diff --git a/README.md b/README.md
index d5ff69b..18ba11f 100644
--- a/README.md
+++ b/README.md
@@ -27,8 +27,8 @@ classifier-free guidance（CFG）策略。
 | 1. Base policy | `64×64` 图像、离散历史 XY、任务提示 | 小型 VLM policy + stop-gradient flow expert | subtask、FAST-lite tokens、连续 `16×2` action chunk | token CE + flow matching |
 | 2. Experience | 一条真实示范轨迹 | 确定性经验构造 | 成功示范、understeer、提前停止、人工纠正 4 类 episode | 无梯度，仅构造 buffer |
 | 3. Value | 图像、离散状态、任务、episode return | 小型独立 value VLM | `201` 个 return bins，期望 value 位于 `[-1,0]` | distributional CE |
-| 4. Advantage | reward、当前/未来 value、intervention mask | n-step estimator + quantile threshold | positive/negative indicator；人工纠正强制 positive | 无梯度，生成策略条件 |
-| 5. Mixed warm-up | 示范与合成经验、advantage 条件 | advantage-conditioned policy | 条件化离散动作和连续 flow | token CE + flow matching |
+| 4. Advantage | reward、学习到的当前/未来 value、intervention mask | critic inference + n-step estimator + quantile threshold | positive/negative indicator；人工纠正强制 positive | value 更新后重新标注策略数据 |
+| 5. Mixed SFT baseline | 示范与合成经验，不提供 advantage | unconditioned mixed policy | 离散动作和连续 flow | token CE + flow matching |
 | 6. RECAP extraction | 更新后的 value、重新标注的全部经验 | policy 从 base checkpoint 再训练 | 改进后的 `π*0.6`-style policy | 条件化 token CE + flow matching |
 | 7. CFG inference | 同一 observation 分别使用 positive 和 omitted 条件 | 两次 flow 前向 + CFG，默认 `β=2` | guided 连续 action chunk | 与真值及 unguided 输出比较 |
 
@@ -43,6 +43,11 @@ advantage token 位于 subtask 之后、动作之前，因此它只改变动作
 读取 teacher-forced FAST-lite 真值动作。flow loss 对 policy/VLM prefix 使用
 stop-gradient，以最小形式模拟论文的 Knowledge Insulation（KI）训练原则。
 
+critic 只读取 observation/state（本单任务实验隐含固定任务提示），不读取
+`demo_success`、失败类型或 intervention source 等结果标签。value 训练完成后，
+程序会对每条 episode 的所有状态重新运行 critic，并用预测值重新计算 n-step
+advantage；最终 RECAP policy 使用该重标注结果，而不是手工 baseline。
+
 ## 与 π0.5 最小项目的重点差异
 
 | π0.5 minimal | π0.6 / `π*0.6` minimal |
@@ -66,7 +71,7 @@ width 为 96，value width 为 64；没有复现这些规模和数据改进。
 - `autonomous_stop`：中途停止；
 - `intervention_correction`：先偏离，再由“人工”轨迹拉回。
 
-默认用 4-step bootstrapped advantage，`gamma=0.99`。训练集的 advantage
+默认用 4-step bootstrapped advantage，按论文使用无折扣 return（`gamma=1.0`）。训练集的 advantage
 第 70 百分位作为 positive 阈值，即约 30% 普通样本为 positive；intervention
 动作无条件标为 positive。这个合成 buffer 只用于检验 RECAP 接口，不能替代
 真实机器人 rollout、任务成功标签或在线人工干预。
@@ -120,7 +125,7 @@ python -m pi06mini.train --config configs/overfit.yaml
 |---|---:|---|
 | base | 300 | 先记住示范的 subtask、FAST-lite 和连续动作 |
 | value | 400 | 记住成功、失败和纠正 episode 的 201-bin return |
-| mixed | 300 | 让策略学会 positive/negative/omitted 三种条件 |
+| mixed | 300 | 训练不带 advantage 的全经验 SFT 对照组 |
 | recap | 700 | 使用合成经验中的 n-step advantage 标签与 30% 条件 dropout 提取最终策略 |
 
 训练应保存 checkpoint、每阶段 loss、value/advantage 诊断和最终评估。CPU 可以
@@ -148,6 +153,8 @@ python -m pi06mini.infer --checkpoint outputs/overfit/checkpoint.pt
 
 在极小数据过拟合场景中，目标是看到 value 能区分合成成功/失败，positive 或
 CFG 策略比 negative/unguided 更接近专家轨迹。它不代表模型具备泛化能力。
+推理链会先由 VLM 预测 subtask token，再将该预测与 advantage indicator 一起
+交给连续 action expert，不使用数据集中的真值 subtask 作为低层条件。
 
 ## 关键配置
 
diff --git a/configs/overfit.yaml b/configs/overfit.yaml
index 92a0a1f..1a47329 100644
--- a/configs/overfit.yaml
+++ b/configs/overfit.yaml
@@ -10,7 +10,7 @@ data:
 
 experience:
   augmentation_factor: 1
-  gamma: 0.99
+  gamma: 1.0
   n_step: 4
   positive_quantile: 0.7
 
