add score-chart utility selector diagnostic

runs/ctt_base_context_obs_learned_dominance_score_chartcompat_utility_task_envclip_k16_train_to_test/command.txt

@@ -0,0 +1 @@
+python scripts/eval_learned_dominance_selector.py --calibration-input runs/ctt_base_context_obs_train_cal_envclip_k16_rollout_comparison/combined_measured_candidates.json --calibration-target-index data/cil_charts_rgb_refs/train/index.json --eval-input runs/ctt_base_context_obs_test_envclip_k16_rollout_comparison/combined_measured_candidates.json --eval-target-index data/cil_charts_rgb_refs/test/index.json --source-index data/cil_charts_rgb_refs/train/index.json --checkpoint-template runs/ctt_residual_base_context_obs_seed{seed}/model.pt --out-dir runs/ctt_base_context_obs_learned_dominance_score_chartcompat_utility_task_envclip_k16_train_to_test --feature-set score_chart_compat --selector-chart-feature-mode base_context_obs --target utility_margin --threshold-scope task --k 16 --bootstrap-samples 300 --no-markdown-report

runs/ctt_base_context_obs_learned_dominance_score_chartcompat_utility_task_envclip_k16_train_to_test/config.yaml

@@ -0,0 +1,18 @@
+bootstrap_samples: 300
+calibration_input: runs/ctt_base_context_obs_train_cal_envclip_k16_rollout_comparison/combined_measured_candidates.json
+calibration_target_index: data/cil_charts_rgb_refs/train/index.json
+checkpoint_template: runs/ctt_residual_base_context_obs_seed{seed}/model.pt
+eval_input: runs/ctt_base_context_obs_test_envclip_k16_rollout_comparison/combined_measured_candidates.json
+eval_target_index: data/cil_charts_rgb_refs/test/index.json
+feature_set: score_chart_compat
+fit_objective: pointwise
+k: 16
+no_markdown_report: True
+out_dir: runs/ctt_base_context_obs_learned_dominance_score_chartcompat_utility_task_envclip_k16_train_to_test
+pairwise_weight: 1.0
+ridge_lambdas: 0,0.01,0.1,1,10,100
+selector_chart_feature_mode: base_context_obs
+source_index: data/cil_charts_rgb_refs/train/index.json
+success_bonus: 1.0
+target: utility_margin
+threshold_scope: task

runs/ctt_base_context_obs_learned_dominance_score_chartcompat_utility_task_envclip_k16_train_to_test/data_hash.txt

@@ -0,0 +1,7 @@
+{
+  "calibration_input": "d79e275b896eb98caa128470d4614b76960cab12407faa157c583224be7abe80",
+  "calibration_target_index": "67973c50072aa87a3b97533b6292b08a8b4d88a95571dc22e91b6e8fd19e0d85",
+  "eval_input": "541d5a928da5f7ee04ca3dea570e04208830dffde431edfe5b5e38b692634045",
+  "eval_target_index": "dee1b5225485df6624747dd4c5074bcf514685b54017d05e698d2051638fbf36",
+  "source_index": "67973c50072aa87a3b97533b6292b08a8b4d88a95571dc22e91b6e8fd19e0d85"
+}

runs/ctt_base_context_obs_learned_dominance_score_chartcompat_utility_task_envclip_k16_train_to_test/eval.log

@@ -0,0 +1,3 @@
+evaluated learned fallback rule on held-out measured rollout rows
+eval_input=runs/ctt_base_context_obs_test_envclip_k16_rollout_comparison/combined_measured_candidates.json
+num_eval_rows=144

runs/ctt_base_context_obs_learned_dominance_score_chartcompat_utility_task_envclip_k16_train_to_test/git_hash.txt

@@ -0,0 +1 @@
+14e81292e2004925f81e0bb4810f6f144bfd87b7

runs/ctt_base_context_obs_learned_dominance_score_chartcompat_utility_task_envclip_k16_train_to_test/metrics_by_task.json

