anhtld commited on
Commit
a7abd6e
·
verified ·
1 Parent(s): 0bf166e

auto-sync 2026-07-03T15:42:31Z workspace (part 3)

Browse files
workspace/runs/ctt_val_proxy_comparison/eval.log CHANGED
@@ -2,4 +2,5 @@ local_atlas: runs/local_atlas_val_proxy
2
  task_memory: runs/task_memory_val_proxy
3
  ctt_residual_full: runs/ctt_residual_full_seed0_val_proxy,runs/ctt_residual_full_seed1_val_proxy,runs/ctt_residual_full_seed2_val_proxy
4
  ctt_residual_base_context: runs/ctt_residual_base_context_seed0_val_proxy
 
5
  ctt_gated_residual_full: runs/ctt_gated_residual_full_seed0_val_proxy,runs/ctt_gated_residual_full_seed1_val_proxy,runs/ctt_gated_residual_full_seed2_val_proxy
 
2
  task_memory: runs/task_memory_val_proxy
3
  ctt_residual_full: runs/ctt_residual_full_seed0_val_proxy,runs/ctt_residual_full_seed1_val_proxy,runs/ctt_residual_full_seed2_val_proxy
4
  ctt_residual_base_context: runs/ctt_residual_base_context_seed0_val_proxy
5
+ ctt_residual_base_context_obs: runs/ctt_residual_base_context_obs_seed0_val_proxy
6
  ctt_gated_residual_full: runs/ctt_gated_residual_full_seed0_val_proxy,runs/ctt_gated_residual_full_seed1_val_proxy,runs/ctt_gated_residual_full_seed2_val_proxy
workspace/runs/ctt_val_proxy_comparison/metrics.json CHANGED
@@ -75,6 +75,24 @@
75
  "split_hash": "9d1dc1c10868f76fa080895df0f2730c173f12b6db14932599d55c2b60859381",
76
  "train_seeds": 1
77
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
78
  {
79
  "beats_local_atlas_support": true,
80
  "candidate_diversity": 0.1163743674776266,
 
75
  "split_hash": "9d1dc1c10868f76fa080895df0f2730c173f12b6db14932599d55c2b60859381",
76
  "train_seeds": 1
77
  },
78
+ {
79
+ "beats_local_atlas_support": true,
80
+ "candidate_diversity": 0.2363007070141511,
81
+ "data_hash": "1d15143588697e89f7fc2f6375b2745e4d479c726c9a947ff7eb7b2705280e1a",
82
+ "mean_positive_distance": 0.44652032547298753,
83
+ "method": "ctt_residual_base_context_obs",
84
+ "negative_near_0p20": 0.020186335403726708,
85
+ "negative_near_0p40": 0.2843300419387376,
86
+ "num_rows": 69,
87
+ "pptc_0p20": 0.17391304347826086,
88
+ "pptc_0p40": 0.5942028985507246,
89
+ "proxy_gate_pass": true,
90
+ "proxy_support_distance": 0.37439867476717087,
91
+ "run_path": "runs/ctt_residual_base_context_obs_seed0_val_proxy",
92
+ "safety_ok_vs_local_atlas": true,
93
+ "split_hash": "9d1dc1c10868f76fa080895df0f2730c173f12b6db14932599d55c2b60859381",
94
+ "train_seeds": 1
95
+ },
96
  {
97
  "beats_local_atlas_support": true,
98
  "candidate_diversity": 0.1163743674776266,
workspace/runs/ctt_val_proxy_comparison/report.md CHANGED
@@ -8,6 +8,7 @@ Safety slack over local-atlas NegativeNear@0.20: `0.0100`
8
  | task_memory | 1 | 69 | 0.3188 | 0.5362 | 0.0175 | 0.9616 | 0.9751 | baseline | `runs/task_memory_val_proxy` |
9
  | ctt_residual_full | 3 | 207 | 0.1981 | 0.6087 | 0.0296 | 0.4509 | 0.2703 | pass | `runs/ctt_residual_full_seed0_val_proxy,runs/ctt_residual_full_seed1_val_proxy,runs/ctt_residual_full_seed2_val_proxy` |
10
  | ctt_residual_base_context | 1 | 69 | 0.1739 | 0.6232 | 0.0182 | 0.4429 | 0.2471 | pass | `runs/ctt_residual_base_context_seed0_val_proxy` |
 
11
  | ctt_gated_residual_full | 3 | 207 | 0.2319 | 0.6135 | 0.0527 | 0.4337 | 0.1164 | fail | `runs/ctt_gated_residual_full_seed0_val_proxy,runs/ctt_gated_residual_full_seed1_val_proxy,runs/ctt_gated_residual_full_seed2_val_proxy` |
12
 
