auto-sync 2026-07-04T03:59:09Z workspace (part 4)

workspace/runs/ctt_val_proxy_comparison/command.txt

@@ -1 +1 @@
-python scripts/build_ctt_proxy_comparison.py --out-dir runs/ctt_val_proxy_comparison
+python scripts/build_ctt_proxy_comparison.py --out-dir runs/ctt_val_proxy_comparison --no-markdown-report

workspace/runs/ctt_val_proxy_comparison/metrics.json

@@ -6,12 +6,15 @@
     {
       "beats_local_atlas_support": false,
       "candidate_diversity": 0.821928094090022,
+      "collapse_rate": 0.06605580409928237,
       "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
+      "mean_negative_distance": 0.6789322236382422,
       "mean_positive_distance": 0.7203264851945577,
       "method": "local_atlas",
       "negative_near_0p20": 0.03683574879227053,
       "negative_near_0p40": 0.20513731437644483,
       "num_rows": 69,
+      "pos_closer_than_neg": 0.5998446814623285,
       "pptc_0p20": 0.4057971014492754,
       "pptc_0p40": 0.6811594202898551,
       "proxy_gate_pass": false,
@@ -24,12 +27,15 @@
     {
       "beats_local_atlas_support": false,
       "candidate_diversity": 0.9751103010950768,
+      "collapse_rate": 0.06256038647342994,
       "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
+      "mean_negative_distance": 0.9219133529525894,
       "mean_positive_distance": 0.9615657234367121,
       "method": "task_memory",
       "negative_near_0p20": 0.017451690821256038,
       "negative_near_0p40": 0.1108695652173913,
       "num_rows": 69,
+      "pos_closer_than_neg": 0.5852941176470589,
       "pptc_0p20": 0.3188405797101449,
       "pptc_0p40": 0.5362318840579711,
       "proxy_gate_pass": false,
@@ -42,12 +48,15 @@
     {
       "beats_local_atlas_support": true,
       "candidate_diversity": 0.27027754304306056,
+      "collapse_rate": 0.06806780176345395,
       "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
+      "mean_negative_distance": 0.5291780433921461,
       "mean_positive_distance": 0.45091308311397055,
       "method": "ctt_residual_full",
       "negative_near_0p20": 0.029584063279715453,
       "negative_near_0p40": 0.289941117658509,
       "num_rows": 207,
+      "pos_closer_than_neg": 0.7351920682803036,
       "pptc_0p20": 0.19806763285024154,
       "pptc_0p40": 0.608695652173913,
       "proxy_gate_pass": true,
@@ -60,12 +69,15 @@
     {
       "beats_local_atlas_support": true,
       "candidate_diversity": 0.24708366700914647,
+      "collapse_rate": 0.06806780176345395,
       "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
+      "mean_negative_distance": 0.5274893750999132,
       "mean_positive_distance": 0.44291597778605085,
       "method": "ctt_residual_base_context",
       "negative_near_0p20": 0.018245341614906832,
       "negative_near_0p40": 0.28596910336040765,
       "num_rows": 69,
+      "pos_closer_than_neg": 0.7576806363571069,
       "pptc_0p20": 0.17391304347826086,
       "pptc_0p40": 0.6231884057971014,
       "proxy_gate_pass": true,
@@ -78,12 +90,15 @@
     {
       "beats_local_atlas_support": true,
       "candidate_diversity": 0.23968052668773945,
+      "collapse_rate": 0.06806780176345395,
       "data_hash": "1d15143588697e89f7fc2f6375b2745e4d479c726c9a947ff7eb7b2705280e1a",
+      "mean_negative_distance": 0.5130644364599437,
       "mean_positive_distance": 0.43468703005975967,
       "method": "ctt_residual_base_context_obs",
       "negative_near_0p20": 0.034336456619065314,
       "negative_near_0p40": 0.30197859708729274,
       "num_rows": 207,
+      "pos_closer_than_neg": 0.7717469893940482,
       "pptc_0p20": 0.24637681159420288,
       "pptc_0p40": 0.6425120772946861,
       "proxy_gate_pass": true,
@@ -96,12 +111,15 @@
     {
       "beats_local_atlas_support": true,
       "candidate_diversity": 0.24174490402345686,
+      "collapse_rate": 0.06830470254383299,
       "data_hash": "67f7b4a692f7c7e71da24378dc71e3399c5d35afc14c494869eba6087b765c42",
+      "mean_negative_distance": 0.5216229740528893,
       "mean_positive_distance": 0.43402743826233176,
       "method": "ctt_residual_base_context_obj",
       "negative_near_0p20": 0.038024455415759766,
       "negative_near_0p40": 0.30944174585478934,
       "num_rows": 207,
+      "pos_closer_than_neg": 0.7307419976537622,
       "pptc_0p20": 0.22705314009661837,
       "pptc_0p40": 0.6425120772946861,
       "proxy_gate_pass": true,
@@ -114,12 +132,15 @@
     {
       "beats_local_atlas_support": true,
       "candidate_diversity": 0.2447676812446944,
+      "collapse_rate": 0.06830470254383299,
       "data_hash": "67f7b4a692f7c7e71da24378dc71e3399c5d35afc14c494869eba6087b765c42",
+      "mean_negative_distance": 0.5333629090899681,
       "mean_positive_distance": 0.4429413077106721,
       "method": "ctt_residual_base_context_obs_obj",
       "negative_near_0p20": 0.02853647962343615,
       "negative_near_0p40": 0.2934249526640831,
       "num_rows": 207,
+      "pos_closer_than_neg": 0.7277074505015682,
       "pptc_0p20": 0.20772946859903382,
       "pptc_0p40": 0.642512077294686,
       "proxy_gate_pass": true,
@@ -132,12 +153,15 @@
     {
       "beats_local_atlas_support": true,
       "candidate_diversity": 0.1163743674776266,
+      "collapse_rate": 0.06806780176345395,
       "data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
+      "mean_negative_distance": 0.5050016174085634,
       "mean_positive_distance": 0.4336709210379143,
       "method": "ctt_gated_residual_full",
       "negative_near_0p20": 0.05269106899541682,
       "negative_near_0p40": 0.33921316203924895,
       "num_rows": 207,
+      "pos_closer_than_neg": 0.7247617850559026,
       "pptc_0p20": 0.2318840579710145,
       "pptc_0p40": 0.6135265700483091,
       "proxy_gate_pass": false,
@@ -149,5 +173,6 @@
     }
   ],
   "safety_slack": 0.01,
+  "schema_version": 2,
   "split_hash": "9d1dc1c10868f76fa080895df0f2730c173f12b6db14932599d55c2b60859381"
 }

