Hugging Face's logo

JulianHJR
/
studentstimulation

Model card Files
xet
 Community
default commited on
Commit
7b93e9e
·
1 Parent(s): 8b11c2b

upload logs and results

Browse files
Files changed (19) hide show
  1. .gitignore +2 -0
  2. logs/01_download.log +22 -0
  3. logs/02_generate.log +51 -0
  4. logs/03_label.log +12 -0
  5. logs/04_routing.log +18 -0
  6. logs/05_select.log +14 -0
  7. logs/06_interaction.log +12 -0
  8. logs/07_residuals.log +203 -0
  9. logs/08_directions.log +272 -0
  10. logs/09_sweep.log +28 -0
  11. logs/12_downstream.log +20 -0
  12. logs/13_analyze.log +4 -0
  13. logs/runall.log +0 -0
  14. results/direction_cosines.json +992 -0
  15. results/downstream_accuracy.json +1534 -0
  16. results/final_report.md +90 -0
  17. results/interaction_summary.json +205 -0
  18. results/sweep_log.jsonl +0 -0
  19. results/sweep_rr.json +778 -0
.gitignore ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ results/*.png
2
+ *.npz
logs/01_download.log ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [2026-04-28 14:18:01] INFO 01_download: Target: /data/home/user-test/jrhu/studentstimulation/data/models/Qwen3-30B-A3B-Thinking-2507
2
+ [2026-04-28 14:18:01] INFO 01_download: Source: Qwen/Qwen3-30B-A3B-Thinking-2507
3
+ [2026-04-28 14:18:02] INFO 01_download: Starting download (Qwen3-30B ~60GB, will take time)...
4
+ [2026-04-28 14:20:18] INFO 01_download: Download complete: /data/home/user-test/jrhu/studentstimulation/data/models/Qwen3-30B-A3B-Thinking-2507
5
+ [2026-04-28 14:26:58] INFO 01_download: Target: /data/home/user-test/jrhu/studentstimulation/data/models/Qwen3-30B-A3B-Thinking-2507
6
+ [2026-04-28 14:26:58] INFO 01_download: Source: Qwen/Qwen3-30B-A3B-Thinking-2507
7
+ [2026-04-28 14:26:58] INFO 01_download: Weights already present, skipping download.
8
+ [2026-04-28 14:47:01] INFO 01_download: Target: /data/home/user-test/jrhu/studentstimulation/data/models/Qwen3-30B-A3B-Thinking-2507
9
+ [2026-04-28 14:47:01] INFO 01_download: Source: Qwen/Qwen3-30B-A3B-Thinking-2507
10
+ [2026-04-28 14:47:01] INFO 01_download: Weights already present, skipping download.
11
+ [2026-04-28 14:51:51] INFO 01_download: Target: /data/home/user-test/jrhu/studentstimulation/data/models/Qwen3-30B-A3B-Thinking-2507
12
+ [2026-04-28 14:51:51] INFO 01_download: Source: Qwen/Qwen3-30B-A3B-Thinking-2507
13
+ [2026-04-28 14:51:51] INFO 01_download: Weights already present, skipping download.
14
+ [2026-04-28 14:54:38] INFO 01_download: Target: /data/home/user-test/jrhu/studentstimulation/data/models/Qwen3-30B-A3B-Thinking-2507
15
+ [2026-04-28 14:54:38] INFO 01_download: Source: Qwen/Qwen3-30B-A3B-Thinking-2507
16
+ [2026-04-28 14:54:38] INFO 01_download: Weights already present, skipping download.
17
+ [2026-04-28 15:13:01] INFO 01_download: Target: /data/home/user-test/jrhu/studentstimulation/data/models/Qwen3-30B-A3B-Thinking-2507
18
+ [2026-04-28 15:13:01] INFO 01_download: Source: Qwen/Qwen3-30B-A3B-Thinking-2507
19
+ [2026-04-28 15:13:01] INFO 01_download: Weights already present, skipping download.