13
  Gate is proxy-only. It is not OutcomePTR or measured rollout success.
 
8
  | task_memory | 1 | 69 | 0.3188 | 0.5362 | 0.0175 | 0.9616 | 0.9751 | baseline | `runs/task_memory_val_proxy` |
9
  | ctt_residual_full | 3 | 207 | 0.1981 | 0.6087 | 0.0296 | 0.4509 | 0.2703 | pass | `runs/ctt_residual_full_seed0_val_proxy,runs/ctt_residual_full_seed1_val_proxy,runs/ctt_residual_full_seed2_val_proxy` |
10
  | ctt_residual_base_context | 1 | 69 | 0.1739 | 0.6232 | 0.0182 | 0.4429 | 0.2471 | pass | `runs/ctt_residual_base_context_seed0_val_proxy` |
11
+ | ctt_residual_base_context_obs | 1 | 69 | 0.1739 | 0.5942 | 0.0202 | 0.4465 | 0.2363 | pass | `runs/ctt_residual_base_context_obs_seed0_val_proxy` |
12
  | ctt_gated_residual_full | 3 | 207 | 0.2319 | 0.6135 | 0.0527 | 0.4337 | 0.1164 | fail | `runs/ctt_gated_residual_full_seed0_val_proxy,runs/ctt_gated_residual_full_seed1_val_proxy,runs/ctt_gated_residual_full_seed2_val_proxy` |
13
 
14
  Gate is proxy-only. It is not OutcomePTR or measured rollout success.
workspace/runs/ctt_val_proxy_comparison/table.tex CHANGED
@@ -7,6 +7,7 @@ local\_atlas & 1 & 69 & 0.4058 & 0.6812 & 0.0368 & 0.7203 & 0.8219 & baseline \\
7
  task\_memory & 1 & 69 & 0.3188 & 0.5362 & 0.0175 & 0.9616 & 0.9751 & baseline \\
8
  ctt\_residual\_full & 3 & 207 & 0.1981 & 0.6087 & 0.0296 & 0.4509 & 0.2703 & pass \\
9
  ctt\_residual\_base\_context & 1 & 69 & 0.1739 & 0.6232 & 0.0182 & 0.4429 & 0.2471 & pass \\
 
10
  ctt\_gated\_residual\_full & 3 & 207 & 0.2319 & 0.6135 & 0.0527 & 0.4337 & 0.1164 & fail \\
11
  \bottomrule
12
  \end{tabular}
 
7
  task\_memory & 1 & 69 & 0.3188 & 0.5362 & 0.0175 & 0.9616 & 0.9751 & baseline \\
8
  ctt\_residual\_full & 3 & 207 & 0.1981 & 0.6087 & 0.0296 & 0.4509 & 0.2703 & pass \\
9
  ctt\_residual\_base\_context & 1 & 69 & 0.1739 & 0.6232 & 0.0182 & 0.4429 & 0.2471 & pass \\
10
+ ctt\_residual\_base\_context\_obs & 1 & 69 & 0.1739 & 0.5942 & 0.0202 & 0.4465 & 0.2363 & pass \\
11
  ctt\_gated\_residual\_full & 3 & 207 & 0.2319 & 0.6135 & 0.0527 & 0.4337 & 0.1164 & fail \\
12
  \bottomrule
13
  \end{tabular}
workspace/runs/summary_ctt.csv CHANGED
@@ -9,6 +9,8 @@ CTT gated residual,runs/ctt_gated_residual_full_seed2_val_proxy,yes,proxy,not me
9
  CTT gated residual,runs/ctt_gated_residual_smoke,yes,train,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,unknown,see run
10
  CTT gated residual,runs/ctt_gated_residual_smoke_proxy,yes,proxy,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.5000,0.7500,0.0573,n/a,unknown,see run
11
  CTT gated residual,runs/ctt_gated_residual_val_proxy,yes,proxy,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.1739,0.5362,0.0281,n/a,47,see run
 
 
12
  CTT residual base-context,runs/ctt_residual_base_context_seed0,yes,train feature=base_context,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,unknown,see run
13
  CTT residual base-context,runs/ctt_residual_base_context_seed0_val_proxy,yes,proxy feature=base_context,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.1739,0.6232,0.0182,n/a,47,see run
14
  CTT residual train,runs/ctt_residual_full_seed0,yes,train,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,unknown,see run
@@ -35,6 +37,7 @@ task memory,runs/task_memory_val_proxy,yes,proxy baseline,not measured,n/a,n/a,n
35
  local atlas,runs/local_atlas_val_proxy,yes,proxy baseline,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.4058,0.6812,0.0368,n/a,47,see run
36
  ctt residual full,runs/ctt_val_proxy_comparison,yes,proxy aggregate gate=True,proxy=True,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.1981,0.6087,0.0296,n/a,3,aggregate over run means
37
  ctt residual base context,runs/ctt_val_proxy_comparison,yes,proxy aggregate gate=True,proxy=True,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.1739,0.6232,0.0182,n/a,1,aggregate over run means
 