workspace/runs/ctt_val_proxy_comparison/metrics_by_task.json

@@ -1 +1,498 @@
-{}
+{
+  "ctt_gated_residual_full": {
+    "LiftPegUpright-v1": {
+      "candidate_diversity": 0.05970962647106902,
+      "collapse_rate": 0.07311091686091686,
+      "mean_negative_distance": 0.3839348728146849,
+      "mean_positive_distance": 0.25572830294882354,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.53125,
+      "pos_closer_than_neg": 0.9444444444444444,
+      "pptc_0p20": 0.5,
+      "pptc_0p40": 0.8888888888888888,
+      "proxy_support_distance": 0.23664649313175856
+    },
+    "PickCube-v1": {
+      "candidate_diversity": 0.2021307744588189,
+      "collapse_rate": 0.10000000000000002,
+      "mean_negative_distance": 0.5829484689774836,
+      "mean_positive_distance": 0.576577907999753,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.3333333333333333,
+      "pos_closer_than_neg": 0.7666666666666666,
+      "pptc_0p20": 0.0,
+      "pptc_0p40": 0.3333333333333333,
+      "proxy_support_distance": 0.4443502650555026
+    },
+    "PullCube-v1": {
+      "candidate_diversity": 0.14847138063951565,
+      "collapse_rate": 0.06404465632406808,
+      "mean_negative_distance": 0.7041678688460474,
+      "mean_positive_distance": 0.592154236746947,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.05861928104575163,
+      "pos_closer_than_neg": 0.6517195767195768,
+      "pptc_0p20": 0.029411764705882353,
+      "pptc_0p40": 0.3137254901960784,
+      "proxy_support_distance": 0.5373627305388711
+    },
+    "PushCube-v1": {
+      "candidate_diversity": 0.05417944485503618,
+      "collapse_rate": 0.06478937728937728,
+      "mean_negative_distance": 0.3979735192222798,
+      "mean_positive_distance": 0.2811870120807798,
+      "negative_near_0p20": 0.08974358974358974,
+      "negative_near_0p40": 0.44871794871794873,
+      "pos_closer_than_neg": 0.625,
+      "pptc_0p20": 0.3076923076923077,
+      "pptc_0p40": 0.9230769230769231,
+      "proxy_support_distance": 0.26109739578233915
+    },
+    "StackCube-v1": {
+      "candidate_diversity": 0.11447220652079974,
+      "collapse_rate": 0.07588217338217343,
+      "mean_negative_distance": 0.3264807588283944,
+      "mean_positive_distance": 0.268244707965138,
+      "negative_near_0p20": 0.16460113960113962,
+      "negative_near_0p40": 0.8038990638990638,
+      "pos_closer_than_neg": 0.7895604395604395,
+      "pptc_0p20": 0.5333333333333333,
+      "pptc_0p40": 0.9333333333333333,
+      "proxy_support_distance": 0.2277182791599637
+    }
+  },
+  "ctt_residual_base_context": {
+    "LiftPegUpright-v1": {
+      "candidate_diversity": 0.14362703522993195,
+      "collapse_rate": 0.07311091686091686,
+      "mean_negative_distance": 0.41497127070807355,
+      "mean_positive_distance": 0.2305104051790113,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.5,
+      "pos_closer_than_neg": 1.0,
+      "pptc_0p20": 0.6666666666666666,
+      "pptc_0p40": 1.0,
+      "proxy_support_distance": 0.2014216883582193
+    },
+    "PickCube-v1": {
+      "candidate_diversity": 0.25660266524569825,
+      "collapse_rate": 0.1,
+      "mean_negative_distance": 0.899796341697192,
+      "mean_positive_distance": 0.9817664760096816,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.0,
+      "pos_closer_than_neg": 0.2,
+      "pptc_0p20": 0.0,
+      "pptc_0p40": 0.0,
+      "proxy_support_distance": 0.7855270291361975
+    },
+    "PullCube-v1": {
+      "candidate_diversity": 0.3216694071900426,
+      "collapse_rate": 0.06404465632406808,
+      "mean_negative_distance": 0.6926077604425354,
+      "mean_positive_distance": 0.5696106164927525,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.02022058823529412,
+      "pos_closer_than_neg": 0.6978174603174603,
+      "pptc_0p20": 0.029411764705882353,
+      "pptc_0p40": 0.38235294117647056,
+      "proxy_support_distance": 0.4743135938351562
+    },
+    "PushCube-v1": {
+      "candidate_diversity": 0.18464687502804178,
+      "collapse_rate": 0.06478937728937728,
+      "mean_negative_distance": 0.4607380704920465,
+      "mean_positive_distance": 0.31700052432677295,
+      "negative_near_0p20": 0.052884615384615384,
+      "negative_near_0p40": 0.3701923076923077,
+      "pos_closer_than_neg": 0.71875,
+      "pptc_0p20": 0.23076923076923078,
+      "pptc_0p40": 0.8461538461538461,
+      "proxy_support_distance": 0.267084673449308
+    },
+    "StackCube-v1": {
+      "candidate_diversity": 0.17288259514532145,
+      "collapse_rate": 0.07588217338217337,
+      "mean_negative_distance": 0.35211110872802126,
+      "mean_positive_distance": 0.3139070522101423,
+      "negative_near_0p20": 0.03809523809523809,
+      "negative_near_0p40": 0.7487912087912088,
+      "pos_closer_than_neg": 0.8025030525030525,
+      "pptc_0p20": 0.26666666666666666,
+      "pptc_0p40": 0.8666666666666667,
+      "proxy_support_distance": 0.2492223974983373
+    }
+  },
+  "ctt_residual_base_context_obj": {
+    "LiftPegUpright-v1": {
+      "candidate_diversity": 0.1493872636722864,
+      "collapse_rate": 0.07311091686091686,
+      "mean_negative_distance": 0.4120512868826152,
+      "mean_positive_distance": 0.26240904635443296,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.4647435897435897,
+      "pos_closer_than_neg": 0.9375,
+      "pptc_0p20": 0.4444444444444444,
+      "pptc_0p40": 0.8888888888888888,
+      "proxy_support_distance": 0.22436331113570077
+    },
+    "PickCube-v1": {
+      "candidate_diversity": 0.2693081365535955,
+      "collapse_rate": 0.10000000000000002,
+      "mean_negative_distance": 0.7942766276118407,
+      "mean_positive_distance": 0.8679002702981619,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.0,
+      "pos_closer_than_neg": 0.3333333333333333,
+      "pptc_0p20": 0.0,
+      "pptc_0p40": 0.0,
+      "proxy_support_distance": 0.7622612941754188
+    },
+    "PullCube-v1": {
+      "candidate_diversity": 0.31298127304146967,
+      "collapse_rate": 0.06404465632406808,
+      "mean_negative_distance": 0.7058399367948929,
+      "mean_positive_distance": 0.5663405352240012,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.03982843137254902,
+      "pos_closer_than_neg": 0.6993386243386243,
+      "pptc_0p20": 0.058823529411764705,
+      "pptc_0p40": 0.4019607843137255,
+      "proxy_support_distance": 0.47487613941275497
+    },
+    "PushCube-v1": {
+      "candidate_diversity": 0.18267389477352383,
+      "collapse_rate": 0.06478937728937728,
+      "mean_negative_distance": 0.45435797530595545,
+      "mean_positive_distance": 0.32555326496747355,
+      "negative_near_0p20": 0.052884615384615384,
+      "negative_near_0p40": 0.375,
+      "pos_closer_than_neg": 0.712673611111111,
+      "pptc_0p20": 0.15384615384615385,
+      "pptc_0p40": 0.8717948717948718,
+      "proxy_support_distance": 0.2722421007397641
+    },
+    "StackCube-v1": {
+      "candidate_diversity": 0.1665761829043618,
+      "collapse_rate": 0.07697191697191702,
+      "mean_negative_distance": 0.32524565684329504,
+      "mean_positive_distance": 0.26785120329886264,
+      "negative_near_0p20": 0.12907916157916158,
+      "negative_near_0p40": 0.8222568172568172,
+      "pos_closer_than_neg": 0.7281339031339031,
+      "pptc_0p20": 0.6,
+      "pptc_0p40": 0.9333333333333333,
+      "proxy_support_distance": 0.20483640069851286
+    }
+  },
+  "ctt_residual_base_context_obs": {
+    "LiftPegUpright-v1": {
+      "candidate_diversity": 0.148958579095827,
+      "collapse_rate": 0.07311091686091686,
+      "mean_negative_distance": 0.4115981169631194,
+      "mean_positive_distance": 0.2362653608389971,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.47061965811965817,
+      "pos_closer_than_neg": 0.9965277777777778,
+      "pptc_0p20": 0.6666666666666666,
+      "pptc_0p40": 1.0,
+      "proxy_support_distance": 0.1957123602918266
+    },
+    "PickCube-v1": {
+      "candidate_diversity": 0.28828954720262945,
+      "collapse_rate": 0.10000000000000002,
+      "mean_negative_distance": 0.7587970286519089,
+      "mean_positive_distance": 0.7901440255768358,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.0,
+      "pos_closer_than_neg": 0.5333333333333333,
+      "pptc_0p20": 0.0,
+      "pptc_0p40": 0.0,
+      "proxy_support_distance": 0.6021893630844198
+    },
+    "PullCube-v1": {
+      "candidate_diversity": 0.3133621097757295,
+      "collapse_rate": 0.06404465632406808,
+      "mean_negative_distance": 0.6841223846741898,
+      "mean_positive_distance": 0.5733947056530471,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.02839052287581699,
+      "pos_closer_than_neg": 0.7119047619047619,
+      "pptc_0p20": 0.0392156862745098,
+      "pptc_0p40": 0.37254901960784315,
+      "proxy_support_distance": 0.4821119993756999
+    },
+    "PushCube-v1": {
+      "candidate_diversity": 0.17226944365330818,
+      "collapse_rate": 0.06478937728937728,
+      "mean_negative_distance": 0.4515933904080564,
+      "mean_positive_distance": 0.31235985654167453,
+      "negative_near_0p20": 0.06891025641025642,
+      "negative_near_0p40": 0.38247863247863245,
+      "pos_closer_than_neg": 0.732638888888889,
+      "pptc_0p20": 0.3333333333333333,
+      "pptc_0p40": 0.9230769230769231,
+      "proxy_support_distance": 0.2586271514743408
+    },
+    "StackCube-v1": {
+      "candidate_diversity": 0.16414005465390855,
+      "collapse_rate": 0.07588217338217343,
+      "mean_negative_distance": 0.33057222184027074,
+      "mean_positive_distance": 0.2819713837511487,
+      "negative_near_0p20": 0.09822547822547821,
+      "negative_near_0p40": 0.8050203500203499,
+      "pos_closer_than_neg": 0.8023656898656899,
+      "pptc_0p20": 0.4888888888888889,
+      "pptc_0p40": 0.9111111111111111,