diff --git a/outputs/overfit/critic_relabeling.json b/outputs/overfit/critic_relabeling.json
new file mode 100644
index 0000000..2486787
--- /dev/null
+++ b/outputs/overfit/critic_relabeling.json
@@ -0,0 +1,88 @@
+{
+  "predicted_values": [
+    [
+      -0.20591823756694794,
+      -0.20782312750816345,
+      -0.20701773464679718,
+      -0.2082844078540802,
+      -0.22784602642059326,
+      -0.2076655775308609,
+      -0.1778903752565384,
+      -0.06647361069917679,
+      -0.028201263397932053,
+      -0.02679925411939621,
+      -0.018520547077059746,
+      -0.018748600035905838,
+      -0.01845482364296913,
+      -0.017860613763332367,
+      -0.01756754145026207,
+      -0.017646076157689095,
+      -0.017747100442647934
+    ],
+    [
+      -0.20591823756694794,
+      -0.20782312750816345,
+      -0.20701773464679718,
+      -0.2082844078540802,
+      -0.1956789344549179,
+      -0.1798555701971054,
+      -0.18803948163986206,
+      -0.1962352693080902,
+      -0.177969291806221,
+      -0.21014070510864258,
+      -0.21469540894031525,
+      -0.218982994556427,
+      -0.170505091547966,
+      -0.16970007121562958,
+      -0.1965235322713852,
+      -0.06712071597576141,
+      -0.026878785341978073
+    ],
+    [
+      -0.20591823756694794,
+      -0.20782312750816345,
+      -0.20701773464679718,
+      -0.2082844078540802,
+      -0.22784602642059326,
+      -0.2076655775308609,
+      -0.18868279457092285,
+      -0.3445330560207367,
+      -0.33078309893608093,
+      -0.3386183977127075,
+      -0.3102336525917053,
+      -0.29943764209747314,
+      -0.2533053159713745,
+      -0.2533053159713745,
+      -0.2533053159713745,
+      -0.2533053159713745,
+      -0.2533053159713745
+    ],
+    [
+      -0.20591823756694794,
+      -0.20782312750816345,
+      -0.20701773464679718,
+      -0.2082844078540802,
+      -0.1956789344549179,
+      -0.1798555701971054,
+      -0.16467693448066711,
+      -0.09015780687332153,
+      -0.07624924182891846,
+      -0.07263262569904327,
+      -0.016775300726294518,
+      -0.01684822142124176,
+      -0.017563480883836746,
+      -0.01873198337852955,
+      -0.019441569223999977,
+      -0.019780147820711136,
+      -0.0206752996891737
+    ]
+  ],
+  "advantage_threshold": 0.02545459009706974,
+  "episode_advantage_threshold": 0.046584345400333405,
+  "episode_labels": [
+    "positive",
+    "negative",
+    "negative",
+    "positive"
+  ]
+}
\ No newline at end of file
diff --git a/outputs/overfit/evaluation.json b/outputs/overfit/evaluation.json
index 7d7880f..90a87ae 100644
--- a/outputs/overfit/evaluation.json
+++ b/outputs/overfit/evaluation.json
@@ -1,32 +1,182 @@
 {
+  "predicted_subtask": "steer_left",
+  "target_subtask": "steer_left",
   "trajectory_metrics": {
-    "mixed": 100.86174011230469,
-    "unconditional": 21.51112937927246,
-    "positive": 0.061239250004291534,
-    "negative": 55.823570251464844,
-    "cfg": 0.24478793144226074,
-    "negative_to_failure_mse": 48.71984100341797
+    "mixed": 5.046791076660156,
+    "unconditional": 330.43865966796875,
+    "positive": 1.3081321716308594,
+    "negative": 351.36962890625,
+    "cfg": 0.06475023925304413,
+    "negative_to_failure_mse": 20.35787582397461
   },
   "values": [
     {
       "source_type": "demo_success",
-      "target": 0.0,
-      "prediction": -0.007052943576127291
+      "targets": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "predictions": [
+        -0.20591823756694794,
+        -0.20782312750816345,
+        -0.20701773464679718,
+        -0.2082844078540802,
+        -0.22784602642059326,
+        -0.2076655775308609,
+        -0.1778903752565384,
+        -0.06647361069917679,
+        -0.028201263397932053,
+        -0.02679925411939621,
+        -0.018520547077059746,
+        -0.018748600035905838,
+        -0.01845482364296913,
+        -0.017860613763332367,
+        -0.01756754145026207,
+        -0.017646076157689095,
+        -0.017747100442647934
+      ],
+      "curve_mse": 0.018043329939246178
     },
     {
       "source_type": "autonomous_understeer",
-      "target": -0.03799404203891754,
-      "prediction": -0.04714635759592056
+      "targets": [
+        -0.03993216156959534,
+        -0.03993084654211998,
+        -0.03992237150669098,
+        -0.03984469920396805,
+        -0.0396307148039341,
+        -0.0392293743789196,
+        -0.03840578347444534,
+        -0.037170957773923874,
+        -0.035448621958494186,
+        -0.03320406377315521,
+        -0.030476942658424377,
+        -0.027057264000177383,
+        -0.022992689162492752,
+        -0.018253253772854805,
+        -0.01285458542406559,
+        -0.006825999356806278,
+        0.0
+      ],
+      "predictions": [
+        -0.20591823756694794,
+        -0.20782312750816345,
+        -0.20701773464679718,
+        -0.2082844078540802,
+        -0.1956789344549179,
+        -0.1798555701971054,
+        -0.18803948163986206,
+        -0.1962352693080902,
+        -0.177969291806221,
+        -0.21014070510864258,
+        -0.21469540894031525,
+        -0.218982994556427,
+        -0.170505091547966,
+        -0.16970007121562958,
+        -0.1965235322713852,
+        -0.06712071597576141,
+        -0.026878785341978073
+      ],
+      "curve_mse": 0.024060551077127457
     },
     {
       "source_type": "autonomous_stop",
-      "target": -0.9509902000427246,
-      "prediction": -0.9405871629714966
+      "targets": [
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -0.9818642139434814,
+        -0.949521541595459,
+        -0.9027397632598877,
+        -0.8413366675376892,
+        -0.7655149698257446,
+        -0.6720305681228638,
+        -0.5647895336151123,
+        -0.444036602973938,
+        -0.30988287925720215,
+        -0.16282053291797638,
+        0.0
+      ],
+      "predictions": [
+        -0.20591823756694794,
+        -0.20782312750816345,
+        -0.20701773464679718,
+        -0.2082844078540802,
+        -0.22784602642059326,
+        -0.2076655775308609,
+        -0.18868279457092285,
+        -0.3445330560207367,
+        -0.33078309893608093,
+        -0.3386183977127075,
+        -0.3102336525917053,
+        -0.29943764209747314,
+        -0.2533053159713745,
+        -0.2533053159713745,
+        -0.2533053159713745,
+        -0.2533053159713745,
+        -0.2533053159713745
+      ],
+      "curve_mse": 0.34516632556915283
     },
     {
       "source_type": "intervention_correction",
-      "target": -0.0051981499418616295,
-      "prediction": -0.0118478462100029
+      "targets": [
+        -0.005220414604991674,
+        -0.005219103302806616,
+        -0.005210627801716328,
+        -0.0051329550333321095,
+        -0.004918968304991722,
+        -0.00456627644598484,
+        -0.003942344803363085,
+        -0.003156546037644148,
+        -0.0022692831698805094,
+        -0.0013850623508915305,
+        -0.0006413022056221962,
+        -0.00012316892389208078,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "predictions": [
+        -0.20591823756694794,
+        -0.20782312750816345,
+        -0.20701773464679718,
+        -0.2082844078540802,
+        -0.1956789344549179,
+        -0.1798555701971054,
+        -0.16467693448066711,
+        -0.09015780687332153,
+        -0.07624924182891846,
+        -0.07263262569904327,
+        -0.016775300726294518,
+        -0.01684822142124176,
+        -0.017563480883836746,
+        -0.01873198337852955,
+        -0.019441569223999977,
+        -0.019780147820711136,
+        -0.0206752996891737
+      ],
+      "curve_mse": 0.016281794756650925
     }
   ]
 }
\ No newline at end of file
diff --git a/outputs/overfit/recap_trajectory_comparison.png b/outputs/overfit/recap_trajectory_comparison.png
index 3b81dd6..638dbc7 100644
Binary files a/outputs/overfit/recap_trajectory_comparison.png and b/outputs/overfit/recap_trajectory_comparison.png differ
diff --git a/outputs/overfit/training_history.json b/outputs/overfit/training_history.json
index 7cdf30f..5d88c5c 100644
--- a/outputs/overfit/training_history.json
+++ b/outputs/overfit/training_history.json
@@ -2,624 +2,624 @@
   {
     "stage": "base",
     "step": 1,
-    "loss": 72.41638946533203,
-    "text_loss": 68.23114013671875,