20
+ [2026-04-29 09:46:14] INFO 01_download: Target: /data/home/user-test/jrhu/studentstimulation/data/models/Qwen3-30B-A3B-Thinking-2507
21
+ [2026-04-29 09:46:14] INFO 01_download: Source: Qwen/Qwen3-30B-A3B-Thinking-2507
22
+ [2026-04-29 09:46:14] INFO 01_download: Weights already present, skipping download.
logs/02_generate.log ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [2026-04-28 14:20:21] INFO 02_generate: n_train=150 n_math_test=50 n_aime=30 n_gpqa=20 max_new_tokens=4096
2
+ [2026-04-28 14:20:25] INFO 02_generate: MATH-500: train=150, holdout_test=50
3
+ [2026-04-28 14:20:29] INFO 02_generate: AIME-24: 30 problems
4
+ [2026-04-28 14:20:29] INFO 02_generate: GPQA-D: 0 problems
5
+ [2026-04-28 14:20:29] INFO 02_generate: Loading model (this will take a few minutes)...
6
+ [2026-04-28 14:20:52] INFO 02_generate: Model loaded.
7
+ [2026-04-28 14:20:52] INFO 02_generate: Training set: 0 already completed, 150 remaining
8
+ [2026-04-28 14:27:01] INFO 02_generate: n_train=150 n_math_test=50 n_aime=30 n_gpqa=20 max_new_tokens=4096
9
+ [2026-04-28 14:27:04] INFO 02_generate: MATH-500: train=150, holdout_test=50
10
+ [2026-04-28 14:27:07] INFO 02_generate: AIME-24: 30 problems
11
+ [2026-04-28 14:27:07] INFO 02_generate: GPQA-D: 0 problems
12
+ [2026-04-28 14:27:07] INFO 02_generate: Loading model (this will take a few minutes)...
13
+ [2026-04-28 14:27:30] INFO 02_generate: Model loaded.
14
+ [2026-04-28 14:27:30] INFO 02_generate: Training set: 0 already completed, 150 remaining
15
+ [2026-04-28 14:47:05] INFO 02_generate: n_train=150 n_math_test=50 n_aime=30 n_gpqa=20 max_new_tokens=4096
16
+ [2026-04-28 14:47:08] INFO 02_generate: MATH-500: train=150, holdout_test=50
17
+ [2026-04-28 14:47:11] INFO 02_generate: AIME-24: 30 problems
18
+ [2026-04-28 14:47:11] INFO 02_generate: GPQA-D: 0 problems
19
+ [2026-04-28 14:47:11] INFO 02_generate: Loading model (this will take a few minutes)...
20
+ [2026-04-28 14:51:55] INFO 02_generate: n_train=150 n_math_test=50 n_aime=30 n_gpqa=20 max_new_tokens=4096
21
+ [2026-04-28 14:51:58] INFO 02_generate: MATH-500: train=150, holdout_test=50
22
+ [2026-04-28 14:52:00] INFO 02_generate: AIME-24: 30 problems
23
+ [2026-04-28 14:52:00] INFO 02_generate: GPQA-D: 0 problems
24
+ [2026-04-28 14:52:00] INFO 02_generate: Loading model (this will take a few minutes)...
25
+ [2026-04-28 14:54:41] INFO 02_generate: n_train=150 n_math_test=50 n_aime=30 n_gpqa=20 max_new_tokens=4096
26
+ [2026-04-28 14:54:44] INFO 02_generate: MATH-500: train=150, holdout_test=50
27
+ [2026-04-28 14:54:47] INFO 02_generate: AIME-24: 30 problems
28
+ [2026-04-28 14:54:47] INFO 02_generate: GPQA-D: 0 problems
29
+ [2026-04-28 14:54:47] INFO 02_generate: Loading model (this will take a few minutes)...
30
+ [2026-04-28 15:13:05] INFO 02_generate: n_train=150 n_math_test=50 n_aime=30 n_gpqa=20 max_new_tokens=4096
31
+ [2026-04-28 15:13:07] INFO 02_generate: MATH-500: train=150, holdout_test=50
32
+ [2026-04-28 15:13:10] INFO 02_generate: AIME-24: 30 problems
33
+ [2026-04-28 15:13:10] INFO 02_generate: GPQA-D: 0 problems
34
+ [2026-04-28 15:13:10] INFO 02_generate: Loading model (this will take a few minutes)...