+      "proxy_support_distance": 0.22536104103963187
+    }
+  },
+  "ctt_residual_base_context_obs_obj": {
+    "LiftPegUpright-v1": {
+      "candidate_diversity": 0.14766830492610186,
+      "collapse_rate": 0.07311091686091686,
+      "mean_negative_distance": 0.41649342498231295,
+      "mean_positive_distance": 0.2643206130229132,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.4671474358974359,
+      "pos_closer_than_neg": 0.9861111111111112,
+      "pptc_0p20": 0.4444444444444444,
+      "pptc_0p40": 1.0,
+      "proxy_support_distance": 0.22818479501118874
+    },
+    "PickCube-v1": {
+      "candidate_diversity": 0.33757465230985995,
+      "collapse_rate": 0.10000000000000002,
+      "mean_negative_distance": 1.1075887642380198,
+      "mean_positive_distance": 1.1680868800298911,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.0,
+      "pos_closer_than_neg": 0.43333333333333335,
+      "pptc_0p20": 0.0,
+      "pptc_0p40": 0.0,
+      "proxy_support_distance": 0.9803242549573113
+    },
+    "PullCube-v1": {
+      "candidate_diversity": 0.3136336919427042,
+      "collapse_rate": 0.06404465632406808,
+      "mean_negative_distance": 0.7162658429465761,
+      "mean_positive_distance": 0.565417223038266,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.029411764705882353,
+      "pos_closer_than_neg": 0.683994708994709,
+      "pptc_0p20": 0.029411764705882353,
+      "pptc_0p40": 0.37254901960784315,
+      "proxy_support_distance": 0.47243182399485906
+    },
+    "PushCube-v1": {
+      "candidate_diversity": 0.181916451774259,
+      "collapse_rate": 0.06478937728937728,
+      "mean_negative_distance": 0.44441461966625945,
+      "mean_positive_distance": 0.32881952335913633,
+      "negative_near_0p20": 0.04487179487179487,
+      "negative_near_0p40": 0.3643162393162393,
+      "pos_closer_than_neg": 0.6822916666666666,
+      "pptc_0p20": 0.23076923076923078,
+      "pptc_0p40": 0.9230769230769231,
+      "proxy_support_distance": 0.27338380647453003
+    },
+    "StackCube-v1": {
+      "candidate_diversity": 0.17579507499334182,
+      "collapse_rate": 0.07697191697191702,
+      "mean_negative_distance": 0.3332039593498869,
+      "mean_positive_distance": 0.28734001912661167,
+      "negative_near_0p20": 0.09237891737891737,
+      "negative_near_0p40": 0.7804884004884005,
+      "pos_closer_than_neg": 0.7293905168905168,
+      "pptc_0p20": 0.5111111111111111,
+      "pptc_0p40": 0.9111111111111111,
+      "proxy_support_distance": 0.225556820839634
+    }
+  },
+  "ctt_residual_full": {
+    "LiftPegUpright-v1": {
+      "candidate_diversity": 0.16663572531428117,
+      "collapse_rate": 0.07311091686091686,
+      "mean_negative_distance": 0.41647353998567643,
+      "mean_positive_distance": 0.2771405265426538,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.468215811965812,
+      "pos_closer_than_neg": 0.90625,
+      "pptc_0p20": 0.4444444444444444,
+      "pptc_0p40": 0.8888888888888888,
+      "proxy_support_distance": 0.225699402965728
+    },
+    "PickCube-v1": {
+      "candidate_diversity": 0.2841252624025221,
+      "collapse_rate": 0.10000000000000002,
+      "mean_negative_distance": 0.7853855604157447,
+      "mean_positive_distance": 0.8107379646608238,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.06666666666666667,
+      "pos_closer_than_neg": 0.7333333333333334,
+      "pptc_0p20": 0.0,
+      "pptc_0p40": 0.3333333333333333,
+      "proxy_support_distance": 0.5973002537013845
+    },
+    "PullCube-v1": {
+      "candidate_diversity": 0.35669628571891027,
+      "collapse_rate": 0.06404465632406808,
+      "mean_negative_distance": 0.7141779455036535,
+      "mean_positive_distance": 0.5869452164166006,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.0196078431372549,
+      "pos_closer_than_neg": 0.6872354497354497,
+      "pptc_0p20": 0.029411764705882353,
+      "pptc_0p40": 0.3235294117647059,
+      "proxy_support_distance": 0.48532493799704673
+    },
+    "PushCube-v1": {
+      "candidate_diversity": 0.1897940504075825,
+      "collapse_rate": 0.06478937728937728,
+      "mean_negative_distance": 0.4571901910242416,
+      "mean_positive_distance": 0.33455829528179704,
+      "negative_near_0p20": 0.03685897435897436,
+      "negative_near_0p40": 0.36217948717948717,
+      "pos_closer_than_neg": 0.7239583333333334,
+      "pptc_0p20": 0.1282051282051282,
+      "pptc_0p40": 0.8717948717948718,
+      "proxy_support_distance": 0.27709473619662045
+    },
+    "StackCube-v1": {
+      "candidate_diversity": 0.18468096572943052,
+      "collapse_rate": 0.07588217338217343,
+      "mean_negative_distance": 0.33657300192660006,
+      "mean_positive_distance": 0.28893509427463027,
+      "negative_near_0p20": 0.10414224664224664,
+      "negative_near_0p40": 0.7836650386650388,
+      "pos_closer_than_neg": 0.7400234025234025,
+      "pptc_0p20": 0.5555555555555556,
+      "pptc_0p40": 0.9333333333333333,
+      "proxy_support_distance": 0.22290972126483205
+    }
+  },
+  "local_atlas": {
+    "LiftPegUpright-v1": {
+      "candidate_diversity": 0.5014669861510727,
+      "collapse_rate": 0.07219516594516594,
+      "mean_negative_distance": 0.4854021065936647,
+      "mean_positive_distance": 0.43646670032063084,
+      "negative_near_0p20": 0.027777777777777776,
+      "negative_near_0p40": 0.24219877344877347,
+      "pos_closer_than_neg": 0.7192460317460317,
+      "pptc_0p20": 0.5,
+      "pptc_0p40": 1.0,
+      "proxy_support_distance": 0.18189470652155995
+    },
+    "PickCube-v1": {
+      "candidate_diversity": 1.0944000325107157,
+      "collapse_rate": 0.1,
+      "mean_negative_distance": 0.6373729677105011,
+      "mean_positive_distance": 0.7798613153680638,
+      "negative_near_0p20": 0.3,
+      "negative_near_0p40": 0.4,
+      "pos_closer_than_neg": 0.6,
+      "pptc_0p20": 1.0,
+      "pptc_0p40": 1.0,
+      "proxy_support_distance": 0.07461591163524217
+    },
+    "PullCube-v1": {
+      "candidate_diversity": 1.0209711156114212,
+      "collapse_rate": 0.06316930618401206,
+      "mean_negative_distance": 0.9870468132552196,
+      "mean_positive_distance": 0.9460291977432415,
+      "negative_near_0p20": 0.003676470588235294,
+      "negative_near_0p40": 0.039460784313725496,
+      "pos_closer_than_neg": 0.6720238095238096,
+      "pptc_0p20": 0.058823529411764705,
+      "pptc_0p40": 0.38235294117647056,
+      "proxy_support_distance": 0.5490009012121931
+    },
+    "PushCube-v1": {
+      "candidate_diversity": 0.692147376654553,
+      "collapse_rate": 0.0641025641025641,
+      "mean_negative_distance": 0.5298147189077234,
+      "mean_positive_distance": 0.5107316898096026,
+      "negative_near_0p20": 0.028846153846153848,
+      "negative_near_0p40": 0.23237179487179485,
+      "pos_closer_than_neg": 0.5859375,
+      "pptc_0p20": 0.7692307692307693,
+      "pptc_0p40": 0.9230769230769231,
+      "proxy_support_distance": 0.21488945685319283
+    },
+    "StackCube-v1": {
+      "candidate_diversity": 0.593260181033124,
+      "collapse_rate": 0.06957264957264958,
+      "mean_negative_distance": 0.40728379791043173,
+      "mean_positive_distance": 0.4999574180225059,
+      "negative_near_0p20": 0.105,
+      "negative_near_0p40": 0.5292521367521368,
+      "pos_closer_than_neg": 0.4584401709401709,
+      "pptc_0p20": 0.8,
+      "pptc_0p40": 1.0,
+      "proxy_support_distance": 0.15391202596431447
+    }
+  },
+  "task_memory": {
+    "LiftPegUpright-v1": {
+      "candidate_diversity": 0.6210576200589012,
+      "collapse_rate": 0.0625,
+      "mean_negative_distance": 0.583368643118901,
+      "mean_positive_distance": 0.5613930072546827,
+      "negative_near_0p20": 0.010416666666666666,
+      "negative_near_0p40": 0.13541666666666666,
+      "pos_closer_than_neg": 0.6875,
+      "pptc_0p20": 0.5,
+      "pptc_0p40": 1.0,
+      "proxy_support_distance": 0.20479276324345116
+    },
+    "PickCube-v1": {
+      "candidate_diversity": 1.2387041849339937,
+      "collapse_rate": 0.06666666666666667,
+      "mean_negative_distance": 0.8095929802695191,
+      "mean_positive_distance": 0.9291008275018162,
+      "negative_near_0p20": 0.26666666666666666,
+      "negative_near_0p40": 0.4,
+      "pos_closer_than_neg": 0.6,
+      "pptc_0p20": 1.0,
+      "pptc_0p40": 1.0,
+      "proxy_support_distance": 0.07461591163524217
+    },
+    "PullCube-v1": {
+      "candidate_diversity": 1.3442506282390456,
+      "collapse_rate": 0.0625,
+      "mean_negative_distance": 1.383142918984482,
+      "mean_positive_distance": 1.3657646662805298,
+      "negative_near_0p20": 0.0,
+      "negative_near_0p40": 0.0,
+      "pos_closer_than_neg": 0.4732142857142857,
+      "pptc_0p20": 0.029411764705882353,
+      "pptc_0p40": 0.08823529411764706,
+      "proxy_support_distance": 0.8291223046675316
+    },
+    "PushCube-v1": {
+      "candidate_diversity": 0.6719745815075286,
+      "collapse_rate": 0.0625,
+      "mean_negative_distance": 0.7351552918288785,
+      "mean_positive_distance": 0.6373056713998045,
+      "negative_near_0p20": 0.014423076923076924,
+      "negative_near_0p40": 0.07211538461538461,
+      "pos_closer_than_neg": 0.609375,
+      "pptc_0p20": 0.38461538461538464,
+      "pptc_0p40": 0.9230769230769231,
+      "proxy_support_distance": 0.26447355510409054
+    },
+    "StackCube-v1": {
+      "candidate_diversity": 0.5251579967031662,
+      "collapse_rate": 0.0625,
+      "mean_negative_distance": 0.5187021685529322,
+      "mean_positive_distance": 0.48864024429118186,
+      "negative_near_0p20": 0.04583333333333333,
+      "negative_near_0p40": 0.36666666666666664,
+      "pos_closer_than_neg": 0.6875,
+      "pptc_0p20": 0.8,
+      "pptc_0p40": 1.0,
+      "proxy_support_distance": 0.17466840356416174
+    }
+  }
+}

