diff --git a/docs/experiment_round33_both.json b/docs/experiment_round33_both.json
new file mode 100644
index 0000000..ac7f50d
--- /dev/null
+++ b/docs/experiment_round33_both.json
@@ -0,0 +1,70 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round33_both_rows.json",
+    "out": "docs/experiment_round33_both.json",
+    "trace": "ml/runs/round33_both.pt",
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.5,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.51953125,
+    "collision_free_success": 0.51953125,
+    "time_ok_1p3": 0.26171875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.5444939136505127,
+    "time_mean_s": 2.4729323387145996,
+    "brakes_mean": 19.97265625,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.2868920564651489,
+    "err_fail_med": 202.12567138671875,
+    "err_success_med": 0.22085000574588776,
+    "dls_share": 0.62109375,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118
+  },
+  "simulation_wall_s": 277.54961390000244
+}
diff --git a/docs/experiment_round33_cap.json b/docs/experiment_round33_cap.json
new file mode 100644
index 0000000..137a473
--- /dev/null
+++ b/docs/experiment_round33_cap.json
@@ -0,0 +1,70 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round33_cap_rows.json",
+    "out": "docs/experiment_round33_cap.json",
+    "trace": "ml/runs/round33_cap.pt",
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5234375,
+    "collision_free_success": 0.5234375,
+    "time_ok_1p3": 0.265625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.615795850753784,
+    "time_mean_s": 2.486567258834839,
+    "brakes_mean": 19.5546875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.27920001745224,
+    "err_fail_med": 204.71360778808594,
+    "err_success_med": 0.21751396358013153,
+    "dls_share": 0.61328125,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118
+  },
+  "simulation_wall_s": 325.18856850000157
+}
diff --git a/docs/experiment_round33_gpu_cycle.json b/docs/experiment_round33_gpu_cycle.json
new file mode 100644
index 0000000..43c665e
--- /dev/null
+++ b/docs/experiment_round33_gpu_cycle.json
@@ -0,0 +1,202 @@
+{
+  "device": "NVIDIA GeForce RTX 3090",
+  "torch": "2.10.0+cu130",
+  "act_scale": 1.0,
+  "cfg": {},
+  "B": 1,
+  "commit_n": 2,
+  "fp64_note": "3090 fp64 \u541e\u5410\u7ea6 fp32 \u7684 1/54\uff1b\u8ba4\u8bc1\u76ee\u524d\u8981\u6c42 fp64",
+  "inference": {
+    "sample_K1_steps1": {
+      "p50": 7.453,
+      "p95": 9.912,
+      "p99": 10.766,
+      "max": 11.109,
+      "reps": 120
+    },
+    "sample_K8_steps1": {
+      "p50": 8.208,
+      "p95": 10.685,
+      "p99": 11.22,
+      "max": 15.224,
+      "reps": 120
+    },
+    "sample_K1_steps2": {
+      "p50": 13.149,
+      "p95": 17.621,
+      "p99": 18.438,
+      "max": 19.296,
+      "reps": 120
+    },
+    "sample_K8_steps2": {
+      "p50": 14.167,
+      "p95": 19.241,
+      "p99": 24.77,
+      "max": 25.703,
+      "reps": 120
+    },
+    "sample_K1_steps4": {
+      "p50": 19.193,