35
+ [2026-04-28 15:13:26] INFO 02_generate: Model loaded.
36
+ [2026-04-28 15:13:26] INFO 02_generate: Training set: 0 already completed, 150 remaining
37
+ [2026-04-28 19:54:41] INFO 02_generate: MATH-500-holdout: 50 problems saved to /data/home/user-test/jrhu/studentstimulation/data/cots/test_math500_holdout.jsonl
38
+ [2026-04-28 19:54:41] INFO 02_generate: AIME-24: 30 problems saved to /data/home/user-test/jrhu/studentstimulation/data/cots/test_aime24.jsonl
39
+ [2026-04-28 19:54:41] INFO 02_generate: ============================================================
40
+ [2026-04-28 19:54:41] INFO 02_generate: Done. Training CoTs saved to /data/home/user-test/jrhu/studentstimulation/data/cots/raw_cots.jsonl
41
+ [2026-04-29 09:46:18] INFO 02_generate: n_train=150 n_math_test=50 n_aime=30 n_gpqa=20 max_new_tokens=4096
42
+ [2026-04-29 09:46:21] INFO 02_generate: MATH-500: train=150, holdout_test=50
43
+ [2026-04-29 09:46:24] INFO 02_generate: AIME-24: 30 problems
44
+ [2026-04-29 09:46:24] INFO 02_generate: GPQA-D: 0 problems
45
+ [2026-04-29 09:46:24] INFO 02_generate: Loading model (this will take a few minutes)...
46
+ [2026-04-29 09:46:45] INFO 02_generate: Model loaded.
47
+ [2026-04-29 09:46:45] INFO 02_generate: Training set: 150 already completed, 0 remaining
48
+ [2026-04-29 09:46:45] INFO 02_generate: MATH-500-holdout: test set already saved at /data/home/user-test/jrhu/studentstimulation/data/cots/test_math500_holdout.jsonl
49
+ [2026-04-29 09:46:45] INFO 02_generate: AIME-24: test set already saved at /data/home/user-test/jrhu/studentstimulation/data/cots/test_aime24.jsonl
50
+ [2026-04-29 09:46:45] INFO 02_generate: ============================================================
51
+ [2026-04-29 09:46:45] INFO 02_generate: Done. Training CoTs saved to /data/home/user-test/jrhu/studentstimulation/data/cots/raw_cots.jsonl
logs/03_label.log ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [2026-04-28 19:54:46] INFO 03_label: Loading tokenizer: /data/home/user-test/jrhu/studentstimulation/data/models/Qwen3-30B-A3B-Thinking-2507
2
+ [2026-04-28 19:54:46] INFO 03_label: Reading CoTs: /data/home/user-test/jrhu/studentstimulation/data/cots/raw_cots.jsonl
3
+ [2026-04-28 19:54:46] INFO 03_label: Got 150 CoTs
4
+ [2026-04-28 19:54:50] INFO 03_label: ============================================================
5
+ [2026-04-28 19:54:50] INFO 03_label: LABELING SUMMARY
6
+ [2026-04-28 19:54:50] INFO 03_label: N CoTs: 150
7
+ [2026-04-28 19:54:50] INFO 03_label: Total plan decision points: 259
8
+ [2026-04-28 19:54:50] INFO 03_label: Total mon decision points: 691
9
+ [2026-04-28 19:54:50] INFO 03_label: Total exec decision points: 2287
10
+ [2026-04-28 19:54:50] INFO 03_label: Total newlines overall: 6479
11
+ [2026-04-28 19:54:50] INFO 03_label: Saved -> /data/home/user-test/jrhu/studentstimulation/data/cots/labeled_cots.jsonl
12
+ [2026-04-29 09:46:50] INFO 03_label: Labeled file exists: /data/home/user-test/jrhu/studentstimulation/data/cots/labeled_cots.jsonl. Skipping.
logs/04_routing.log ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [2026-04-28 19:54:54] INFO 04_routing: Reading labeled CoTs: /data/home/user-test/jrhu/studentstimulation/data/cots/labeled_cots.jsonl
2
+ [2026-04-28 19:54:54] INFO 04_routing: Got 150 CoTs
3
+ [2026-04-28 19:54:54] INFO 04_routing: Will produce 3 shards of up to 50 CoTs each
4
+ [2026-04-28 19:54:54] INFO 04_routing: Shards to compute: 3 / 3
5
+ [2026-04-28 19:54:54] INFO 04_routing: Loading model...