workspace/runs/ctt_val_proxy_comparison/table.tex

@@ -1,15 +1,15 @@
 % Auto-generated by scripts/build_ctt_proxy_comparison.py
-\begin{tabular}{lrrrrrrrc}
+\begin{tabular}{lrrrrrrrrrrc}
 \toprule
-Method & Seeds & Rows & PPTC@0.20 & PPTC@0.40 & Neg@0.20 & MeanPosDist & Diversity & Gate \\
+Method & Seeds & Rows & PPTC@0.20 & PPTC@0.40 & Neg@0.20 & Neg@0.40 & Pos<Neg & MeanPos & MeanNeg & Collapse & Gate \\
 \midrule
-local\_atlas & 1 & 69 & 0.4058 & 0.6812 & 0.0368 & 0.7203 & 0.8219 & baseline \\
-task\_memory & 1 & 69 & 0.3188 & 0.5362 & 0.0175 & 0.9616 & 0.9751 & baseline \\
-ctt\_residual\_full & 3 & 207 & 0.1981 & 0.6087 & 0.0296 & 0.4509 & 0.2703 & pass \\
-ctt\_residual\_base\_context & 1 & 69 & 0.1739 & 0.6232 & 0.0182 & 0.4429 & 0.2471 & pass \\
-ctt\_residual\_base\_context\_obs & 3 & 207 & 0.2464 & 0.6425 & 0.0343 & 0.4347 & 0.2397 & pass \\
-ctt\_residual\_base\_context\_obj & 3 & 207 & 0.2271 & 0.6425 & 0.0380 & 0.4340 & 0.2417 & pass \\
-ctt\_residual\_base\_context\_obs\_obj & 3 & 207 & 0.2077 & 0.6425 & 0.0285 & 0.4429 & 0.2448 & pass \\
-ctt\_gated\_residual\_full & 3 & 207 & 0.2319 & 0.6135 & 0.0527 & 0.4337 & 0.1164 & fail \\
+local\_atlas & 1 & 69 & 0.4058 & 0.6812 & 0.0368 & 0.2051 & 0.5998 & 0.7203 & 0.6789 & 0.0661 & baseline \\
+task\_memory & 1 & 69 & 0.3188 & 0.5362 & 0.0175 & 0.1109 & 0.5853 & 0.9616 & 0.9219 & 0.0626 & baseline \\
+ctt\_residual\_full & 3 & 207 & 0.1981 & 0.6087 & 0.0296 & 0.2899 & 0.7352 & 0.4509 & 0.5292 & 0.0681 & pass \\
+ctt\_residual\_base\_context & 1 & 69 & 0.1739 & 0.6232 & 0.0182 & 0.2860 & 0.7577 & 0.4429 & 0.5275 & 0.0681 & pass \\
+ctt\_residual\_base\_context\_obs & 3 & 207 & 0.2464 & 0.6425 & 0.0343 & 0.3020 & 0.7717 & 0.4347 & 0.5131 & 0.0681 & pass \\
+ctt\_residual\_base\_context\_obj & 3 & 207 & 0.2271 & 0.6425 & 0.0380 & 0.3094 & 0.7307 & 0.4340 & 0.5216 & 0.0683 & pass \\
+ctt\_residual\_base\_context\_obs\_obj & 3 & 207 & 0.2077 & 0.6425 & 0.0285 & 0.2934 & 0.7277 & 0.4429 & 0.5334 & 0.0683 & pass \\
+ctt\_gated\_residual\_full & 3 & 207 & 0.2319 & 0.6135 & 0.0527 & 0.3392 & 0.7248 & 0.4337 & 0.5050 & 0.0681 & fail \\
 \bottomrule
 \end{tabular}