-    "flow_loss": 2.0926246643066406
+    "loss": 70.63429260253906,
+    "text_loss": 68.23114776611328,
+    "flow_loss": 1.2015743255615234
   },
   {
     "stage": "base",
     "step": 20,
-    "loss": 4.305421829223633,
-    "text_loss": 1.9023208618164062,
-    "flow_loss": 1.2015506029129028
+    "loss": 3.8942408561706543,
+    "text_loss": 1.7884641885757446,
+    "flow_loss": 1.0528883934020996
   },
   {
     "stage": "base",
     "step": 40,
-    "loss": 2.912196159362793,
-    "text_loss": 0.8887173533439636,
-    "flow_loss": 1.0117393732070923
+    "loss": 3.1883139610290527,
+    "text_loss": 0.8112223148345947,
+    "flow_loss": 1.188545823097229
   },
   {
     "stage": "base",
     "step": 60,
-    "loss": 2.184612512588501,
-    "text_loss": 0.5364341735839844,
-    "flow_loss": 0.8240891695022583
+    "loss": 2.8662467002868652,
+    "text_loss": 0.5043408274650574,
+    "flow_loss": 1.1809529066085815
   },
   {
     "stage": "base",
     "step": 80,
-    "loss": 1.578565239906311,
-    "text_loss": 0.23946531116962433,
-    "flow_loss": 0.6695499420166016
+    "loss": 2.9415433406829834,
+    "text_loss": 0.2767218053340912,
+    "flow_loss": 1.3324108123779297
   },
   {
     "stage": "base",
     "step": 100,
-    "loss": 1.5856558084487915,
-    "text_loss": 0.04227600246667862,
-    "flow_loss": 0.7716898918151855
+    "loss": 1.23435640335083,
+    "text_loss": 0.013344191014766693,
+    "flow_loss": 0.6105061173439026
   },
   {
     "stage": "base",
     "step": 120,
-    "loss": 0.3423895239830017,
-    "text_loss": 0.0004414900322444737,
-    "flow_loss": 0.1709740161895752
+    "loss": 0.6844875812530518,
+    "text_loss": 0.1804233342409134,
+    "flow_loss": 0.2520321309566498
   },
   {
     "stage": "base",
     "step": 140,
-    "loss": 0.24176472425460815,
-    "text_loss": 1.8706961782299913e-05,
-    "flow_loss": 0.12087301164865494
+    "loss": 1.1563541889190674,
+    "text_loss": 6.113824201747775e-05,
+    "flow_loss": 0.5781465172767639
   },
   {
     "stage": "base",
     "step": 160,
-    "loss": 0.3000774383544922,
-    "text_loss": 7.637506496394053e-05,
-    "flow_loss": 0.15000052750110626
+    "loss": 0.49136143922805786,
+    "text_loss": 0.00014380548964254558,
+    "flow_loss": 0.24560882151126862
   },
   {
     "stage": "base",
     "step": 180,
-    "loss": 0.6410971879959106,
-    "text_loss": 1.6521895304322243e-05,
-    "flow_loss": 0.32054033875465393
+    "loss": 0.5463516712188721,
+    "text_loss": 1.483415599068394e-05,
+    "flow_loss": 0.2731684148311615
   },
   {
     "stage": "base",
     "step": 200,
-    "loss": 0.6132829189300537,
-    "text_loss": 9.676185072748922e-06,
-    "flow_loss": 0.30663663148880005
+    "loss": 0.21821065247058868,
+    "text_loss": 4.533280844043475e-06,
+    "flow_loss": 0.10910306125879288
   },
   {
     "stage": "base",
     "step": 220,
-    "loss": 0.3431224524974823,
-    "text_loss": 7.356796231761109e-06,
-    "flow_loss": 0.17155754566192627
+    "loss": 0.2135496884584427,
+    "text_loss": 3.412749947528937e-06,
+    "flow_loss": 0.10677313804626465
   },
   {
     "stage": "base",
     "step": 240,
-    "loss": 0.28047603368759155,
-    "text_loss": 5.987634267512476e-06,
-    "flow_loss": 0.1402350217103958
+    "loss": 0.15570451319217682,
+    "text_loss": 3.027882712558494e-06,
+    "flow_loss": 0.07785074412822723
   },
   {
     "stage": "base",
     "step": 260,
-    "loss": 0.13324393332004547,
-    "text_loss": 4.9011541705112904e-06,
-    "flow_loss": 0.06661951541900635
+    "loss": 0.17475822567939758,
+    "text_loss": 2.816716460074531e-06,
+    "flow_loss": 0.08737770467996597
   },
   {
     "stage": "base",
     "step": 280,
-    "loss": 0.1635851263999939,
-    "text_loss": 4.0462705328536686e-06,
-    "flow_loss": 0.08179053664207458
+    "loss": 0.12843455374240875,
+    "text_loss": 2.6055499802168924e-06,
+    "flow_loss": 0.06421597301959991
   },
   {
     "stage": "base",
     "step": 300,
-    "loss": 0.15081387758255005,
-    "text_loss": 3.4570459774840856e-06,
-    "flow_loss": 0.07540521025657654
+    "loss": 0.3167646825313568,
+    "text_loss": 2.387570930295624e-06,
+    "flow_loss": 0.1583811491727829
   },
   {
     "stage": "value",
     "step": 1,
-    "loss": 5.427414894104004,
-    "prediction": -0.5049723982810974,
+    "loss": 5.466859817504883,
+    "prediction": -0.5042269229888916,
     "target": 0.0
   },
   {
     "stage": "value",
     "step": 20,
-    "loss": 2.24937105178833,
-    "prediction": -0.3925527334213257,
-    "target": -0.0051981499418616295
+    "loss": 3.274851083755493,
+    "prediction": -0.3775767683982849,
+    "target": -0.005210627801716328
   },
   {
     "stage": "value",
     "step": 40,
-    "loss": 1.728512167930603,
-    "prediction": -0.32079368829727173,
-    "target": -0.0051981499418616295
+    "loss": 2.3394997119903564,
+    "prediction": -0.2888708710670471,
+    "target": -0.00456627644598484
   },
   {
     "stage": "value",
     "step": 60,
-    "loss": 1.4823004007339478,
-    "prediction": -0.3057250678539276,
-    "target": -0.0051981499418616295
+    "loss": 0.9264469146728516,
+    "prediction": -0.2154604196548462,
+    "target": -0.0022692831698805094
   },
   {
     "stage": "value",
     "step": 80,
-    "loss": 1.1117331981658936,
-    "prediction": -0.3553512394428253,
-    "target": -0.0051981499418616295
+    "loss": 0.756015956401825,
+    "prediction": -0.19315782189369202,
+    "target": -0.00012316892389208078
   },
   {
     "stage": "value",
     "step": 100,
-    "loss": 0.2580283582210541,
-    "prediction": -0.11316870898008347,
-    "target": -0.0051981499418616295
+    "loss": 0.6257156729698181,
+    "prediction": -0.18343542516231537,
+    "target": 0.0
   },
   {
     "stage": "value",
     "step": 120,
-    "loss": 0.12432077527046204,
-    "prediction": -0.054820768535137177,
-    "target": -0.0051981499418616295
+    "loss": 1.3138372898101807,
+    "prediction": -0.37475356459617615,
+    "target": -0.005220414604991674
   },
   {
     "stage": "value",
     "step": 140,
-    "loss": 0.08916354924440384,
-    "prediction": -0.04263461381196976,
-    "target": -0.0051981499418616295
+    "loss": 1.4211702346801758,
+    "prediction": -0.1804044544696808,
+    "target": -0.0051329550333321095
   },
   {
     "stage": "value",
     "step": 160,
-    "loss": 0.07165656983852386,
-    "prediction": -0.034443676471710205,
-    "target": -0.0051981499418616295
+    "loss": 1.448256254196167,
+    "prediction": -0.2506459355354309,
+    "target": -0.003942344803363085
   },
   {
     "stage": "value",
     "step": 180,
-    "loss": 0.05974685773253441,
-    "prediction": -0.029927225783467293,
-    "target": -0.0051981499418616295
+    "loss": 0.33143094182014465,
+    "prediction": -0.08220178633928299,
+    "target": -0.0013850623508915305
   },
   {
     "stage": "value",
     "step": 200,
-    "loss": 0.05075548216700554,
-    "prediction": -0.026268303394317627,
-    "target": -0.0051981499418616295
+    "loss": 0.2722547948360443,
+    "prediction": -0.06349469721317291,
+    "target": 0.0
   },
   {
     "stage": "value",
     "step": 220,
-    "loss": 0.043695785105228424,
-    "prediction": -0.023351771757006645,
-    "target": -0.0051981499418616295
+    "loss": 0.22779540717601776,
+    "prediction": -0.05417971312999725,
+    "target": 0.0
   },
   {
     "stage": "value",
     "step": 240,
-    "loss": 0.03802856057882309,
-    "prediction": -0.021020548418164253,
-    "target": -0.0051981499418616295
+    "loss": 1.5308982133865356,
+    "prediction": -0.3449951708316803,
+    "target": -0.005219103302806616
   },
   {
     "stage": "value",
     "step": 260,
-    "loss": 0.03340449184179306,
-    "prediction": -0.019105376675724983,
-    "target": -0.0051981499418616295
+    "loss": 1.1563409566879272,
+    "prediction": -0.2026708871126175,
+    "target": -0.004918968304991722
   },
   {
     "stage": "value",
     "step": 280,
-    "loss": 0.02958041802048683,
-    "prediction": -0.017514187842607498,
-    "target": -0.0051981499418616295
+    "loss": 1.2716314792633057,
+    "prediction": -0.07276482880115509,
+    "target": -0.003156546037644148
   },
   {
     "stage": "value",
     "step": 300,
-    "loss": 0.02638067863881588,