+      "p95": 22.338,
+      "p99": 23.666,
+      "max": 25.787,
+      "reps": 120
+    },
+    "sample_K8_steps4": {
+      "p50": 21.509,
+      "p95": 24.732,
+      "p99": 35.655,
+      "max": 41.514,
+      "reps": 120
+    },
+    "sample_K1_steps8": {
+      "p50": 40.658,
+      "p95": 55.162,
+      "p99": 60.394,
+      "max": 60.398,
+      "reps": 120
+    },
+    "sample_K8_steps8": {
+      "p50": 40.416,
+      "p95": 47.997,
+      "p99": 62.49,
+      "max": 71.976,
+      "reps": 120
+    },
+    "sample_K1_steps16": {
+      "p50": 72.429,
+      "p95": 90.188,
+      "p99": 107.536,
+      "max": 112.674,
+      "reps": 120
+    },
+    "sample_K8_steps16": {
+      "p50": 79.893,
+      "p95": 99.933,
+      "p99": 105.746,
+      "max": 106.078,
+      "reps": 120
+    }
+  },
+  "encode_once": {
+    "p50": 2.265,
+    "p95": 4.001,
+    "p99": 6.086,
+    "max": 6.136,
+    "reps": 120
+  },
+  "tiny_op_ms": 0.028,
+  "sync_only_ms": 0.005,
+  "components": {
+    "features_f32": {
+      "p50": 12.443,
+      "p95": 14.227,
+      "p99": 19.678,
+      "max": 20.531,
+      "reps": 120
+    },
+    "endpoint_score_K8_f32": {
+      "p50": 4.278,
+      "p95": 5.841,
+      "p99": 6.322,
+      "max": 8.883,
+      "reps": 120
+    },
+    "dls_fb_1step_f64": {
+      "p50": 7.191,
+      "p95": 12.26,
+      "p99": 14.642,
+      "max": 15.384,
+      "reps": 120
+    },
+    "dls_fb_1step_f32": {
+      "p50": 7.619,
+      "p95": 10.142,
+      "p99": 11.597,
+      "max": 12.022,
+      "reps": 120
+    },
+    "discrete_cert_4pts_f32": {
+      "p50": 8.214,
+      "p95": 13.365,
+      "p99": 14.235,
+      "max": 14.991,
+      "reps": 120
+    },
+    "cert_n8_f64_2roots": {
+      "p50": 11.265,
+      "p95": 14.64,
+      "p99": 16.788,
+      "max": 17.761,
+      "reps": 120
+    },
+    "cert_n16_f64_2roots": {
+      "p50": 14.574,
+      "p95": 16.731,
+      "p99": 18.459,
+      "max": 22.346,
+      "reps": 120
+    },
+    "cert_n32_f64_2roots": {
+      "p50": 21.528,
+      "p95": 24.746,
+      "p99": 28.037,
+      "max": 28.066,
+      "reps": 120
+    }
+  },
+  "cycle_K8_steps16_discrete": {
+    "p50": 116.474,
+    "p95": 134.474,
+    "p99": 142.593,
+    "max": 152.768,
+    "reps": 120
+  },
+  "cycle_K8_steps4_discrete": {
+    "p50": 55.919,
+    "p95": 65.767,
+    "p99": 74.809,
+    "max": 78.059,
+    "reps": 120
+  },
+  "cycle_K1_steps4_discrete": {
+    "p50": 51.087,
+    "p95": 58.614,
+    "p99": 62.326,
+    "max": 63.919,
+    "reps": 120
+  },
+  "cycle_K1_steps1_discrete": {
+    "p50": 34.493,
+    "p95": 42.078,
+    "p99": 43.906,
+    "max": 50.268,
+    "reps": 120
+  },
+  "cycle_K8_steps16_cert_n8": {
+    "p50": 132.027,
+    "p95": 150.463,
+    "p99": 160.83,
+    "max": 192.586,
+    "reps": 120
+  },
+  "cycle_K8_steps4_cert_n8": {
+    "p50": 64.812,
+    "p95": 79.192,
+    "p99": 89.629,
+    "max": 95.837,
+    "reps": 120
+  },
+  "cycle_K1_steps4_cert_n16": {
+    "p50": 61.967,