6
+ [2026-04-28 19:55:08] INFO 04_routing: Model loaded. VRAM: 58285 MB
7
+ [2026-04-28 19:55:08] INFO 04_routing: [shard 0] processing records 0:50
8
+ [2026-04-28 19:55:33] INFO 04_routing: [shard 0] saved /data/home/user-test/jrhu/studentstimulation/data/routing/shard_0000.pt: 50 CoTs, 133223 total tokens, 0.20 GB
9
+ [2026-04-28 19:55:33] INFO 04_routing: [shard 1] processing records 50:100
10
+ [2026-04-28 19:55:57] INFO 04_routing: [shard 1] saved /data/home/user-test/jrhu/studentstimulation/data/routing/shard_0001.pt: 50 CoTs, 125162 total tokens, 0.19 GB
11
+ [2026-04-28 19:55:57] INFO 04_routing: [shard 2] processing records 100:150
12
+ [2026-04-28 19:56:23] INFO 04_routing: [shard 2] saved /data/home/user-test/jrhu/studentstimulation/data/routing/shard_0002.pt: 50 CoTs, 147952 total tokens, 0.23 GB
13
+ [2026-04-28 19:56:23] INFO 04_routing: ============================================================
14
+ [2026-04-28 19:56:23] INFO 04_routing: Routing capture complete.
15
+ [2026-04-29 09:46:54] INFO 04_routing: Reading labeled CoTs: /data/home/user-test/jrhu/studentstimulation/data/cots/labeled_cots.jsonl
16
+ [2026-04-29 09:46:54] INFO 04_routing: Got 150 CoTs
17
+ [2026-04-29 09:46:54] INFO 04_routing: Will produce 3 shards of up to 50 CoTs each
18
+ [2026-04-29 09:46:54] INFO 04_routing: All shards already exist. Skipping.
logs/05_select.log ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [2026-04-28 19:56:25] INFO 05_select: Loading routing shards...
2
+ [2026-04-28 19:56:26] INFO 05_select: Total tokens: 406337
3
+ [2026-04-28 19:56:26] INFO 05_select: Loading labels...
4
+ [2026-04-28 19:56:26] INFO 05_select: Global indices: plan=259, mon=691, exec=2287
5
+ [2026-04-28 19:56:26] INFO 05_select: Computing expert selection scores...
6
+ [2026-04-28 19:56:26] INFO 05_select: Top-32 planning experts saved: /data/home/user-test/jrhu/studentstimulation/data/checkpoints/top_experts_planning.json
7
+ [2026-04-28 19:56:26] INFO 05_select: Top-32 monitoring experts saved: /data/home/user-test/jrhu/studentstimulation/data/checkpoints/top_experts_monitoring.json
8
+ [2026-04-28 19:56:26] INFO 05_select: Target layers: planning=[0, 3, 4, 8, 9, 10, 11, 12, 13, 14, 15, 17, 18, 20, 21, 22, 24, 27, 28, 30, 32, 34, 36, 37, 38, 42, 47]
9
+ [2026-04-28 19:56:26] INFO 05_select: monitoring=[0, 2, 3, 6, 7, 8, 10, 13, 14, 15, 17, 20, 21, 23, 24, 25, 26, 27, 29, 31, 33, 35, 36, 37, 38, 39, 45]
10
+ [2026-04-28 19:56:26] INFO 05_select: union=[0, 2, 3, 4, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 17, 18, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 42, 45, 47]
11
+ [2026-04-28 19:56:26] INFO 05_select: Plotting routing heatmaps...
12
+ [2026-04-28 19:56:27] INFO 05_select: Saved raw stats -> routing_stats.npz
13
+ [2026-04-28 19:56:27] INFO 05_select: Done.
14
+ [2026-04-29 09:46:55] INFO 05_select: Top-experts already saved. Skipping.
logs/06_interaction.log ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [2026-04-28 19:56:29] INFO 06_interaction: Jaccard overlap (top-K experts): 0.000
2
+ [2026-04-28 19:56:29] INFO 06_interaction: Loading routing shards for PMI...