workspace/runs/paper_ctt_audit/command.txt

@@ -1 +1 @@
-python scripts/audit_ctt_paper_artifacts.py --repo-root . --paper latex/main.tex --out-dir runs/paper_ctt_audit
+python scripts/audit_ctt_paper_artifacts.py 

workspace/runs/summary_ctt.csv

@@ -206,6 +206,10 @@ CTT residual test dominance tau0,runs/ctt_base_context_obs_dominance_val_to_test
 "CTT residual test learned dominance (context, utility_margin)",runs/ctt_base_context_obs_learned_dominance_context_tanh_val_to_test,yes,measured dominance K=8 coverage=0.3403 fallback=0.6597 target=utility_margin features=context,"fail:selected<47.45,oracle<50,support_gap>7,selector_gap>3,OutcomePTR<=V0oracle",0.3333,0.3125,0.3819,0.7292,0.6913,0.3366,0.3681,0.1806,0.4167,n/a,n/a,n/a,n/a,n/a,3,see run
 "CTT residual test learned dominance (context, utility_margin)",runs/ctt_base_context_obs_learned_dominance_context_train_to_test,yes,measured dominance K=8 coverage=0.3403 fallback=0.6597 target=utility_margin features=context,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2917,0.2986,0.5139,0.7292,0.4885,0.4836,0.2639,0.2639,0.5347,n/a,n/a,n/a,n/a,n/a,3,see run
 "CTT residual test learned dominance (context, utility_margin)",runs/ctt_base_context_obs_learned_dominance_context_val_to_test,yes,measured dominance K=8 coverage=0.5069 fallback=0.4931 target=utility_margin features=context,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2917,0.3264,0.5139,0.7292,0.4885,0.4355,0.2639,0.2431,0.5347,n/a,n/a,n/a,n/a,n/a,3,see run
+"CTT residual test learned dominance (score_chart_compat, success_weighted_margin)",runs/ctt_base_context_obs_learned_dominance_score_chartcompat_successbonus2_task_envclip_k16_train_to_test,yes,measured dominance K=16 coverage=0.5486 fallback=0.4514 target=success_weighted_margin features=score_chart_compat,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2917,0.3194,0.5694,0.7292,0.3761,0.4580,0.2014,0.2500,0.5486,n/a,n/a,n/a,n/a,0.0302,3,see run
+"CTT residual test learned dominance (score_chart_compat, utility_margin)",runs/ctt_base_context_obs_learned_dominance_score_chartcompat_utility_task_envclip_k16_train_to_test,yes,measured dominance K=16 coverage=0.6389 fallback=0.3611 target=utility_margin features=score_chart_compat,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2917,0.3264,0.5694,0.7292,0.3761,0.4654,0.2014,0.2500,0.5486,n/a,n/a,n/a,n/a,0.0221,3,see run
+"CTT residual test learned dominance (score_context_chart_compat, success_weighted_margin)",runs/ctt_base_context_obs_learned_dominance_score_context_chartcompat_successbonus2_task_envclip_k16_train_to_test,yes,measured dominance K=16 coverage=0.6111 fallback=0.3889 target=success_weighted_margin features=score_context_chart_compat,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2917,0.3403,0.5694,0.7292,0.3761,0.4248,0.2014,0.2292,0.5486,n/a,n/a,n/a,n/a,0.0365,3,see run
+"CTT residual test learned dominance (score_context_chart_compat, utility_margin)",runs/ctt_base_context_obs_learned_dominance_score_context_chartcompat_utility_task_envclip_k16_train_to_test,yes,measured dominance K=16 coverage=0.6458 fallback=0.3542 target=utility_margin features=score_context_chart_compat,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2917,0.3264,0.5694,0.7292,0.3761,0.4655,0.2014,0.2500,0.5486,n/a,n/a,n/a,n/a,0.0188,3,see run
 "CTT residual test learned dominance (source_evidence, utility_margin)",runs/ctt_base_context_obs_learned_dominance_source_envclip_k16_train_to_test,yes,measured dominance K=16 coverage=0.5903 fallback=0.4097 target=utility_margin features=source_evidence,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2917,0.3056,0.5694,0.7292,0.3761,0.5187,0.2014,0.2847,0.5486,n/a,n/a,n/a,n/a,n/a,3,see run
 "CTT residual test learned dominance (source_evidence, utility_margin)",runs/ctt_base_context_obs_learned_dominance_source_envclip_train_to_test,yes,measured dominance K=8 coverage=0.2361 fallback=0.7639 target=utility_margin features=source_evidence,"fail:selected<47.45,oracle<50,support_gap>7,selector_gap>3",0.2917,0.3333,0.4792,0.7292,0.5381,0.3674,0.2917,0.2014,0.5069,n/a,n/a,n/a,n/a,n/a,3,see run
 "CTT residual test learned dominance (source_evidence, success)",runs/ctt_base_context_obs_learned_dominance_source_success_envclip_k16_train_to_test,yes,measured dominance K=16 coverage=0.5139 fallback=0.4861 target=success features=source_evidence,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2917,0.3056,0.5694,0.7292,0.3761,0.4804,0.2014,0.2708,0.5486,n/a,n/a,n/a,n/a,n/a,3,see run

workspace/scripts/build_ctt_proxy_comparison.py

@@ -43,6 +43,11 @@ def main(argv: list[str] | None = None) -> int:
     parser = argparse.ArgumentParser(description="Build CTT validation proxy comparison table.")
     parser.add_argument("--out-dir", type=Path, default=Path("runs/ctt_val_proxy_comparison"))
     parser.add_argument("--safety-slack", type=float, default=0.01)
+    parser.add_argument(
+        "--no-markdown-report",
+        action="store_true",
+        help="Do not write report.md; the persistent prose summary lives in README.md.",
+    )
     args = parser.parse_args(argv)
 