+    "p95": 73.219,
+    "p99": 79.053,
+    "max": 83.18,
+    "reps": 120
+  },
+  "f32_vs_f64": {
+    "points": 3,
+    "max_abs_err_mm": 0.0,
+    "sign_flips": 0
+  }
+}
\ No newline at end of file
diff --git a/docs/experiment_round33_gpu_cycle_steps1.json b/docs/experiment_round33_gpu_cycle_steps1.json
new file mode 100644
index 0000000..ad77aa1
--- /dev/null
+++ b/docs/experiment_round33_gpu_cycle_steps1.json
@@ -0,0 +1,143 @@
+{
+  "device": "NVIDIA GeForce RTX 3090",
+  "B": 1,
+  "steps_used_by_eval": 1,
+  "note": "eval_cl \u4ece\u4e0d\u4f20 steps\uff0crollout \u9ed8\u8ba4 steps=1\uff08MeanFlow \u4e00\u6b65\u751f\u6210\uff09\uff1b\u6240\u4ee5\u771f\u5b9e\u5468\u671f\u6309 steps=1 \u6d4b",
+  "parts": {
+    "features_total": {
+      "p50": 12.589,
+      "p95": 18.334,
+      "p99": 20.217,
+      "max": 23.649,
+      "reps": 120
+    },
+    "features_gm_groups_only": {
+      "p50": 6.817,
+      "p95": 8.771,
+      "p99": 10.254,
+      "max": 17.169,
+      "reps": 120
+    },
+    "features_gaps40_all_groups": {
+      "p50": 7.245,
+      "p95": 10.501,
+      "p99": 13.794,
+      "max": 14.508,
+      "reps": 120
+    },
+    "features_fk_only": {
+      "p50": 3.708,
+      "p95": 5.024,
+      "p99": 5.991,
+      "max": 6.301,
+      "reps": 120
+    },
+    "sample_k_K8_steps1": {
+      "p50": 7.799,
+      "p95": 11.655,
+      "p99": 13.342,
+      "max": 16.212,
+      "reps": 120
+    },
+    "sample_k_K1_steps1": {
+      "p50": 7.205,
+      "p95": 8.935,
+      "p99": 9.615,
+      "max": 13.92,
+      "reps": 120
+    },
+    "sample_k_K4_steps1": {
+      "p50": 7.834,
+      "p95": 10.354,
+      "p99": 14.744,
+      "max": 15.241,
+      "reps": 120
+    },
+    "dls_fb_f64": {
+      "p50": 6.729,
+      "p95": 9.297,
+      "p99": 11.186,
+      "max": 12.064,
+      "reps": 120
+    },
+    "discrete_cert_4pts": {
+      "p50": 8.603,
+      "p95": 11.181,
+      "p99": 12.464,
+      "max": 15.799,
+      "reps": 120
+    },
+    "cert_uniform_n8_2roots_f64": {
+      "p50": 11.479,
+      "p95": 19.221,
+      "p99": 20.942,
+      "max": 21.183,
+      "reps": 120
+    },
+    "cert_uniform_n16_2roots_f64": {
+      "p50": 16.138,
+      "p95": 21.972,
+      "p99": 23.898,
+      "max": 25.793,
+      "reps": 120
+    }
+  },
+  "cycles": {
+    "cycle_K8_full_discrete": {
+      "p50": 40.676,
+      "p95": 51.742,
+      "p99": 59.782,
+      "max": 62.074,
+      "reps": 120
+    },
+    "cycle_K4_full_discrete": {
+      "p50": 43.997,
+      "p95": 56.145,
+      "p99": 59.905,
+      "max": 65.882,
+      "reps": 120
+    },
+    "cycle_K1_full_discrete": {
+      "p50": 37.87,
+      "p95": 45.754,
+      "p99": 49.394,
+      "max": 53.139,
+      "reps": 120
+    },
+    "cycle_K8_noscore_discrete": {
+      "p50": 37.427,
+      "p95": 49.602,
+      "p99": 60.777,
+      "max": 70.139,
+      "reps": 120
+    },
+    "cycle_K8_full_cert_n8": {
+      "p50": 43.984,
+      "p95": 65.795,
+      "p99": 75.53,