-    "prediction": -0.016178548336029053,
-    "target": -0.0051981499418616295
+    "loss": 0.3126322329044342,
+    "prediction": -0.07691694051027298,
+    "target": -0.0006413022056221962
   },
   {
     "stage": "value",
     "step": 320,
-    "loss": 0.023675711825489998,
-    "prediction": -0.01504577323794365,
-    "target": -0.0051981499418616295
+    "loss": 0.1044526994228363,
+    "prediction": -0.03439018875360489,
+    "target": 0.0
   },
   {
     "stage": "value",
     "step": 340,
-    "loss": 0.021368399262428284,
-    "prediction": -0.014076616615056992,
-    "target": -0.0051981499418616295
+    "loss": 0.1302744299173355,
+    "prediction": -0.03440919518470764,
+    "target": 0.0
   },
   {
     "stage": "value",
     "step": 360,
-    "loss": 0.019383536651730537,
-    "prediction": -0.013240991160273552,
-    "target": -0.0051981499418616295
+    "loss": 1.5437527894973755,
+    "prediction": -0.24599501490592957,
+    "target": -0.005210627801716328
   },
   {
     "stage": "value",
     "step": 380,
-    "loss": 0.017663555219769478,
-    "prediction": -0.01251534465700388,
-    "target": -0.0051981499418616295
+    "loss": 1.4080818891525269,
+    "prediction": -0.3180582821369171,
+    "target": -0.00456627644598484
   },
   {
     "stage": "value",
     "step": 400,
-    "loss": 0.016163267195224762,
-    "prediction": -0.011881118640303612,
-    "target": -0.0051981499418616295
+    "loss": 0.8151843547821045,
+    "prediction": -0.07982967793941498,
+    "target": -0.0022692831698805094
   },
   {
     "stage": "mixed",
     "step": 1,
-    "loss": 0.11788154393434525,
-    "text_loss": 3.4366105410299497e-06,
-    "flow_loss": 0.05893905460834503
+    "loss": 0.05100901424884796,
+    "text_loss": 2.3773534394422313e-06,
+    "flow_loss": 0.02550331875681877
   },
   {
     "stage": "mixed",
     "step": 20,
-    "loss": 4.7238688468933105,
-    "text_loss": 2.6195578575134277,
-    "flow_loss": 1.0521554946899414
+    "loss": 3.583022117614746,
+    "text_loss": 2.2782084941864014,
+    "flow_loss": 0.6524068117141724
   },
   {
     "stage": "mixed",
     "step": 40,
-    "loss": 5.051268577575684,
-    "text_loss": 1.2949998378753662,
-    "flow_loss": 1.8781343698501587
+    "loss": 2.2450661659240723,
+    "text_loss": 1.1263315677642822,
+    "flow_loss": 0.5593672394752502
   },
   {
     "stage": "mixed",
     "step": 60,
-    "loss": 1.6957576274871826,
-    "text_loss": 0.8319522142410278,
-    "flow_loss": 0.4319027364253998
+    "loss": 2.665435552597046,
+    "text_loss": 0.9095665812492371,
+    "flow_loss": 0.8779345154762268
   },
   {
     "stage": "mixed",
     "step": 80,
-    "loss": 2.1900432109832764,
-    "text_loss": 0.8061568140983582,
-    "flow_loss": 0.6919432282447815
+    "loss": 0.9744815826416016,
+    "text_loss": 0.7057558298110962,
+    "flow_loss": 0.1343628615140915
   },
   {
     "stage": "mixed",
     "step": 100,
-    "loss": 2.2935876846313477,
-    "text_loss": 0.5968648791313171,
-    "flow_loss": 0.8483614325523376
+    "loss": 2.257481575012207,
+    "text_loss": 0.7136633992195129,
+    "flow_loss": 0.7719091176986694
   },
   {
     "stage": "mixed",
     "step": 120,
-    "loss": 1.4990849494934082,
-    "text_loss": 0.6506502032279968,
-    "flow_loss": 0.4242174029350281
+    "loss": 1.557675838470459,
+    "text_loss": 0.6338224411010742,
+    "flow_loss": 0.46192666888237
   },
   {
     "stage": "mixed",
     "step": 140,
-    "loss": 1.6056029796600342,
-    "text_loss": 0.5977785587310791,
-    "flow_loss": 0.5039122104644775
+    "loss": 1.2216718196868896,
+    "text_loss": 0.7448893785476685,
+    "flow_loss": 0.2383912056684494
   },
   {
     "stage": "mixed",
     "step": 160,
-    "loss": 1.3810169696807861,
-    "text_loss": 0.48245665431022644,
-    "flow_loss": 0.44928014278411865
+    "loss": 0.5911489725112915,
+    "text_loss": 0.27954256534576416,
+    "flow_loss": 0.15580320358276367
   },
   {
     "stage": "mixed",
     "step": 180,
-    "loss": 1.1724852323532104,
-    "text_loss": 0.3958587944507599,
-    "flow_loss": 0.3883132338523865
+    "loss": 2.7055318355560303,
+    "text_loss": 0.25718608498573303,
+    "flow_loss": 1.224172830581665
   },
   {
     "stage": "mixed",
     "step": 200,
-    "loss": 1.2829020023345947,
-    "text_loss": 0.18238645792007446,
-    "flow_loss": 0.5502578020095825
+    "loss": 1.3802378177642822,
+    "text_loss": 0.32982689142227173,
+    "flow_loss": 0.5252054929733276
   },
   {
     "stage": "mixed",
     "step": 220,
-    "loss": 0.7328482866287231,
-    "text_loss": 0.1418272703886032,
-    "flow_loss": 0.29551050066947937
+    "loss": 1.1629283428192139,
+    "text_loss": 0.26602569222450256,
+    "flow_loss": 0.44845134019851685
   },
   {
     "stage": "mixed",
     "step": 240,
-    "loss": 1.196341872215271,
-    "text_loss": 0.28424081206321716,
-    "flow_loss": 0.4560505449771881
+    "loss": 0.5630282759666443,
+    "text_loss": 0.2572731077671051,
+    "flow_loss": 0.1528775840997696
   },
   {
     "stage": "mixed",
     "step": 260,
-    "loss": 4.298240661621094,
-    "text_loss": 0.257877379655838,
-    "flow_loss": 2.020181655883789
+    "loss": 1.1026716232299805,
+    "text_loss": 0.15438926219940186,
+    "flow_loss": 0.4741412103176117
   },
   {
     "stage": "mixed",
     "step": 280,
-    "loss": 1.3184027671813965,
-    "text_loss": 0.07380203157663345,
-    "flow_loss": 0.622300386428833
+    "loss": 1.9438565969467163,
+    "text_loss": 0.18404552340507507,
+    "flow_loss": 0.8799055218696594
   },
   {
     "stage": "mixed",
     "step": 300,
-    "loss": 2.4886393547058105,
-    "text_loss": 0.30716001987457275,
-    "flow_loss": 1.0907396078109741
+    "loss": 0.8524055480957031,
+    "text_loss": 0.1450427919626236,
+    "flow_loss": 0.35368138551712036
   },
   {
     "stage": "recap",
     "step": 1,
-    "loss": 10.11633586883545,
-    "text_loss": 9.866013526916504,
-    "flow_loss": 0.12516121566295624
+    "loss": 10.587621688842773,
+    "text_loss": 9.856688499450684,
+    "flow_loss": 0.3654666244983673
   },
   {
     "stage": "recap",
     "step": 20,
-    "loss": 3.1665196418762207,
-    "text_loss": 2.23919415473938,
-    "flow_loss": 0.4636627733707428
+    "loss": 3.289060354232788,
+    "text_loss": 2.240044355392456,
+    "flow_loss": 0.524507999420166
   },
   {
     "stage": "recap",
     "step": 40,
-    "loss": 3.1297008991241455,
-    "text_loss": 1.4351346492767334,
-    "flow_loss": 0.847283124923706
+    "loss": 2.791560173034668,
+    "text_loss": 1.4383518695831299,
+    "flow_loss": 0.676604151725769
   },
   {
     "stage": "recap",
     "step": 60,
-    "loss": 2.608584403991699,
-    "text_loss": 1.171854853630066,
-    "flow_loss": 0.7183647751808167
+    "loss": 1.8424029350280762,
+    "text_loss": 1.1395618915557861,
+    "flow_loss": 0.351420521736145
   },
   {
     "stage": "recap",
     "step": 80,
-    "loss": 3.5118634700775146,
-    "text_loss": 0.9280703067779541,
-    "flow_loss": 1.2918965816497803
+    "loss": 1.8679100275039673,
+    "text_loss": 0.9174721837043762,
+    "flow_loss": 0.47521892189979553
   },
   {
     "stage": "recap",
     "step": 100,
-    "loss": 2.7196712493896484,
-    "text_loss": 0.8991937041282654,
-    "flow_loss": 0.9102387428283691
+    "loss": 2.929778814315796,
+    "text_loss": 0.9240415096282959,
+    "flow_loss": 1.00286865234375
   },
   {
     "stage": "recap",
     "step": 120,
-    "loss": 1.1335762739181519,
-    "text_loss": 0.7848832011222839,
-    "flow_loss": 0.17434655129909515
+    "loss": 1.4325169324874878,
+    "text_loss": 0.7645139694213867,
+    "flow_loss": 0.33400148153305054
   },
   {
     "stage": "recap",
     "step": 140,
-    "loss": 2.5729737281799316,
-    "text_loss": 0.5885628461837769,
-    "flow_loss": 0.9922055006027222
+    "loss": 1.4750593900680542,
+    "text_loss": 0.7536309957504272,
+    "flow_loss": 0.3607141971588135
   },
   {
     "stage": "recap",
     "step": 160,
-    "loss": 1.9800522327423096,
-    "text_loss": 0.6485287547111511,
-    "flow_loss": 0.6657617688179016
+    "loss": 1.5970324277877808,
+    "text_loss": 1.017958164215088,
+    "flow_loss": 0.28953713178634644
   },
   {
     "stage": "recap",
     "step": 180,