     rows = [
@@ -74,6 +79,7 @@ def main(argv: list[str] | None = None) -> int:
     out_dir.mkdir(parents=True, exist_ok=True)
     payload = {
         "report_type": "ctt_val_proxy_comparison",
+        "schema_version": 2,
         "baseline": "local_atlas",
         "safety_slack": args.safety_slack,
         "rows": rows,
@@ -81,10 +87,21 @@ def main(argv: list[str] | None = None) -> int:
         "split_hash": _first(rows, "split_hash"),
     }
     (out_dir / "metrics.json").write_text(json.dumps(payload, indent=2, sort_keys=True) + "\n")
-    (out_dir / "metrics_by_task.json").write_text("{}\n")
-    (out_dir / "metrics_by_seed.json").write_text("{}\n")
+    (out_dir / "metrics_by_task.json").write_text(
+        json.dumps(_grouped_metrics(DEFAULT_RUNS, group_key="task_id"), indent=2, sort_keys=True)
+        + "\n"
+    )
+    (out_dir / "metrics_by_seed.json").write_text(
+        json.dumps(_grouped_metrics(DEFAULT_RUNS, group_key="seed"), indent=2, sort_keys=True)
+        + "\n"
+    )
     (out_dir / "table.tex").write_text(_table(rows) + "\n")
-    (out_dir / "report.md").write_text(_report(rows, args.safety_slack) + "\n")
+    _write_report_artifact(
+        out_dir,
+        rows,
+        safety_slack=args.safety_slack,
+        no_markdown_report=args.no_markdown_report,
+    )
     (out_dir / "train.log").write_text("comparison artifact; no training\n")
     (out_dir / "eval.log").write_text(
         "\n".join(f"{row['method']}: {row['run_path']}" for row in rows) + "\n"
@@ -113,7 +130,10 @@ def _row(method: str, run_dirs: list[Path]) -> dict[str, Any]:
         "negative_near_0p20": _mean_across(summaries, "negative_near_at_16_thr_0p20"),
         "negative_near_0p40": _mean_across(summaries, "negative_near_at_16_thr_0p40"),
         "mean_positive_distance": _mean_across(summaries, "mean_positive_distance_at_16"),
+        "mean_negative_distance": _mean_across(summaries, "mean_negative_distance_at_16"),
+        "pos_closer_than_neg": _mean_across(summaries, "pos_closer_than_neg_at_16"),
         "candidate_diversity": _mean_across(summaries, "candidate_diversity_at_16"),
+        "collapse_rate": _mean_across(summaries, "collapse_rate_at_16"),
         "proxy_support_distance": _mean_across(summaries, "proxy_support_distance_at_16"),
         "data_hash": payloads[0].get("data_hash"),
         "split_hash": payloads[0].get("target_split_hash", payloads[0].get("split_hash")),
@@ -132,17 +152,19 @@ def _mean_across(summaries: list[dict[str, Any]], key: str) -> float:
 def _table(rows: list[dict[str, Any]]) -> str:
     lines = [
         "% Auto-generated by scripts/build_ctt_proxy_comparison.py",
-        "\\begin{tabular}{lrrrrrrrc}",
+        "\\begin{tabular}{lrrrrrrrrrrc}",
         "\\toprule",
-        "Method & Seeds & Rows & PPTC@0.20 & PPTC@0.40 & Neg@0.20 & MeanPosDist & Diversity & Gate \\\\",
+        "Method & Seeds & Rows & PPTC@0.20 & PPTC@0.40 & Neg@0.20 & Neg@0.40 & Pos<Neg & MeanPos & MeanNeg & Collapse & Gate \\\\",
         "\\midrule",
     ]
     for row in rows:
         lines.append(
             f"{_latex_escape(row['method'])} & {row['train_seeds']} & {row['num_rows']} & "
             f"{_fmt(row['pptc_0p20'])} & {_fmt(row['pptc_0p40'])} & "
-            f"{_fmt(row['negative_near_0p20'])} & {_fmt(row['mean_positive_distance'])} & "
-            f"{_fmt(row['candidate_diversity'])} & {_gate(row)} \\\\"
+            f"{_fmt(row['negative_near_0p20'])} & {_fmt(row['negative_near_0p40'])} & "
+            f"{_fmt(row['pos_closer_than_neg'])} & {_fmt(row['mean_positive_distance'])} & "
+            f"{_fmt(row['mean_negative_distance'])} & {_fmt(row['collapse_rate'])} & "
+            f"{_gate(row)} \\\\"
         )
     lines.extend(["\\bottomrule", "\\end{tabular}"])
     return "\n".join(lines)
@@ -154,15 +176,17 @@ def _report(rows: list[dict[str, Any]], safety_slack: float) -> str:
         "",
         f"Safety slack over local-atlas NegativeNear@0.20: `{safety_slack:.4f}`",
         "",
-        "| Method | Seeds | Rows | PPTC@0.20 | PPTC@0.40 | Neg@0.20 | MeanPosDist | Diversity | Gate | Run |",
-        "| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | --- | --- |",
+        "| Method | Seeds | Rows | PPTC@0.20 | PPTC@0.40 | Neg@0.20 | Neg@0.40 | Pos<Neg | MeanPos | MeanNeg | Diversity | Collapse | Gate | Run |",
+        "| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | --- | --- |",
     ]
     for row in rows:
         lines.append(
             f"| {row['method']} | {row['train_seeds']} | {row['num_rows']} | "
             f"{_fmt(row['pptc_0p20'])} | "
             f"{_fmt(row['pptc_0p40'])} | {_fmt(row['negative_near_0p20'])} | "
-            f"{_fmt(row['mean_positive_distance'])} | {_fmt(row['candidate_diversity'])} | "
+            f"{_fmt(row['negative_near_0p40'])} | {_fmt(row['pos_closer_than_neg'])} | "
+            f"{_fmt(row['mean_positive_distance'])} | {_fmt(row['mean_negative_distance'])} | "
+            f"{_fmt(row['candidate_diversity'])} | {_fmt(row['collapse_rate'])} | "
             f"{_gate(row)} | `{row['run_path']}` |"
         )
     lines.append("")
@@ -170,6 +194,66 @@ def _report(rows: list[dict[str, Any]], safety_slack: float) -> str:
     return "\n".join(lines)
 
 
+def _write_report_artifact(
+    out_dir: Path,
+    rows: list[dict[str, Any]],
+    *,
+    safety_slack: float,
+    no_markdown_report: bool,
+) -> None:
+    report_path = out_dir / "report.md"
+    if no_markdown_report:
+        if report_path.exists():
+            report_path.unlink()
+        return
+    report_path.write_text(_report(rows, safety_slack) + "\n")
+
+
+def _grouped_metrics(
+    run_specs: list[tuple[str, list[Path]]],
+    *,
+    group_key: str,
+) -> dict[str, dict[str, dict[str, float]]]:
+    output: dict[str, dict[str, dict[str, float]]] = {}
+    for method, run_dirs in run_specs:
+        if not run_dirs or not all((run_dir / "metrics.json").exists() for run_dir in run_dirs):
+            continue
+        rows: list[dict[str, Any]] = []
+        for run_dir in run_dirs:
+            payload = json.loads((run_dir / "metrics.json").read_text())
+            rows.extend(payload.get("rows", []))
+        if not rows:
+            continue
+        method_groups: dict[str, dict[str, list[float]]] = {}
+        for row in rows:
+            group = str(row.get(group_key, "unknown"))
+            metrics = method_groups.setdefault(group, {})
+            for source_key, target_key in _row_metric_key_map().items():
+                value = row.get(source_key)
+                if isinstance(value, (int, float)) and math.isfinite(float(value)):
+                    metrics.setdefault(target_key, []).append(float(value))
+        output[method] = {
+            group: {name: sum(values) / len(values) for name, values in sorted(metrics.items())}
+            for group, metrics in sorted(method_groups.items())
+        }
+    return output
+
+
+def _row_metric_key_map() -> dict[str, str]:
+    return {
+        "pptc_at_16_thr_0p20": "pptc_0p20",
+        "pptc_at_16_thr_0p40": "pptc_0p40",
+        "negative_near_at_16_thr_0p20": "negative_near_0p20",
+        "negative_near_at_16_thr_0p40": "negative_near_0p40",
+        "mean_positive_distance_at_16": "mean_positive_distance",
+        "mean_negative_distance_at_16": "mean_negative_distance",
+        "pos_closer_than_neg_at_16": "pos_closer_than_neg",
+        "candidate_diversity_at_16": "candidate_diversity",
+        "collapse_rate_at_16": "collapse_rate",
+        "proxy_support_distance_at_16": "proxy_support_distance",
+    }
+
+
 def _fmt(value: Any) -> str:
     if not isinstance(value, (int, float)) or not math.isfinite(float(value)):
         return "n/a"