38
  ctt gated residual full,runs/ctt_val_proxy_comparison,yes,proxy aggregate gate=False,proxy=False,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.2319,0.6135,0.0527,n/a,3,aggregate over run means
39
  Utility energy smoke,runs/utility_energy_smoke,train-only,calibration diagnostic,not deployment,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.1316,train split,see run
40
  CTT residual test utility-energy dominance,runs/ctt_dominance_utility_energy_smoke_val_to_test,yes,measured dominance K=8 coverage=0.2847 fallback=0.7153,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2847,0.2222,0.5069,0.7292,0.4833,0.5744,0.2639,0.3264,0.5278,n/a,n/a,n/a,n/a,3,see run
 
9
  CTT gated residual,runs/ctt_gated_residual_smoke,yes,train,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,unknown,see run
10
  CTT gated residual,runs/ctt_gated_residual_smoke_proxy,yes,proxy,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.5000,0.7500,0.0573,n/a,unknown,see run
11
  CTT gated residual,runs/ctt_gated_residual_val_proxy,yes,proxy,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.1739,0.5362,0.0281,n/a,47,see run
12
+ CTT residual base-context-observation,runs/ctt_residual_base_context_obs_seed0,yes,train feature=base_context_obs,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,unknown,see run
13
+ CTT residual base-context-observation,runs/ctt_residual_base_context_obs_seed0_val_proxy,yes,proxy feature=base_context_obs,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.1739,0.5942,0.0202,n/a,47,see run
14
  CTT residual base-context,runs/ctt_residual_base_context_seed0,yes,train feature=base_context,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,unknown,see run
15
  CTT residual base-context,runs/ctt_residual_base_context_seed0_val_proxy,yes,proxy feature=base_context,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.1739,0.6232,0.0182,n/a,47,see run
16
  CTT residual train,runs/ctt_residual_full_seed0,yes,train,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,unknown,see run
 
37
  local atlas,runs/local_atlas_val_proxy,yes,proxy baseline,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.4058,0.6812,0.0368,n/a,47,see run
38
  ctt residual full,runs/ctt_val_proxy_comparison,yes,proxy aggregate gate=True,proxy=True,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.1981,0.6087,0.0296,n/a,3,aggregate over run means
39
  ctt residual base context,runs/ctt_val_proxy_comparison,yes,proxy aggregate gate=True,proxy=True,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.1739,0.6232,0.0182,n/a,1,aggregate over run means
40
+ ctt residual base context obs,runs/ctt_val_proxy_comparison,yes,proxy aggregate gate=True,proxy=True,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.1739,0.5942,0.0202,n/a,1,aggregate over run means
41
  ctt gated residual full,runs/ctt_val_proxy_comparison,yes,proxy aggregate gate=False,proxy=False,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.2319,0.6135,0.0527,n/a,3,aggregate over run means
42
  Utility energy smoke,runs/utility_energy_smoke,train-only,calibration diagnostic,not deployment,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,0.1316,train split,see run
43
  CTT residual test utility-energy dominance,runs/ctt_dominance_utility_energy_smoke_val_to_test,yes,measured dominance K=8 coverage=0.2847 fallback=0.7153,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2847,0.2222,0.5069,0.7292,0.4833,0.5744,0.2639,0.3264,0.5278,n/a,n/a,n/a,n/a,3,see run
workspace/runs/summary_ctt.md CHANGED
@@ -12,6 +12,8 @@
12
  | CTT gated residual | train; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | `runs/ctt_gated_residual_smoke` |
13
  | CTT gated residual | proxy; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.5000 | 0.7500 | 0.0573 | n/a | `runs/ctt_gated_residual_smoke_proxy` |
14
  | CTT gated residual | proxy; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.1739 | 0.5362 | 0.0281 | n/a | `runs/ctt_gated_residual_val_proxy` |
 
 
15
  | CTT residual base-context | train feature=base_context; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | `runs/ctt_residual_base_context_seed0` |
16
  | CTT residual base-context | proxy feature=base_context; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.1739 | 0.6232 | 0.0182 | n/a | `runs/ctt_residual_base_context_seed0_val_proxy` |
17
  | CTT residual train | train; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | `runs/ctt_residual_full_seed0` |
@@ -38,6 +40,7 @@
38
  | local atlas | proxy baseline; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.4058 | 0.6812 | 0.0368 | n/a | `runs/local_atlas_val_proxy` |
39
  | ctt residual full | proxy aggregate gate=True; gate: proxy=True | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.1981 | 0.6087 | 0.0296 | n/a | `runs/ctt_val_proxy_comparison` |
40
  | ctt residual base context | proxy aggregate gate=True; gate: proxy=True | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.1739 | 0.6232 | 0.0182 | n/a | `runs/ctt_val_proxy_comparison` |
 