+      "max": 95.452,
+      "reps": 120
+    },
+    "cycle_K8_full_cert_n16": {
+      "p50": 51.054,
+      "p95": 74.862,
+      "p99": 95.768,
+      "max": 106.48,
+      "reps": 120
+    },
+    "cycle_K8_fused_estimate": {
+      "p50": 24.054,
+      "p95": 26.595,
+      "p99": 30.17,
+      "max": 33.284,
+      "reps": 120
+    },
+    "cycle_K8_fused_cert_n8": {
+      "p50": 34.274,
+      "p95": 46.329,
+      "p99": 72.737,
+      "max": 79.022,
+      "reps": 120
+    }
+  }
+}
\ No newline at end of file
diff --git a/docs/experiment_round33_margin.json b/docs/experiment_round33_margin.json
new file mode 100644
index 0000000..f49486c
--- /dev/null
+++ b/docs/experiment_round33_margin.json
@@ -0,0 +1,70 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round33_margin_rows.json",
+    "out": "docs/experiment_round33_margin.json",
+    "trace": "ml/runs/round33_margin.pt",
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 0,
+    "gap_margin": 0.5,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5625,
+    "collision_free_success": 0.5625,
+    "time_ok_1p3": 0.37890625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.448458433151245,
+    "time_mean_s": 2.2555553913116455,
+    "brakes_mean": 20.046875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.18239997327327728,
+    "err_fail_med": 225.54278564453125,
+    "err_success_med": 0.18544802069664001,
+    "dls_share": 0.62890625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.6600004434585571,
+    "speed_ratio_p95": 1.660003662109375,
+    "speed_ratio_max": 1.660003662109375,
+    "over_vmax_rate": 1.0,
+    "accel_ratio_p95": 1.5889430046081543,
+    "accel_ratio_max": 1.6600035429000854
+  },
+  "simulation_wall_s": 298.24517190000915
+}
diff --git a/docs/探索日志.md b/docs/探索日志.md
index 4aa9ebf..f38458b 100644
--- a/docs/探索日志.md
+++ b/docs/探索日志.md
@@ -650,9 +650,50 @@
 - **两个新开关**：`inc_cap=1` 在**反馈叠加之后**再按每轴 VMAX·DT 限幅（此前 `inc=clamp(act,±1)·astep` 之后又加 `gain×fb`，所以 1.66× 是结构性的）；`gap_margin>0` 让安全层要求间隙 > 余量而非 >0，并把宽相位 cap 自动抬到 `margin+0.5`——不抬的话 `gaps(allq, 0.5)` 的截顶值 0.5 会让任何 >0.5 的余量门限恒假。`col`（真实碰撞计数）仍按 `gap≤0` 判，不随余量改变，所以两者不会互相顶替。
 - **部署侧同病**（只读核查，未改 JS）：`webapp/js/ik-policy.js:355` 也是 `inc[s][j] += fbGain*fbStep[j]` 在 ±1·STEP 之后叠加，所以浏览器控制器同样能超 VMAX；JS 的余量是 `o.margin=5` 的**投影式**护栏（`project()` 每轴 ≤projMax，只处理 gap<margin 的近距对），与本轮的"低于余量就刹停"不是同一机制；`ACC=5.0` 只用于刹停段长度，不是每步加速度上限（§9.2 也没给数值化的加速度门槛，所以 `accel_ratio` 只作变体间比较）。
 - **间隙分布的口径坑**：第30轮 `plans['gaps']` 是在 cap=0.5 下记录的，所以"最小间隙 ≥1 mm 的题占 0"是截顶假象。可说的是：逐题最小记录间隙 p5 **0.0005**、p25 0.0064、中位 **0.046** mm，57.4% 的题曾低于 0.5 mm——即策略实际贴障碍飞，而查看器护栏是 5 mm。要把 0.5 以上那段分开，必须抬 cap 重算（本轮后续做）。
-- **四组单变量对照**（同 256 题、同 seed、其余按第30轮配方）：base / `inc_cap=1` / `gap_margin=0.5` / 两者同开。结果与加密复核见本节末尾的补充提交。
+- **四组单变量对照**（同 256 题、同 seed、其余按第30轮配方）：base / `inc_cap=1` / `gap_margin=0.5` / 两者同开。结果见下表。
+
+- **四组单变量对照结果**（同 256 题、同 seed、K=8、commit 2、fb 2.2、fb_block_gain 0.5、dls_safety 1、select endpoint、replan empty、partial 1；`docs/experiment_round33_{base,cap,margin,both}.json`）：