-    "loss": 1.5937684774398804,
-    "text_loss": 0.6764931678771973,
-    "flow_loss": 0.45863765478134155
+    "loss": 1.6522973775863647,
+    "text_loss": 0.922364354133606,
+    "flow_loss": 0.3649665117263794
   },
   {
     "stage": "recap",
     "step": 200,
-    "loss": 1.588905692100525,
-    "text_loss": 1.1673603057861328,
-    "flow_loss": 0.21077267825603485
+    "loss": 1.3916192054748535,
+    "text_loss": 0.9831637740135193,
+    "flow_loss": 0.2042277306318283
   },
   {
     "stage": "recap",
     "step": 220,
-    "loss": 1.1653995513916016,
-    "text_loss": 0.5840165019035339,
-    "flow_loss": 0.2906915545463562
+    "loss": 1.138975739479065,
+    "text_loss": 0.8730899691581726,
+    "flow_loss": 0.13294290006160736
   },
   {
     "stage": "recap",
     "step": 240,
-    "loss": 2.6636130809783936,
-    "text_loss": 1.3158458471298218,
-    "flow_loss": 0.6738836169242859
+    "loss": 1.502185344696045,
+    "text_loss": 0.8015594482421875,
+    "flow_loss": 0.3503129780292511
   },
   {
     "stage": "recap",
     "step": 260,
-    "loss": 0.5980206727981567,
-    "text_loss": 0.42834123969078064,
-    "flow_loss": 0.08483970910310745
+    "loss": 0.8756294846534729,
+    "text_loss": 0.31519001722335815,
+    "flow_loss": 0.2802197337150574
   },
   {
     "stage": "recap",
     "step": 280,
-    "loss": 1.0666983127593994,
-    "text_loss": 0.47541406750679016,
-    "flow_loss": 0.2956421375274658
+    "loss": 0.779049277305603,
+    "text_loss": 0.3690524101257324,
+    "flow_loss": 0.2049984186887741
   },
   {
     "stage": "recap",
     "step": 300,
-    "loss": 0.8535268306732178,
-    "text_loss": 0.4074874520301819,
-    "flow_loss": 0.22301967442035675
+    "loss": 0.7172976732254028,
+    "text_loss": 0.23720531165599823,
+    "flow_loss": 0.2400461882352829
   },
   {
     "stage": "recap",
     "step": 320,
-    "loss": 5.488016605377197,
-    "text_loss": 0.851451575756073,
-    "flow_loss": 2.3182826042175293
+    "loss": 2.9586782455444336,
+    "text_loss": 0.7825477719306946,
+    "flow_loss": 1.088065266609192
   },
   {
     "stage": "recap",
     "step": 340,
-    "loss": 0.7862600088119507,
-    "text_loss": 0.37299302220344543,
-    "flow_loss": 0.20663349330425262
+    "loss": 1.745682954788208,
+    "text_loss": 0.43472984433174133,
+    "flow_loss": 0.6554765701293945
   },
   {
     "stage": "recap",
     "step": 360,
-    "loss": 1.3418667316436768,
-    "text_loss": 0.7838777899742126,
-    "flow_loss": 0.27899447083473206
+    "loss": 1.5331687927246094,
+    "text_loss": 0.7225565314292908,
+    "flow_loss": 0.4053061604499817
   },
   {
     "stage": "recap",
     "step": 380,
-    "loss": 0.7067349553108215,
-    "text_loss": 0.15938280522823334,
-    "flow_loss": 0.2736760675907135
+    "loss": 1.6411525011062622,
+    "text_loss": 0.14736497402191162,
+    "flow_loss": 0.7468937635421753
   },
   {
     "stage": "recap",
     "step": 400,
-    "loss": 0.4985463619232178,
-    "text_loss": 0.2566893398761749,
-    "flow_loss": 0.12092851102352142
+    "loss": 0.5169777870178223,
+    "text_loss": 0.24374954402446747,
+    "flow_loss": 0.1366141140460968
   },
   {
     "stage": "recap",
     "step": 420,
-    "loss": 0.5810034871101379,
-    "text_loss": 0.24579037725925446,
-    "flow_loss": 0.16760656237602234
+    "loss": 0.6074326038360596,
+    "text_loss": 0.11582576483488083,
+    "flow_loss": 0.24580343067646027
   },
   {
     "stage": "recap",
     "step": 440,
-    "loss": 0.40272825956344604,
-    "text_loss": 0.19410930573940277,
-    "flow_loss": 0.10430946946144104
+    "loss": 0.7017235159873962,
+    "text_loss": 0.3753448724746704,
+    "flow_loss": 0.16318932175636292
   },
   {
     "stage": "recap",
     "step": 460,
-    "loss": 0.7170483469963074,
-    "text_loss": 0.32609468698501587,
-    "flow_loss": 0.19547683000564575
+    "loss": 1.029306411743164,
+    "text_loss": 0.2204148769378662,
+    "flow_loss": 0.40444573760032654
   },
   {
     "stage": "recap",
     "step": 480,
-    "loss": 0.3233756422996521,
-    "text_loss": 0.08512496948242188,
-    "flow_loss": 0.11912534385919571
+    "loss": 0.9227437376976013,
+    "text_loss": 0.21806900203227997,
+    "flow_loss": 0.3523373603820801
   },
   {
     "stage": "recap",
     "step": 500,
-    "loss": 0.625529408454895,
-    "text_loss": 0.23650097846984863,
-    "flow_loss": 0.1945142149925232
+    "loss": 0.7456694841384888,
+    "text_loss": 0.3062073886394501,
+    "flow_loss": 0.21973104774951935
   },
   {
     "stage": "recap",
     "step": 520,
-    "loss": 1.645598292350769,
-    "text_loss": 0.874957263469696,
-    "flow_loss": 0.3853205144405365
+    "loss": 1.2862324714660645,
+    "text_loss": 0.6478312015533447,
+    "flow_loss": 0.3192006051540375
   },
   {
     "stage": "recap",
     "step": 540,
-    "loss": 0.8086838722229004,
-    "text_loss": 0.31192541122436523,
-    "flow_loss": 0.24837924540042877
+    "loss": 0.7051770687103271,
+    "text_loss": 0.23054330050945282,
+    "flow_loss": 0.23731689155101776
   },
   {
     "stage": "recap",
     "step": 560,
-    "loss": 1.9226100444793701,
-    "text_loss": 0.8500325083732605,
-    "flow_loss": 0.5362887382507324
+    "loss": 1.9938247203826904,
+    "text_loss": 1.1214075088500977,
+    "flow_loss": 0.4362086057662964
   },
   {
     "stage": "recap",
     "step": 580,
-    "loss": 2.138998031616211,
-    "text_loss": 0.7942134141921997,
-    "flow_loss": 0.6723923683166504
+    "loss": 1.6423680782318115,
+    "text_loss": 1.1029378175735474,
+    "flow_loss": 0.2697151303291321
   },
   {
     "stage": "recap",
     "step": 600,
-    "loss": 0.504305362701416,
-    "text_loss": 0.13201656937599182,
-    "flow_loss": 0.1861443817615509
+    "loss": 1.3713027238845825,
+    "text_loss": 0.1559465080499649,
+    "flow_loss": 0.6076781153678894
   },
   {
     "stage": "recap",
     "step": 620,
-    "loss": 0.5794067978858948,
-    "text_loss": 0.26564836502075195,
-    "flow_loss": 0.1568792164325714
+    "loss": 0.3619869351387024,
+    "text_loss": 0.06553713977336884,
+    "flow_loss": 0.14822489023208618
   },
   {
     "stage": "recap",
     "step": 640,
-    "loss": 0.8541183471679688,
-    "text_loss": 0.154631569981575,
-    "flow_loss": 0.34974339604377747
+    "loss": 0.5116170048713684,
+    "text_loss": 0.07361899316310883,
+    "flow_loss": 0.2189989984035492
   },
   {
     "stage": "recap",
     "step": 660,
-    "loss": 0.16945119202136993,
-    "text_loss": 0.056921038776636124,
-    "flow_loss": 0.056265074759721756
+    "loss": 0.8094859719276428,
+    "text_loss": 0.07687854766845703,
+    "flow_loss": 0.3663037121295929
   },
   {
     "stage": "recap",
     "step": 680,
-    "loss": 0.47458726167678833,
-    "text_loss": 0.04364287480711937,
-    "flow_loss": 0.21547219157218933
+    "loss": 1.4220974445343018,
+    "text_loss": 0.13382284343242645,
+    "flow_loss": 0.6441373229026794
   },
   {
     "stage": "recap",
     "step": 700,
-    "loss": 0.18534913659095764,
-    "text_loss": 0.05316398665308952,
-    "flow_loss": 0.06609257310628891
+    "loss": 0.40469539165496826,
+    "text_loss": 0.04384676367044449,
+    "flow_loss": 0.18042431771755219
   }
 ]
\ No newline at end of file
diff --git a/outputs/overfit/value_comparison.png b/outputs/overfit/value_comparison.png
index 5b4143a..587fa5c 100644
Binary files a/outputs/overfit/value_comparison.png and b/outputs/overfit/value_comparison.png differ
diff --git a/src/pi06mini/data.py b/src/pi06mini/data.py
index ccb1bdc..4c92fe0 100644
--- a/src/pi06mini/data.py
+++ b/src/pi06mini/data.py
@@ -214,7 +214,7 @@ class Pi06ExperienceDataset(Dataset[dict[str, Any]]):
         history_steps: int = 8,
         action_horizon: int = 16,
         n_step: int = 4,
-        gamma: float = 0.99,
+        gamma: float = 1.0,
         positive_quantile: float = 0.7,
         seed: int = 7,
         action_statistics: ActionStatistics | dict[str, Sequence[float]] | None = None,
@@ -229,6 +229,9 @@ class Pi06ExperienceDataset(Dataset[dict[str, Any]]):
         self.seed = int(seed)
         self.state_dim = self.history_steps * 2
         self.action_dim = 2
+        self.n_step = int(n_step)
+        self.gamma = float(gamma)
+        self.positive_quantile = float(positive_quantile)
 