41
  | ctt gated residual full | proxy aggregate gate=False; gate: proxy=False | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.2319 | 0.6135 | 0.0527 | n/a | `runs/ctt_val_proxy_comparison` |
42
  | Utility energy smoke | calibration diagnostic; gate: not deployment | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.1316 | `runs/utility_energy_smoke` |
43
  | CTT residual test utility-energy dominance | measured dominance K=8 coverage=0.2847 fallback=0.7153; gate: fail:selected<47.45,support_gap>7,selector_gap>3 | 0.2847 | 0.2222 | 0.5069 | 0.7292 | 0.4833 | 0.5744 | 0.2639 | 0.3264 | 0.5278 | n/a | n/a | n/a | n/a | `runs/ctt_dominance_utility_energy_smoke_val_to_test` |
 
12
  | CTT gated residual | train; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | `runs/ctt_gated_residual_smoke` |
13
  | CTT gated residual | proxy; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.5000 | 0.7500 | 0.0573 | n/a | `runs/ctt_gated_residual_smoke_proxy` |
14
  | CTT gated residual | proxy; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.1739 | 0.5362 | 0.0281 | n/a | `runs/ctt_gated_residual_val_proxy` |
15
+ | CTT residual base-context-observation | train feature=base_context_obs; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | `runs/ctt_residual_base_context_obs_seed0` |
16
+ | CTT residual base-context-observation | proxy feature=base_context_obs; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.1739 | 0.5942 | 0.0202 | n/a | `runs/ctt_residual_base_context_obs_seed0_val_proxy` |
17
  | CTT residual base-context | train feature=base_context; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | `runs/ctt_residual_base_context_seed0` |
18
  | CTT residual base-context | proxy feature=base_context; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.1739 | 0.6232 | 0.0182 | n/a | `runs/ctt_residual_base_context_seed0_val_proxy` |
19
  | CTT residual train | train; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | `runs/ctt_residual_full_seed0` |
 
40
  | local atlas | proxy baseline; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.4058 | 0.6812 | 0.0368 | n/a | `runs/local_atlas_val_proxy` |
41
  | ctt residual full | proxy aggregate gate=True; gate: proxy=True | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.1981 | 0.6087 | 0.0296 | n/a | `runs/ctt_val_proxy_comparison` |
42
  | ctt residual base context | proxy aggregate gate=True; gate: proxy=True | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.1739 | 0.6232 | 0.0182 | n/a | `runs/ctt_val_proxy_comparison` |
43
+ | ctt residual base context obs | proxy aggregate gate=True; gate: proxy=True | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.1739 | 0.5942 | 0.0202 | n/a | `runs/ctt_val_proxy_comparison` |
44
  | ctt gated residual full | proxy aggregate gate=False; gate: proxy=False | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.2319 | 0.6135 | 0.0527 | n/a | `runs/ctt_val_proxy_comparison` |
45
  | Utility energy smoke | calibration diagnostic; gate: not deployment | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.1316 | `runs/utility_energy_smoke` |
46
  | CTT residual test utility-energy dominance | measured dominance K=8 coverage=0.2847 fallback=0.7153; gate: fail:selected<47.45,support_gap>7,selector_gap>3 | 0.2847 | 0.2222 | 0.5069 | 0.7292 | 0.4833 | 0.5744 | 0.2639 | 0.3264 | 0.5278 | n/a | n/a | n/a | n/a | `runs/ctt_dominance_utility_energy_smoke_val_to_test` |
workspace/scripts/build_ctt_proxy_comparison.py CHANGED
@@ -20,6 +20,10 @@ DEFAULT_RUNS = [
20
  "ctt_residual_base_context",
21
  sorted(Path("runs").glob("ctt_residual_base_context_seed*_val_proxy")),
22
  ),
 
 
 
 
23
  (
24
  "ctt_gated_residual_full",
25
  sorted(Path("runs").glob("ctt_gated_residual_full_seed*_val_proxy")),
 
20
  "ctt_residual_base_context",
21
  sorted(Path("runs").glob("ctt_residual_base_context_seed*_val_proxy")),
22
  ),
23
+ (
24
+ "ctt_residual_base_context_obs",
25
+ sorted(Path("runs").glob("ctt_residual_base_context_obs_seed*_val_proxy")),
26
+ ),
27
  (
28
  "ctt_gated_residual_full",
29
  sorted(Path("runs").glob("ctt_gated_residual_full_seed*_val_proxy")),