+
+  | 变体 | 到达 | 无碰撞到达 | time_ok_1p3 | 真实碰撞 | 刹停/题 | 速度比 p50 | 超 VMAX 题占比 | 2 s 误差比 |
+  |---|---|---|---|---|---|---|---|---|
+  | base | 147/256 = 0.5742 | 0.5742 | 0.3906 | 0 | 19.25 | 1.660 | 1.000 | 0.2100 |
+  | `--inc_cap 1` | 134/256 = 0.5234 | 0.5234 | 0.2656 | 0 | 19.55 | 1.000 | 0.000 | 0.2790 |
+  | `--gap_margin 0.5` | 144/256 = 0.5625 | 0.5625 | 0.3789 | 0 | 20.05 | 1.660 | 1.000 | 0.1817 |
+  | 两者同开 | 133/256 = 0.5195 | 0.5195 | 0.2617 | 0 | 19.97 | 1.000 | 0.000 | 0.2870 |
+
+  - **速度限幅是有代价的硬约束**：每轴增量限回 VMAX 后到达 −5.1 pp、time_ok −12.5 pp、2 s 误差比 0.210→0.279。也就是说 base 的 57.4% 里**有一部分是用超速换来的**：反馈在限幅之外叠加，实际替模型补了速度。不能把 base 当作"满足速度约束的成绩"。
+  - **机理**：base 速度比 p50=1.660 恰等于 `1+0.3×fb_gain 2.2`；限幅后刹停几乎不变（19.25→19.55），但 time_ok 掉得比到达更多，说明损失集中在**逼近目标末段被剥夺了超速修正**，不是更容易刹停。所以出路不是调低增益（第29/30轮已测：增益小到达就掉），而是把该由前馈承担的速度还给模型——在限幅后的分布上重采数据（DAgger，需批准），或用不违反速度约束的前馈（#14 的余量投影 + 按时限打分）。
+  - **0.5 mm 余量门很便宜**：只少 3 题、刹停 +0.8/题，2 s 误差比反而更好（0.182）。但这不等于裕度够用——它只说明"要求 0.5 mm 不明显拖慢"。5 mm（查看器护栏）是否同样便宜，取决于策略能否找到 ≥5 mm 的路径，必须抬 cap 重算真实间隙分布才能回答（第30轮 `plans['gaps']` 在 0.5 mm 截顶，算不出这一条）。
+  - 四组 `collision_rate` 都是 0，但这仍是端点/中点离散口径；`gap_margin` 不改 `col` 判据（仍 `gap≤0`），余量与碰撞计数是两条独立的线，不互相顶替。
 - **回归**：`python -m unittest tests.test_eval_cl` **30 项通过**（新增 4 项）：未限幅时实测速度比确实 =3.4/1.2>1（灵敏度）、限幅后恰为 1.0、0.3 mm 间隙在余量 0 时通过而余量 0.5 时刹停、5 mm 间隙在余量 0.5 时仍通过（证明 cap 抬升生效，否则该测试会恒假）。
 
+- **GPU 上的一个控制周期（按 GPU 口径实测）**：诊断档与逐项分解见 `docs/experiment_round33_gpu_cycle.json`，真实配方（steps=1）见 `docs/experiment_round33_gpu_cycle_steps1.json`，RTX 3090、B=1、float32 推理 + float64 认证，每项 20 次预热 + 120 次计时并 `cuda.synchronize`。
+
+  - **前提更正（重要）**：`eval_cl` 从不传 `steps`，`rollout` 默认 `steps=1`，而 `ml/model/policy.py:156` 写明"steps=1 即一步生成，steps>1 是诊断用"。所以第27轮以来的所有评估成绩一直是**单步推理**；本节先用 steps=16 测出的"当前配方 116 ms ≈ 8.6 Hz"是口径错误，已作废并在下面用真实配方重测。
+  - **真实配方（steps=1, K=8）的分解**：
+
+  | 环节 | P50 | P95 | P99 |
+  |---|---|---|---|
+  | 推理 `sample_k` steps=1，K=1 / 4 / 8 | 7.11 / 7.08 / 7.19 | 9.45 / 10.17 / 9.95 | 11.91 / 14.00 / 11.34 |
+  | 特征 83 维（总 / 其中宽相位间隙 / 其中 FK） | 13.11 / 8.18 / 4.00 | 19.57 / 16.57 / 7.51 | 21.17 / 19.43 / 8.47 |
+  | 目标反馈 DLS 1 步 | 6.53 | 9.32 | 12.62 |
+  | 现有离散安全层 4 点（cap 1.0） | 7.41 | 11.34 | 13.80 |
+  | `certify_uniform` 2 根 n=8 / n=16 | 9.70 / 16.07 | 14.40 / 20.95 | 15.75 / 24.42 |
+
+  | 组合周期（同一位姿，顺序测；跨行比较受漂移影响，只作量级） | P50 | P95 | P99 |
+  |---|---|---|---|
+  | K=8 + 离散安全层（真实配方） | **36.68 → 27.3 Hz** | 43.27 | 49.17 |
+  | K=4 / K=1 + 离散安全层 | 37.71 / 40.36 | 55.17 / 48.46 | 76.53 / 52.06 |
+  | K=8 + 连续认证 n=8 | 45.34 → 22.1 Hz | 49.68 | 52.20 |
+  | K=8 + 连续认证 n=16 | 49.55 → 20.2 Hz | 56.85 | 59.06 |
+  | （对照）诊断档 steps=16 的同一周期 | 116.5 | 134.5 | 142.6 |
+
+  - **归因更正（先证仪表再下结论）**：空算子 P50 0.028 ms、空 `synchronize` 0.005 ms、一次 `encode` 2.27 ms，所以**不是启动/同步地板**。先前口算的"5–10 ms 是 GPU 启动开销"错了，在此更正：特征 13.1 ms、认证 7.4 ms、DLS 6.5 ms 是 B=1 下真实的小算子工作量——即 **GPU 在单臂批量=1 时利用率极低**，而不是被框架拖慢。