workspace/scripts/eval_ctt_proxy.py

@@ -39,6 +39,11 @@ def main(argv: list[str] | None = None) -> int:
     parser.add_argument("--thresholds", default="0.20,0.40")
     parser.add_argument("--max-target-charts", type=int, default=64)
     parser.add_argument("--neighbors", type=int, default=8)
+    parser.add_argument(
+        "--no-markdown-report",
+        action="store_true",
+        help="Do not write report.md; the persistent prose summary lives in README.md.",
+    )
     args = parser.parse_args(argv)
 
     thresholds = [float(item) for item in args.thresholds.split(",") if item.strip()]
@@ -159,7 +164,7 @@ def main(argv: list[str] | None = None) -> int:
     (out_dir / "eval.log").write_text("\n".join(log_lines) + "\n")
     (out_dir / "train.log").write_text("see checkpoint run\n")
     (out_dir / "table.tex").write_text(_table(summary) + "\n")
-    (out_dir / "report.md").write_text(_report(summary, args.k) + "\n")
+    _write_report_artifact(out_dir, summary, k=args.k, no_markdown_report=args.no_markdown_report)
     print(json.dumps({"out_dir": str(out_dir), "num_rows": len(rows)}, indent=2))
     return 0
 
@@ -271,5 +276,20 @@ def _report(summary: dict[str, Any], k: int) -> str:
     return "\n".join(lines)
 
 
+def _write_report_artifact(
+    out_dir: Path,
+    summary: dict[str, Any],
+    *,
+    k: int,
+    no_markdown_report: bool,
+) -> None:
+    report_path = out_dir / "report.md"
+    if no_markdown_report:
+        if report_path.exists():
+            report_path.unlink()
+        return
+    report_path.write_text(_report(summary, k) + "\n")
+
+
 if __name__ == "__main__":
     raise SystemExit(main())

workspace/scripts/eval_learned_dominance_selector.py

@@ -82,17 +82,30 @@ CHART_COMPAT_NAMES = [
     "target_obj_norm",
     "source_obj_norm",
 ]
+SCORE_SHAPE_NAMES = [
+    "candidate_score_rank_fraction",
+    "candidate_score_softmax_prob",
+    "candidate_score_gap_to_best",
+    "candidate_score_gap_to_second_best",
+    "candidate_score_gap_to_prev_higher",
+    "candidate_score_gap_to_next_lower",
+    "candidate_score_percentile",
+    "candidate_score_top_margin",
+]
 FEATURE_SET_CHOICES = (
     "basic",
     "tangent",
     "context",
     "context_tangent",
+    "score_context",
     "source_evidence",
     "tangent_source_evidence",
     "context_source_evidence",
     "context_tangent_source_evidence",
     "chart_compat",
     "chart_tangent_compat",
+    "score_chart_compat",
+    "score_context_chart_compat",
     "chart_source_compat",
     "chart_tangent_source_compat",
 )
@@ -149,6 +162,15 @@ def main(argv: list[str] | None = None) -> int:
             "candidate success to prioritize the lexicographic success/progress utility."
         ),
     )
+    parser.add_argument(
+        "--success-bonus",
+        type=float,
+        default=1.0,
+        help=(
+            "Bonus multiplier for candidate_success when --target=success_weighted_margin. "
+            "Default preserves the original +1 success bonus."
+        ),
+    )
     parser.add_argument(
         "--threshold-scope",
         choices=("global", "task"),
@@ -175,6 +197,11 @@ def main(argv: list[str] | None = None) -> int:
         help="Relative weight for pairwise rows when --fit-objective=hybrid_pairwise.",
     )
     parser.add_argument("--bootstrap-samples", type=int, default=1000)
+    parser.add_argument(
+        "--no-markdown-report",
+        action="store_true",
+        help="Do not write report.md; useful when the workspace is kept README-only.",
+    )
     args = parser.parse_args(argv)
 
     if args.k <= 0:
@@ -184,6 +211,8 @@ def main(argv: list[str] | None = None) -> int:
         parser.error("--ridge-lambdas must contain non-negative values")
     if args.pairwise_weight <= 0.0:
         parser.error("--pairwise-weight must be positive")
+    if args.success_bonus < 0.0:
+        parser.error("--success-bonus must be non-negative")
 
     out_dir = args.out_dir
     out_dir.mkdir(parents=True, exist_ok=True)
@@ -236,6 +265,7 @@ def main(argv: list[str] | None = None) -> int:
         selector_target_charts=selector_calibration_charts,
         selector_source_charts=selector_source_charts,
         selector_chart_feature_mode=args.selector_chart_feature_mode,
+        success_bonus=args.success_bonus,
     )
     eval_dataset = _candidate_dataset(
         eval_rows,
@@ -248,6 +278,7 @@ def main(argv: list[str] | None = None) -> int:
         selector_target_charts=selector_eval_charts,
         selector_source_charts=selector_source_charts,
         selector_chart_feature_mode=args.selector_chart_feature_mode,
+        success_bonus=args.success_bonus,
     )
     best = _fit_select_ridge(
         calibration_dataset,
@@ -307,6 +338,7 @@ def main(argv: list[str] | None = None) -> int:
             args.selector_chart_feature_mode if _uses_chart_compat(args.feature_set) else None
         ),
         "target": args.target,
+        "success_bonus": args.success_bonus,
         "fit_objective": args.fit_objective,
         "pairwise_weight": args.pairwise_weight,
         "threshold_scope": args.threshold_scope,
@@ -360,7 +392,7 @@ def main(argv: list[str] | None = None) -> int:
         + "\n"
     )
     (out_dir / "table.tex").write_text(_table(metrics) + "\n")
-    (out_dir / "report.md").write_text(_report(metrics) + "\n")
+    _write_report_artifact(out_dir, metrics, no_markdown_report=args.no_markdown_report)
     (out_dir / "train.log").write_text(
         "trained ridge dominance calibrator on calibration measured rows only\n"
         f"fit_objective={args.fit_objective}\n"
@@ -399,6 +431,7 @@ def _candidate_dataset(
     selector_target_charts: dict[str, Any] | None = None,
     selector_source_charts: dict[str, Any] | None = None,
     selector_chart_feature_mode: str = "base_context_obs_obj",
+    success_bonus: float = 1.0,
 ) -> dict[str, Any]:
     source_evidence = source_evidence or {}
     selector_target_charts = selector_target_charts or {}
@@ -422,6 +455,7 @@ def _candidate_dataset(
             continue
         score_mean = sum(scores) / len(scores)
         score_std = math.sqrt(sum((score - score_mean) ** 2 for score in scores) / len(scores)) + 1.0e-6
+        score_shape = _score_shape_matrix(scores)
         for candidate_index, score in enumerate(scores):
             source_chart_id = str(source_chart_ids[candidate_index]) if candidate_index < len(source_chart_ids) else ""
             tangent = np.asarray(
@@ -454,6 +488,7 @@ def _candidate_dataset(
                     selector_source_charts.get(source_chart_id),
                     chart_feature_mode=selector_chart_feature_mode,
                 ),
+                score_shape=score_shape[candidate_index],
                 num_candidates=len(scores),
                 feature_set=feature_set,
             )
@@ -472,6 +507,7 @@ def _candidate_dataset(
                         target,
                         utility_margin=target_margin,
                         candidate_success=successes[candidate_index],
+                        success_bonus=success_bonus,
                     ),
                     "measured_utility_margin": target_margin,
                     "candidate_utility": utilities[candidate_index],
@@ -508,6 +544,8 @@ def _feature_names(feature_set: str) -> list[str]:
         *[f"abs_tangent_{index:02d}" for index in range(21)],
     ]
     names = list(BASIC_FEATURE_NAMES)
+    if _uses_score_shape(feature_set):
+        names.extend(SCORE_SHAPE_NAMES)
     if _uses_context(feature_set):
         names.extend(context_names)
     if _uses_tangent(feature_set):
@@ -535,6 +573,7 @@ def _candidate_feature(
     context: dict[str, Any] | None = None,
     source_evidence: np.ndarray | None = None,
     chart_compat: np.ndarray | None = None,
+    score_shape: np.ndarray | None = None,
 ) -> np.ndarray:
     tangent = np.asarray(tangent, dtype=float).reshape(-1)
     if tangent.size < 21:
@@ -561,6 +600,10 @@ def _candidate_feature(
     if feature_set == "basic":
         return basic
     parts = [basic]
+    if _uses_score_shape(feature_set):
+        if score_shape is None:
+            score_shape = np.zeros(len(SCORE_SHAPE_NAMES), dtype=float)
+        parts.append(np.asarray(score_shape, dtype=float).reshape(-1))
     if _uses_context(feature_set):
         parts.append(_context_feature(context or {}))
     if _uses_tangent(feature_set):
@@ -584,6 +627,16 @@ def _uses_context(feature_set: str) -> bool:
         "context_tangent",
         "context_source_evidence",
         "context_tangent_source_evidence",
+        "score_context",
+        "score_context_chart_compat",
+    }
+
+
+def _uses_score_shape(feature_set: str) -> bool:
+    return feature_set in {
+        "score_context",
+        "score_chart_compat",
+        "score_context_chart_compat",
     }
 