3
+ [2026-04-28 19:56:29] INFO 06_interaction: Computing PMI over 259 planning decision points for same-layer (plan_expert, mon_expert) pairs...
4
+ [2026-04-28 19:56:29] INFO 06_interaction: Computed PMI for 19 same-layer pairs
5
+ [2026-04-28 19:56:30] INFO 06_interaction: Saved interaction heatmap: /data/home/user-test/jrhu/studentstimulation/data/results/interaction_heatmap.png
6
+ [2026-04-28 19:56:30] INFO 06_interaction: Saved interaction summary: /data/home/user-test/jrhu/studentstimulation/data/results/interaction_summary.json
7
+ [2026-04-29 09:46:57] INFO 06_interaction: Jaccard overlap (top-K experts): 0.000
8
+ [2026-04-29 09:46:57] INFO 06_interaction: Loading routing shards for PMI...
9
+ [2026-04-29 09:46:57] INFO 06_interaction: Computing PMI over 259 planning decision points for same-layer (plan_expert, mon_expert) pairs...
10
+ [2026-04-29 09:46:57] INFO 06_interaction: Computed PMI for 19 same-layer pairs
11
+ [2026-04-29 09:46:58] INFO 06_interaction: Saved interaction heatmap: /data/home/user-test/jrhu/studentstimulation/data/results/interaction_heatmap.png
12
+ [2026-04-29 09:46:58] INFO 06_interaction: Saved interaction summary: /data/home/user-test/jrhu/studentstimulation/data/results/interaction_summary.json
logs/07_residuals.log ADDED
@@ -0,0 +1,203 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [2026-04-28 19:56:34] INFO 07_residuals: Target layers (39): [0, 2, 3, 4, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 17, 18, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 42, 45, 47]
2
+ [2026-04-28 19:56:34] INFO 07_residuals: Got 150 labeled CoTs
3
+ [2026-04-28 19:56:34] INFO 07_residuals: Loading model...
4
+ [2026-04-28 19:56:50] INFO 07_residuals: Model loaded. VRAM: 58292 MB
5
+ [2026-04-28 19:58:15] INFO 07_residuals: Concatenating captures...
6
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 0 cat plan shape (259, 2048)
7
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 0 cat mon shape (691, 2048)
8
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 0 cat exec shape (2287, 2048)
9
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 0 cat all_nl shape (6479, 2048)
10
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 0 cat non_nl shape (3000, 2048)
11
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 2 cat plan shape (259, 2048)
12
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 2 cat mon shape (691, 2048)
13
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 2 cat exec shape (2287, 2048)
14
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 2 cat all_nl shape (6479, 2048)
15
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 2 cat non_nl shape (3000, 2048)
16
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 3 cat plan shape (259, 2048)
17
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 3 cat mon shape (691, 2048)
18
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 3 cat exec shape (2287, 2048)
19
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 3 cat all_nl shape (6479, 2048)
20
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 3 cat non_nl shape (3000, 2048)
21
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 4 cat plan shape (259, 2048)
22
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 4 cat mon shape (691, 2048)
23
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 4 cat exec shape (2287, 2048)
24
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 4 cat all_nl shape (6479, 2048)
25
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 4 cat non_nl shape (3000, 2048)
26
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 6 cat plan shape (259, 2048)
27
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 6 cat mon shape (691, 2048)
28
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 6 cat exec shape (2287, 2048)
29
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 6 cat all_nl shape (6479, 2048)
30
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 6 cat non_nl shape (3000, 2048)
31
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 7 cat plan shape (259, 2048)
32
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 7 cat mon shape (691, 2048)
33
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 7 cat exec shape (2287, 2048)
34
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 7 cat all_nl shape (6479, 2048)
35
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 7 cat non_nl shape (3000, 2048)
36
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 8 cat plan shape (259, 2048)
37
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 8 cat mon shape (691, 2048)
38
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 8 cat exec shape (2287, 2048)
39
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 8 cat all_nl shape (6479, 2048)
40
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 8 cat non_nl shape (3000, 2048)
41
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 9 cat plan shape (259, 2048)
42
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 9 cat mon shape (691, 2048)
43
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 9 cat exec shape (2287, 2048)
44
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 9 cat all_nl shape (6479, 2048)
45
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 9 cat non_nl shape (3000, 2048)
46
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 10 cat plan shape (259, 2048)
47
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 10 cat mon shape (691, 2048)
48
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 10 cat exec shape (2287, 2048)
49
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 10 cat all_nl shape (6479, 2048)
50
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 10 cat non_nl shape (3000, 2048)
51
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 11 cat plan shape (259, 2048)