         episodes: list[ExperienceEpisode] = []
         episode_record_indices: list[int] = []
@@ -275,6 +278,12 @@ class Pi06ExperienceDataset(Dataset[dict[str, Any]]):
             )
         )
 
+        self.value_histories = torch.stack(
+            [self._episode_value_histories(episode) for episode in self.episodes]
+        )
+        terminal = torch.zeros(len(self.episodes), 1, dtype=self.normalized_returns.dtype)
+        self.value_targets = torch.cat((self.normalized_returns, terminal), dim=1)
+
         if action_statistics is None:
             self.action_normalizer = ActionNormalizer.fit(
                 [episode.history_xy for episode in episodes],
@@ -287,6 +296,58 @@ class Pi06ExperienceDataset(Dataset[dict[str, Any]]):
         else:
             self.action_normalizer = ActionNormalizer.from_dict(action_statistics)
 
+    def _episode_value_histories(self, episode: ExperienceEpisode) -> torch.Tensor:
+        """Create observed state windows for every step and the terminal state."""
+
+        path = torch.cat((episode.history_xy, episode.future_xy), dim=0)
+        windows: list[torch.Tensor] = []
+        history_end = len(episode.history_xy)
+        for step in range(len(episode.future_xy) + 1):
+            end = history_end + step
+            start = max(0, end - self.history_steps)
+            window = path[start:end]
+            if len(window) < self.history_steps:
+                pad = window[:1].expand(self.history_steps - len(window), -1)
+                window = torch.cat((pad, window), dim=0)
+            windows.append(window[-self.history_steps :])
+        return torch.stack(windows)
+
+    def relabel_from_value_predictions(
+        self,
+        predicted_values: torch.Tensor,
+        *,
+        force_interventions_positive: bool = True,
+    ) -> None:
+        """Recompute RECAP indicators from the learned critic predictions."""
+
+        predictions = torch.as_tensor(predicted_values, dtype=torch.float32)
+        expected = (len(self.episodes), self.action_horizon + 1)
+        if predictions.shape != expected:
+            raise ValueError(f"predicted_values must have shape {expected}")
+        rewards = stack_episode_field(self.episodes, "rewards")
+        self.advantages = n_step_advantages(
+            rewards,
+            predictions,
+            n_step=self.n_step,
+            gamma=self.gamma,
+        )
+        intervention_masks = stack_episode_field(self.episodes, "intervention_mask").bool()
+        step_interventions = intervention_masks if force_interventions_positive else None
+        self.positive_steps, self.advantage_threshold = quantile_advantage_labels(
+            self.advantages,
+            quantile=self.positive_quantile,
+            intervention_mask=step_interventions,
+        )
+        episode_advantages = self.advantages.mean(dim=-1)
+        episode_interventions = (
+            intervention_masks.any(dim=-1) if force_interventions_positive else None
+        )
+        self.positive_episodes, self.episode_advantage_threshold = quantile_advantage_labels(
+            episode_advantages,
+            quantile=self.positive_quantile,
+            intervention_mask=episode_interventions,
+        )
+
     @property
     def action_statistics(self) -> ActionStatistics:
         return self.action_normalizer.statistics
@@ -334,9 +395,10 @@ class Pi06ExperienceDataset(Dataset[dict[str, Any]]):
             "future_xy": episode.future_xy.clone(),
             "rewards": episode.rewards.clone(),
             "returns": self.normalized_returns[episode_index].clone(),
-            # The critic predicts the beginning-of-episode return for this
-            # minimal trajectory-level buffer; the full vector remains
-            # available for diagnostics and advantage recomputation.
+            "value_histories": self.value_histories[episode_index].clone(),
+            "value_targets": self.value_targets[episode_index].clone(),
+            # Scalar aliases are retained for simple diagnostics; value training
+            # and RECAP relabeling use the complete per-state vectors above.
             "return_target": self.normalized_returns[episode_index, 0].clone(),
             "value_target": self.normalized_returns[episode_index, 0].clone(),
             "advantages": self.advantages[episode_index].clone(),
diff --git a/src/pi06mini/infer.py b/src/pi06mini/infer.py
index a46a3fc..04fbd82 100644
--- a/src/pi06mini/infer.py
+++ b/src/pi06mini/infer.py
@@ -15,6 +15,29 @@ from .tokenizer import MiniTokenizer
 from .value import ValueConfig, ValueModel
 