 
@@ -613,6 +666,8 @@ def _uses_chart_compat(feature_set: str) -> bool:
     return feature_set in {
         "chart_compat",
         "chart_tangent_compat",
+        "score_chart_compat",
+        "score_context_chart_compat",
         "chart_source_compat",
         "chart_tangent_source_compat",
     }
@@ -864,6 +919,44 @@ def _min_rms_distance(tangent: np.ndarray, candidates: np.ndarray) -> float:
     return float(np.sqrt(np.mean(diff * diff, axis=1)).min())
 
 
+def _score_shape_matrix(scores: list[float]) -> np.ndarray:
+    """Deployment-visible row-relative score features for each candidate."""
+    score_array = np.asarray(scores, dtype=float).reshape(-1)
+    if score_array.size == 0:
+        return np.zeros((0, len(SCORE_SHAPE_NAMES)), dtype=float)
+    order = sorted(range(score_array.size), key=lambda index: (-float(score_array[index]), index))
+    ranks = np.zeros(score_array.size, dtype=float)
+    for rank, index in enumerate(order):
+        ranks[index] = float(rank)
+    sorted_scores = score_array[order]
+    best = float(sorted_scores[0])
+    second = float(sorted_scores[1]) if sorted_scores.size > 1 else best
+    denom = max(1.0, float(score_array.size - 1))
+    shifted = score_array - float(np.max(score_array))
+    exp_scores = np.exp(np.clip(shifted, -60.0, 60.0))
+    softmax = exp_scores / max(float(exp_scores.sum()), 1.0e-12)
+
+    rows: list[list[float]] = []
+    for index, score in enumerate(score_array):
+        rank = int(ranks[index])
+        prev_higher = sorted_scores[rank - 1] if rank > 0 else score
+        next_lower = sorted_scores[rank + 1] if rank + 1 < sorted_scores.size else score
+        percentile = float(np.mean(score_array <= score))
+        rows.append(
+            [
+                float(rank) / denom,
+                float(softmax[index]),
+                float(score - best),
+                float(score - second),
+                float(score - prev_higher),
+                float(score - next_lower),
+                percentile,
+                float(best - second),
+            ]
+        )
+    return np.asarray(rows, dtype=float)
+
+
 def _context_feature(context: dict[str, Any]) -> np.ndarray:
     target_task = str(context.get("target_task_id", ""))
     source_task = str(context.get("source_task_id", ""))
@@ -891,13 +984,14 @@ def _target_value(
     *,
     utility_margin: float,
     candidate_success: float,
+    success_bonus: float = 1.0,
 ) -> float:
     if target == "utility_margin":
         return float(utility_margin)
     if target == "success":
         return float(candidate_success)
     if target == "success_weighted_margin":
-        return float(utility_margin) + float(candidate_success)
+        return float(utility_margin) + float(success_bonus) * float(candidate_success)
     raise ValueError(f"unknown target: {target}")
 
 
@@ -1446,6 +1540,20 @@ def _report(metrics: dict[str, Any]) -> str:
     return "\n".join(lines)
 
 
+def _write_report_artifact(
+    out_dir: Path,
+    metrics: dict[str, Any],
+    *,
+    no_markdown_report: bool = False,
+) -> None:
+    report_path = out_dir / "report.md"
+    if no_markdown_report:
+        if report_path.exists():
+            report_path.unlink()
+        return
+    report_path.write_text(_report(metrics) + "\n")
+
+
 def _write_provenance(out_dir: Path, args: argparse.Namespace) -> None:
     (out_dir / "config.yaml").write_text(
         "\n".join(f"{key}: {value}" for key, value in sorted(vars(args).items())) + "\n"

workspace/scripts/eval_nonlinear_dominance_selector.py

@@ -105,6 +105,11 @@ def main(argv: list[str] | None = None) -> int:
     parser.add_argument("--selection-frac", type=float, default=0.35)
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--bootstrap-samples", type=int, default=1000)
+    parser.add_argument(
+        "--no-markdown-report",
+        action="store_true",
+        help="Do not write report.md; useful when the workspace is kept README-only.",
+    )
     args = parser.parse_args(argv)
 
     if args.k <= 0:
@@ -309,7 +314,7 @@ def main(argv: list[str] | None = None) -> int:
         + "\n"
     )
     (out_dir / "table.tex").write_text(_table(metrics) + "\n")
-    (out_dir / "report.md").write_text(_report(metrics) + "\n")
+    _write_report_artifact(out_dir, metrics, no_markdown_report=args.no_markdown_report)
     (out_dir / "train.log").write_text(
         "trained nonlinear dominance selector on calibration-fit rows only\n"
         f"selected_model_type={best['model_type']}\n"
@@ -599,6 +604,20 @@ def _report(metrics: dict[str, Any]) -> str:
     return "\n".join(lines)
 
 
+def _write_report_artifact(
+    out_dir: Path,
+    metrics: dict[str, Any],
+    *,
+    no_markdown_report: bool = False,
+) -> None:
+    report_path = out_dir / "report.md"
+    if no_markdown_report:
+        if report_path.exists():
+            report_path.unlink()
+        return
+    report_path.write_text(_report(metrics) + "\n")
+
+
 def _write_provenance(out_dir: Path, args: argparse.Namespace) -> None:
     (out_dir / "config.yaml").write_text(
         "\n".join(f"{key}: {value}" for key, value in sorted(vars(args).items())) + "\n"