@@ -0,0 +1,137 @@
+{
+  "LiftPegUpright-v1": {
+    "base_success": 0.2,
+    "base_utility": 0.9498705387115478,
+    "coverage": 0.6666666666666666,
+    "execute_generated": 0.6666666666666666,
+    "fallback_rate": 0.3333333333333333,
+    "hidden_chart_oracle_success": 1.0,
+    "hidden_chart_oracle_utility": 2.0,
+    "outcome_ptr": 0.5333333333333333,
+    "pairwise_causal_calibration_accuracy": 0.5494360697557629,
+    "pairwise_causal_calibration_confidence": 0.5321483671612611,
+    "pairwise_causal_calibration_ece": 0.08779692862966537,
+    "pairwise_causal_calibration_pairs": 94.8,
+    "predicted_margin": 0.09934494198504117,
+    "proposal_oracle_success": 0.6,
+    "proposal_oracle_utility": 1.4122399787108104,
+    "selected_candidate_index": 7.2,
+    "selected_success": 0.26666666666666666,
+    "selected_success_gain_over_base": 0.06666666666666667,
+    "selected_utility": 0.9396337101856868,
+    "selected_utility_gain_over_base": -0.010236828525861105,
+    "selector_regret": 0.47343848447004955,
+    "success_selector_gap": 0.3333333333333333,
+    "success_support_gap": 0.4,
+    "support_gap": 0.5877600212891897,
+    "tau": 0.028235521236802175
+  },
+  "PickCube-v1": {
+    "base_success": 0.0,
+    "base_utility": 0.5839591026306152,
+    "coverage": 0.0,
+    "execute_generated": 0.0,
+    "fallback_rate": 1.0,
+    "hidden_chart_oracle_success": 1.0,
+    "hidden_chart_oracle_utility": 2.0,
+    "outcome_ptr": 0.0,
+    "pairwise_causal_calibration_accuracy": 0.3796296296296296,
+    "pairwise_causal_calibration_confidence": 0.5397879799473912,
+    "pairwise_causal_calibration_ece": 0.16015835031776168,
+    "pairwise_causal_calibration_pairs": 36.0,
+    "predicted_margin": -0.2678354586918643,
+    "proposal_oracle_success": 0.0,
+    "proposal_oracle_utility": 0.13425676027933756,
+    "selected_candidate_index": 1.3333333333333333,
+    "selected_success": 0.0,
+    "selected_success_gain_over_base": 0.0,
+    "selected_utility": 0.5839591026306152,
+    "selected_utility_gain_over_base": 0.0,
+    "selector_regret": 0.0,
+    "success_selector_gap": 0.0,
+    "success_support_gap": 1.0,
+    "support_gap": 1.8657432397206624,
+    "tau": 0.5025584143631503
+  },
+  "PullCube-v1": {
+    "base_success": 0.36,
+    "base_utility": 0.7751704612234608,
+    "coverage": 0.8133333333333334,
+    "execute_generated": 0.8133333333333334,
+    "fallback_rate": 0.18666666666666668,
+    "hidden_chart_oracle_success": 0.52,
+    "hidden_chart_oracle_utility": 1.1467878860235214,
+    "outcome_ptr": 0.4666666666666667,
+    "pairwise_causal_calibration_accuracy": 0.5439315887395845,
+    "pairwise_causal_calibration_confidence": 0.5315713773974821,
+    "pairwise_causal_calibration_ece": 0.1202240575596694,
+    "pairwise_causal_calibration_pairs": 92.32,
+    "predicted_margin": 0.0587216273047277,
+    "proposal_oracle_success": 0.5333333333333333,
+    "proposal_oracle_utility": 1.1093241602430741,
+    "selected_candidate_index": 6.16,
+    "selected_success": 0.32,
+    "selected_success_gain_over_base": -0.04,
+    "selected_utility": 0.6822500004371007,
+    "selected_utility_gain_over_base": -0.09292046078636,
+    "selector_regret": 0.4528969661643108,
+    "success_selector_gap": 0.22666666666666666,
+    "success_support_gap": 0.02666666666666667,
+    "support_gap": 0.1075624164690574,
+    "tau": -0.2012384152043576
+  },
+  "PushCube-v1": {
+    "base_success": 0.16666666666666666,
+    "base_utility": 0.3924041787783305,
+    "coverage": 0.7777777777777778,
+    "execute_generated": 0.7777777777777778,
+    "fallback_rate": 0.2222222222222222,
+    "hidden_chart_oracle_success": 0.8333333333333334,
+    "hidden_chart_oracle_utility": 1.7015800600250561,
+    "outcome_ptr": 0.8333333333333334,
+    "pairwise_causal_calibration_accuracy": 0.5408758312375324,
+    "pairwise_causal_calibration_confidence": 0.5154259344288218,
+    "pairwise_causal_calibration_ece": 0.09497784791093052,
+    "pairwise_causal_calibration_pairs": 97.44444444444444,
+    "predicted_margin": 0.32846837901195985,
+    "proposal_oracle_success": 0.8888888888888888,
+    "proposal_oracle_utility": 1.7857814149724112,
+    "selected_candidate_index": 5.222222222222222,
+    "selected_success": 0.5,
+    "selected_success_gain_over_base": 0.3333333333333333,
+    "selected_utility": 1.0611259018381436,
+    "selected_utility_gain_over_base": 0.6687217230598131,
+    "selector_regret": 0.7246555131342676,
+    "success_selector_gap": 0.3888888888888889,
+    "success_support_gap": 0.1111111111111111,
+    "support_gap": 0.21421858502758873,
+    "tau": 0.0742008167559442
+  },
+  "StackCube-v1": {
+    "base_success": 0.2727272727272727,
+    "base_utility": 0.7763947045261209,
+    "coverage": 0.21212121212121213,
+    "execute_generated": 0.21212121212121213,
+    "fallback_rate": 0.7878787878787878,
+    "hidden_chart_oracle_success": 1.0,
+    "hidden_chart_oracle_utility": 2.0,
+    "outcome_ptr": 0.6363636363636364,
+    "pairwise_causal_calibration_accuracy": 0.4871933998564077,
+    "pairwise_causal_calibration_confidence": 0.5389990977485094,
+    "pairwise_causal_calibration_ece": 0.10659066999978553,
+    "pairwise_causal_calibration_pairs": 67.45454545454545,
+    "predicted_margin": 0.03761618987646351,
+    "proposal_oracle_success": 0.5151515151515151,
+    "proposal_oracle_utility": 1.1567951707225856,
+    "selected_candidate_index": 5.606060606060606,
+    "selected_success": 0.30303030303030304,
+    "selected_success_gain_over_base": 0.030303030303030304,
+    "selected_utility": 0.769150389414845,
+    "selected_utility_gain_over_base": -0.007244315111275875,
+    "selector_regret": 0.3912377664537141,
+    "success_selector_gap": 0.21212121212121213,
+    "success_support_gap": 0.48484848484848486,
+    "support_gap": 0.8432048292774142,
+    "tau": 0.13235082030965997
+  }
+}