 
+@torch.no_grad()
+def predict_subtask_token(
+    model: Pi06Mini,
+    tokenizer: MiniTokenizer,
+    image: torch.Tensor,
+    history_xy: torch.Tensor,
+    action_horizon: int,
+    device: torch.device,
+) -> int:
+    """Autoregressively produce the semantic token used by low-level control."""
+
+    query = tokenizer.build_subtask_query(history_xy)[None].to(device)
+    zero_actions = torch.zeros(1, action_horizon, model.config.action_dim, device=device)
+    output = model(
+        image,
+        query,
+        zero_actions,
+        torch.zeros(1, device=device),
+        generated_token_start=query.shape[1],
+    )
+    return int(output.text_logits[0, -1].argmax().cpu())
+
+
 @torch.no_grad()
 def evaluate(checkpoint_path: str) -> tuple[Path, Path]:
     checkpoint=torch.load(checkpoint_path,map_location="cpu",weights_only=False)
@@ -35,13 +58,18 @@ def evaluate(checkpoint_path: str) -> tuple[Path, Path]:
     failure=next(dataset[i] for i in range(len(dataset)) if "stop" in dataset[i]["source_type"])
     image=demo["image"][None].to(device)
 
+    predicted_subtask = predict_subtask_token(
+        recap, tokenizer, image, demo["history_xy"], dataset.action_horizon, device
+    )
+
     def condition(model:Pi06Mini, advantage:str|None, x:torch.Tensor,t:torch.Tensor)->torch.Tensor:
-        seq=tokenizer.build_policy_sequence(
-            demo["history_xy"],demo["reasoning"],demo["actions"],advantage,
-            metadata=None,
-        )
-        ids=seq.input_ids[:seq.action_start_index][None].to(device)
-        return model(image,ids,x,t,generated_token_start=seq.generated_token_start).vector_field
+        ids = tokenizer.build_action_condition(
+            demo["history_xy"], predicted_subtask, advantage
+        )[None].to(device)
+        return model(
+            image, ids, x, t,
+            generated_token_start=tokenizer.build_subtask_query(demo["history_xy"]).numel(),
+        ).vector_field
 
     def sample_field(fn,seed:int=2):
         torch.manual_seed(seed)
@@ -77,16 +105,38 @@ def evaluate(checkpoint_path: str) -> tuple[Path, Path]:
 
     value_rows=[]
     for i in range(len(dataset)):
-        s=dataset[i];ids=tokenizer.build_value_input(s["history_xy"],s["source_type"])[None].to(device)
-        prediction=float(value(s["image"][None].to(device),ids).expectation.cpu())
-        value_rows.append({"source_type":s["source_type"],"target":float(s["value_target"]),"prediction":prediction})
-    fig,ax=plt.subplots(figsize=(8,4));x=np.arange(len(value_rows));w=.35
-    ax.bar(x-w/2,[r["target"] for r in value_rows],w,label="return target")
-    ax.bar(x+w/2,[r["prediction"] for r in value_rows],w,label="value prediction")
-    ax.set_xticks(x,[r["source_type"] for r in value_rows],rotation=20,ha="right");ax.set_ylim(-1.05,.05);ax.grid(True,axis="y",alpha=.3);ax.legend();ax.set_title("201-bin distributional value overfit")
+        s=dataset[i]
+        histories=s["value_histories"]
+        ids=torch.stack([tokenizer.build_value_input(history) for history in histories]).to(device)
+        images=s["image"][None].expand(len(histories),-1,-1,-1).to(device)
+        predictions=value(images,ids).expectation.cpu()
+        targets=s["value_targets"]
+        value_rows.append({
+            "source_type":s["source_type"],
+            "targets":targets.tolist(),
+            "predictions":predictions.tolist(),
+            "curve_mse":float(torch.mean((predictions-targets).square())),
+        })
+    fig,axes=plt.subplots(2,2,figsize=(10,7),sharex=True,sharey=True)
+    for ax,row in zip(axes.flat,value_rows):
+        ax.plot(row["targets"],"o-",label="return target")
+        ax.plot(row["predictions"],"x--",label="value prediction")
+        ax.set_title(f'{row["source_type"]} MSE={row["curve_mse"]:.3f}')
+        ax.set_ylim(-1.05,.05);ax.grid(True,alpha=.3)
+    axes[0,0].legend(fontsize=8);fig.supxlabel("episode step");fig.supylabel("normalized value")
     value_path=outdir/"value_comparison.png";fig.tight_layout();fig.savefig(value_path,dpi=170);plt.close(fig)
-    write_json(outdir/"evaluation.json",{"trajectory_metrics":metrics,"values":value_rows})
-    print({"trajectory_metrics":metrics,"values":value_rows})
+    try:
+        predicted_subtask_label = tokenizer.subtasks.decode(predicted_subtask)
+    except ValueError:
+        predicted_subtask_label = f"token_{predicted_subtask}"
+    evaluation = {
+        "predicted_subtask": predicted_subtask_label,
+        "target_subtask": tokenizer.subtasks.classify(demo["reasoning"]),
+        "trajectory_metrics": metrics,
+        "values": value_rows,
+    }
+    write_json(outdir/"evaluation.json",evaluation)
+    print(evaluation)
     return trajectory_path,value_path
 
 
diff --git a/src/pi06mini/tokenizer.py b/src/pi06mini/tokenizer.py
index fec7dbc..caf346f 100644
--- a/src/pi06mini/tokenizer.py
+++ b/src/pi06mini/tokenizer.py
@@ -191,15 +191,18 @@ class MiniTokenizer:
         parts.append(query)
         return torch.cat(parts)
 
-    def build_value_input(self, history_xy: Tensor, metadata: str | None = None) -> Tensor:
-        """Return image marker plus discretized state, without policy query tokens."""
+    def build_value_input(self, history_xy: Tensor) -> Tensor:
+        """Return observation tokens for the critic without outcome metadata.
+
+        RECAP's value function estimates ``V(o, task)``.  Experience provenance
+        and success/failure labels are training targets, not observable inputs;
+        including them would let the critic solve the toy buffer by leakage.
+        The single-task minimal experiment keeps the task command implicit.
+        """
 
         coordinate_tokens = self.coordinates.encode(history_xy).reshape(-1).cpu()
         specials = torch.tensor([self.BOS, self.IMAGE, self.STATE], dtype=torch.long)
-        parts = [specials, coordinate_tokens]
-        metadata_token = self.encode_metadata(metadata)
-        if metadata_token is not None: parts.append(metadata_token.reshape(1))
-        return torch.cat(parts)
+        return torch.cat((specials, coordinate_tokens))
 
     # Backward-compatible name used by the pi0.5 data utilities.
     build_prefix = build_observation
@@ -216,6 +219,28 @@ class MiniTokenizer:
             return None
         raise ValueError("advantage must be positive, negative, omitted, bool, or None")
 