+  - **推理不是瓶颈，几何才是**：单步推理 7.2 ms，且候选数几乎免费（K 1→8 不涨）；一个周期里几何类工作合计约 27 ms（特征宽相位 8.2 + 特征 FK 4.0 + 认证 7.4 + DLS 6.5），并且这几次调用各自重复做 `cum_xf` 与宽相位筛选。所以 50 Hz 的路径是**合并几何查询**（`features()` 已支持 `gm=`/`xf=` 复用，`gm_groups` 文档保证与整表 `gaps` 数值一致），而不是继续压推理。
+  - **连续认证的增量成本可接受**：在同一位姿上 n=8 只比现有离散层贵约 8.7 ms（整周期 36.7→45.3 ms），n=16 贵约 12.9 ms。也就是说"段内连续证明"在 GPU 上不是奢侈项，真正卡住 20 ms 的是那 ~27 ms 的重复几何。合并后再看是否进得了 20 ms——这一步正在做交错复测（跨行顺序测的数字受漂移影响，不作最终结论）。
+  - **50 Hz 结论（当前状态）**：真实配方一个规划周期 **36.7 ms（P99 49.2）≈ 27 Hz**，仍不满足"P99<20 ms"；且必须记住这是在**不限幅**（可超速 1.66×）的策略分布下测的——限幅后的到达率代价见上表，两者不能混着谈。
+  - f32 与 f64 的间隙在 3 个点上完全相同（最大偏差 0.0 mm、无符号翻转），但 **3 个点不足以支持"认证可以改用 float32"**，需按真实轨迹批量对比才能下结论。
+
 ## 分支清理（2026-10-20）
 
 - `feat/fb-mix`（8e5431c）：反馈混合方案的中间分支，**已验证是 `exp/dagger-deploy` 的祖先**（所有提交都在后者历史里），本地/远端分支删除，内容保留在 exp/dagger-deploy。
diff --git a/ml/common/kin_torch.py b/ml/common/kin_torch.py
index b894ecc..205a6a9 100644
--- a/ml/common/kin_torch.py
+++ b/ml/common/kin_torch.py
@@ -280,6 +280,8 @@ class Kin:
         """
         if self.dtype != torch.float64 or a.dtype != torch.float64 or b.dtype != torch.float64:
             raise ValueError('segment certification requires float64')
+        if a.shape != b.shape or a.ndim != 2 or a.shape[1] != 6:
+            raise ValueError('a and b must be [count,6] with identical shape')
         # 预算给成张量时 max_depth + 1 会在 uint8 这类窄类型上回绕成 0，循环整个跳过，把碰撞段当安全返回。
         if type(max_queries) is not int or type(max_depth) is not int:
             raise ValueError('certification budgets must be plain ints')
@@ -336,6 +338,8 @@ class Kin:
             raise ValueError('segment certification requires float64')
         if type(n) is not int or n < 1 or eps <= 0 or (cap is not None and cap <= eps):
             raise ValueError('invalid subdivision count or tolerance')
+        if a.shape != b.shape or a.ndim != 2 or a.shape[1] != 6:
+            raise ValueError('a and b must be [count,6] with identical shape')
         count = a.shape[0]
         inside = (((a >= self.lo) & (a <= self.hi) & (b >= self.lo) & (b <= self.hi)) | self.cont).all(-1)
         bound = (b - a).abs() @ self.cert_lever.T / n                       # [count,15] 子段界
diff --git a/tests/test_eval_cl.py b/tests/test_eval_cl.py
index 46b685a..0469170 100644
--- a/tests/test_eval_cl.py
+++ b/tests/test_eval_cl.py
@@ -437,6 +437,15 @@ class SegmentCertificateTest(unittest.TestCase):
                 self.assertTrue(result['unresolved'].item())
                 self.assertEqual(result['queries'].item(), 2)
 
+    def test_shape_mismatch_between_endpoints_is_rejected(self):
+        a = torch.zeros(1, 6, dtype=torch.float64)
+        b = torch.zeros(2, 6, dtype=torch.float64)
+        kin = self.ThinObstacle()
+        for method in (kin.certify_segments, kin.certify_uniform):
+            with self.subTest(method=method.__name__):
+                with self.assertRaisesRegex(ValueError, 'identical shape'):
+                    method(a, b)
+
     def test_budget_exhaustion_scales_with_gap_and_bound(self):
         class ConstantGap(self.ThinObstacle):
             def gaps(self, q, cap):