "immediate": false}}。
+同读出prior与两固定顺序是匹配控制；history:*保留已发布原读出/MBR及旧学习方法，不能冒称这些历史方法都在本次相同训练规模重训。
+三折分别及合并96唯一组报告，不能把相互重叠的训练折当独立训练证据。原评价集已查看，组bootstrap为描述性，不能称新盲测。
+真假反馈、STOP、仅公共预测、删更新/替换/跳过控制与必要延迟见证均保存；没有全知oracle策略分数。
+
+核验：3705个训练读出状态及独立标量特征/正规方程，6636个训练动作转移及独立特征/即时标签、续策DP与有序枚举、RBF正规方程。
+390模型-组的自主/反馈干预及260固定前缀组次双实现核对，8988个完整评测决策。
+总236.503443 CPU秒，0新增GPU/VLM；历史Hyper01采集费用不被缓存复用抹除。
+
+records.tar.gz含inputs/provenance、四fit（读出状态、训练转移/标签、固定训练前缀、拟合检查）、freeze、130逐组文件、summary和四父发布回执。
+manifest提供逐件SHA；只含可公开数值任务，不含原图、原始私有文本、凭据或身份信息。
+模型model.json含公共settings、四范围的读出与动作参数及固定问数，不含环境truth/responses。
+数据和模型五件均从不可变revision逐字节回读。上游恢复位置在父回执，输入SHA在config；原图留既有受限仓库。
+复现：OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 PYTHONPATH=code python code/train_setground_kilogram_joint_readout.py --results <父来源目录> --root <新根> --config code/configs/setground_kilogram_joint_readout.json --revision 557746bf16d40fa8a1e638e35a2de7342de3bc52。
+部署用setground_kilogram_aligned_residual.AlignedActor(public_group(group), settings, model)，按scope选择prior/immediate/continuation；观察答案仅在实际选题后加入。

snapshots/2026-10-27-matched-readout-action-v1/model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54b4b8fbcda8e732c082f838a70121035c2932d3638ee57d9e2bdc05ba5d1bbc
+size 11315803