+    def build_subtask_query(self, history_xy: Tensor) -> Tensor:
+        """Observation prefix whose final token predicts the next subtask."""
+
+        return self.build_observation(history_xy, metadata=None)
+
+    def build_action_condition(
+        self,
+        history_xy: Tensor,
+        subtask_token: int | Tensor,
+        advantage: Advantage | bool | None,
+    ) -> Tensor:
+        """Build the inference prefix ``observation -> subtask -> advantage``."""
+
+        pieces = [
+            self.build_observation(history_xy, metadata=None),
+            torch.as_tensor(subtask_token, dtype=torch.long).reshape(1).cpu(),
+        ]
+        advantage_token = self.encode_advantage(advantage)
+        if advantage_token is not None:
+            pieces.append(advantage_token.reshape(1))
+        return torch.cat(pieces)
+
     @staticmethod
     def _orthonormal_dct_matrix(
         horizon: int, *, device: torch.device, dtype: torch.dtype
diff --git a/src/pi06mini/train.py b/src/pi06mini/train.py
index f9d033d..9df30f8 100644
--- a/src/pi06mini/train.py
+++ b/src/pi06mini/train.py
@@ -63,6 +63,27 @@ def policy_step(
     }
 
 
+@torch.no_grad()
+def predict_dataset_values(
+    value: ValueModel,
+    dataset: Pi06ExperienceDataset,
+    tokenizer: MiniTokenizer,
+    device: torch.device,
+) -> torch.Tensor:
+    """Run the learned critic over every state used for RECAP relabeling."""
+
+    value.eval()
+    predictions: list[torch.Tensor] = []
+    for episode_index in range(len(dataset.episodes)):
+        sample = dataset[episode_index * dataset.augmentation_factor]
+        histories = sample["value_histories"]
+        ids = torch.stack([tokenizer.build_value_input(history) for history in histories])
+        images = sample["image"][None].expand(len(histories), -1, -1, -1)
+        output = value(images.to(device), ids.to(device))
+        predictions.append(output.expectation.cpu())
+    return torch.stack(predictions)
+
+
 def train(config_path: str) -> Path:
     cfg = load_config(config_path)
     root = Path(__file__).resolve().parents[2]
@@ -73,7 +94,8 @@ def train(config_path: str) -> Path:
         manifest, image_size=cfg["data"]["image_size"],
         augmentation_factor=cfg["experience"]["augmentation_factor"],
         action_horizon=cfg["data"]["action_horizon"], n_step=cfg["experience"]["n_step"],
-        gamma=cfg["experience"]["gamma"], positive_quantile=cfg["experience"]["positive_quantile"],
+        gamma=cfg["experience"].get("gamma", 1.0),
+        positive_quantile=cfg["experience"]["positive_quantile"],
     )
     tokenizer = MiniTokenizer()
     base, value = build_models(cfg, tokenizer)
@@ -99,11 +121,31 @@ def train(config_path: str) -> Path:
     value.train()
     for step in range(1, cfg["train"]["value_steps"] + 1):
         sample = dataset[(step - 1) % len(dataset)]
-        ids = tokenizer.build_value_input(sample["history_xy"], sample["source_type"])[None].to(device)
-        output = value(sample["image"][None].to(device), ids, target_values=sample["value_target"][None].to(device))
+        time_index = (step - 1) % (dataset.action_horizon + 1)
+        history_xy = sample["value_histories"][time_index]
+        ids = tokenizer.build_value_input(history_xy)[None].to(device)
+        target = sample["value_targets"][time_index][None].to(device)
+        output = value(sample["image"][None].to(device), ids, target_values=target)
         value_opt.zero_grad(set_to_none=True); output.loss.backward(); value_opt.step()
         if step == 1 or step % cfg["train"]["log_every"] == 0:
-            row={"stage":"value","step":step,"loss":float(output.loss.detach().cpu()),"prediction":float(output.expectation.detach().cpu()),"target":float(sample["value_target"])};history.append(row);print(json.dumps(row))
+            row={"stage":"value","step":step,"loss":float(output.loss.detach().cpu()),"prediction":float(output.expectation.detach().cpu()),"target":float(target.cpu())};history.append(row);print(json.dumps(row))
+
+    critic_values = predict_dataset_values(value, dataset, tokenizer, device)
+    dataset.relabel_from_value_predictions(
+        critic_values,
+        force_interventions_positive=cfg["recap"].get(
+            "force_interventions_positive", True
+        ),
+    )
+    write_json(outdir / "critic_relabeling.json", {
+        "predicted_values": critic_values.tolist(),
+        "advantage_threshold": float(dataset.advantage_threshold),
+        "episode_advantage_threshold": float(dataset.episode_advantage_threshold),
+        "episode_labels": [
+            "positive" if bool(label) else "negative"
+            for label in dataset.positive_episodes
+        ],
+    })
 
     def fit_policy(stage: str, steps: int, conditioned: bool) -> Pi06Mini:
         model, _ = build_models(cfg, tokenizer); model.load_state_dict(base_state); model.to(device).train()
diff --git a/tests/test_experience_data.py b/tests/test_experience_data.py
index a3d0a20..69a82f0 100644
--- a/tests/test_experience_data.py
+++ b/tests/test_experience_data.py
@@ -71,6 +71,8 @@ def test_experience_dataset_contract_and_intervention_label(tmp_path: Path) -> N
     assert intervention["returns"].shape == (16,)
     assert intervention["return_target"].shape == ()
     assert intervention["value_target"].shape == ()
+    assert intervention["value_histories"].shape == (17, 8, 2)
+    assert intervention["value_targets"].shape == (17,)
     assert intervention["advantages"].shape == (16,)
     assert intervention["advantage_label"] == "positive"
     assert intervention["reasoning"] == "Steer left."
@@ -78,3 +80,21 @@ def test_experience_dataset_contract_and_intervention_label(tmp_path: Path) -> N
         intervention["actions"], intervention["history_xy"]
     )
     torch.testing.assert_close(decoded, intervention["future_xy"], atol=1e-5, rtol=1e-5)
+
+
+def test_learned_value_predictions_drive_advantage_relabeling(tmp_path: Path) -> None:
+    manifest = _make_source_project(tmp_path / "source")
+    dataset = Pi06ExperienceDataset(manifest, image_size=32)
+    before = dataset.advantages.clone()
+    predicted = torch.zeros(len(dataset.episodes), dataset.action_horizon + 1)
+    predicted[:, :-1] = torch.linspace(-0.9, -0.1, dataset.action_horizon)
+
+    dataset.relabel_from_value_predictions(predicted)
+
+    assert not torch.allclose(dataset.advantages, before)
+    intervention_index = next(
+        index
+        for index, episode in enumerate(dataset.episodes)
+        if episode.source_type == "intervention_correction"
+    )
+    assert bool(dataset.positive_episodes[intervention_index])
diff --git a/tests/test_tokenizer.py b/tests/test_tokenizer.py
index 8bd686c..a99aeb8 100644
--- a/tests/test_tokenizer.py
+++ b/tests/test_tokenizer.py
@@ -37,6 +37,28 @@ def test_omitted_advantage_is_absent_not_a_third_condition_token() -> None:
     assert tokenizer.ADVANTAGE_NEGATIVE not in sequence.input_ids
 
 
+def test_value_input_does_not_encode_experience_outcome() -> None:
+    tokenizer = MiniTokenizer()
+    history = torch.zeros(8, 2)
+    value_input = tokenizer.build_value_input(history)
+
+    assert tokenizer.METADATA_DEMO not in value_input
+    assert tokenizer.METADATA_UNDERSTEER not in value_input
+    assert tokenizer.METADATA_STOP not in value_input
+    assert tokenizer.METADATA_INTERVENTION not in value_input
+
+
+def test_predicted_subtask_can_be_connected_to_action_condition() -> None:
+    tokenizer = MiniTokenizer()
+    history = torch.zeros(8, 2)
+    predicted = tokenizer.subtasks.encode("steer_left")
+    condition = tokenizer.build_action_condition(history, predicted, "positive")
+
+    assert condition[-2].item() == predicted
+    assert condition[-1].item() == tokenizer.ADVANTAGE_POSITIVE
+    assert tokenizer.ACTION_START not in condition
+
+
 def test_labels_predict_subtask_and_fast_but_not_external_advantage() -> None:
     tokenizer = MiniTokenizer()
     positive = tokenizer.build_policy_sequence(