runs/ctt_base_context_obs_learned_dominance_score_chartcompat_utility_task_envclip_k16_train_to_test/split_hash.txt

@@ -0,0 +1,14 @@
+{
+  "calibration_target_index": {
+    "content_hash": "67f7b4a692f7c7e71da24378dc71e3399c5d35afc14c494869eba6087b765c42",
+    "retrieval_index_allowed": true,
+    "split": "train",
+    "split_hash": "867a94a1c67686c082765671a7bf8d74a9d633085e1b2619c83197c60bf58372"
+  },
+  "eval_target_index": {
+    "content_hash": "0c38d73dc9bc4ab756e12006671b5d12f1a23358b281e54b3f6a9ae3cbf6e9af",
+    "retrieval_index_allowed": false,
+    "split": "test",
+    "split_hash": "84151df9e88c219b76751677558cc5bbbc19082bf358cea7330158e9e403e25f"
+  }
+}

runs/ctt_base_context_obs_learned_dominance_score_chartcompat_utility_task_envclip_k16_train_to_test/table.tex

@@ -0,0 +1,8 @@
+% Auto-generated by scripts/eval_learned_dominance_selector.py
+\begin{tabular}{lrrrrrrrrr}
+\toprule
+Rows & Coverage & Fallback & Base succ. & Selected succ. & Oracle succ. & OutcomePTR & Succ. support gap & Succ. selector gap & Cal. ECE \\
+\midrule
+144 & 0.6389 & 0.3611 & 0.2917 & 0.3264 & 0.5694 & 0.5486 & 0.2014 & 0.2500 & 0.0221 \\
+\bottomrule
+\end{tabular}

runs/ctt_base_context_obs_learned_dominance_score_chartcompat_utility_task_envclip_k16_train_to_test/train.log

@@ -0,0 +1,5 @@
+trained ridge dominance calibrator on calibration measured rows only
+fit_objective=pointwise
+selected_lambda=100.0
+tau={"LiftPegUpright-v1": 0.028236, "PickCube-v1": 0.502558, "PullCube-v1": -0.201238, "PushCube-v1": 0.074201, "StackCube-v1": 0.132351, "__global__": 0.018529}
+calibration_input=runs/ctt_base_context_obs_train_cal_envclip_k16_rollout_comparison/combined_measured_candidates.json