52
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 11 cat mon shape (691, 2048)
53
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 11 cat exec shape (2287, 2048)
54
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 11 cat all_nl shape (6479, 2048)
55
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 11 cat non_nl shape (3000, 2048)
56
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 12 cat plan shape (259, 2048)
57
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 12 cat mon shape (691, 2048)
58
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 12 cat exec shape (2287, 2048)
59
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 12 cat all_nl shape (6479, 2048)
60
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 12 cat non_nl shape (3000, 2048)
61
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 13 cat plan shape (259, 2048)
62
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 13 cat mon shape (691, 2048)
63
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 13 cat exec shape (2287, 2048)
64
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 13 cat all_nl shape (6479, 2048)
65
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 13 cat non_nl shape (3000, 2048)
66
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 14 cat plan shape (259, 2048)
67
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 14 cat mon shape (691, 2048)
68
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 14 cat exec shape (2287, 2048)
69
+ [2026-04-28 19:58:15] INFO 07_residuals: layer 14 cat all_nl shape (6479, 2048)
70
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 14 cat non_nl shape (3000, 2048)
71
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 15 cat plan shape (259, 2048)
72
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 15 cat mon shape (691, 2048)
73
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 15 cat exec shape (2287, 2048)
74
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 15 cat all_nl shape (6479, 2048)
75
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 15 cat non_nl shape (3000, 2048)
76
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 17 cat plan shape (259, 2048)
77
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 17 cat mon shape (691, 2048)
78
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 17 cat exec shape (2287, 2048)
79
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 17 cat all_nl shape (6479, 2048)
80
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 17 cat non_nl shape (3000, 2048)
81
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 18 cat plan shape (259, 2048)
82
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 18 cat mon shape (691, 2048)
83
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 18 cat exec shape (2287, 2048)
84
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 18 cat all_nl shape (6479, 2048)
85
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 18 cat non_nl shape (3000, 2048)
86
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 20 cat plan shape (259, 2048)
87
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 20 cat mon shape (691, 2048)
88
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 20 cat exec shape (2287, 2048)
89
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 20 cat all_nl shape (6479, 2048)
90
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 20 cat non_nl shape (3000, 2048)
91
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 21 cat plan shape (259, 2048)
92
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 21 cat mon shape (691, 2048)
93
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 21 cat exec shape (2287, 2048)
94
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 21 cat all_nl shape (6479, 2048)
95
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 21 cat non_nl shape (3000, 2048)
96
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 22 cat plan shape (259, 2048)
97
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 22 cat mon shape (691, 2048)
98
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 22 cat exec shape (2287, 2048)
99
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 22 cat all_nl shape (6479, 2048)
100
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 22 cat non_nl shape (3000, 2048)
101
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 23 cat plan shape (259, 2048)
102
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 23 cat mon shape (691, 2048)
103
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 23 cat exec shape (2287, 2048)
104
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 23 cat all_nl shape (6479, 2048)
105
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 23 cat non_nl shape (3000, 2048)
106
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 24 cat plan shape (259, 2048)
107
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 24 cat mon shape (691, 2048)
108
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 24 cat exec shape (2287, 2048)
109
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 24 cat all_nl shape (6479, 2048)
110
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 24 cat non_nl shape (3000, 2048)
111
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 25 cat plan shape (259, 2048)
112
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 25 cat mon shape (691, 2048)
113
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 25 cat exec shape (2287, 2048)
114
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 25 cat all_nl shape (6479, 2048)
115
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 25 cat non_nl shape (3000, 2048)
116
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 26 cat plan shape (259, 2048)
117
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 26 cat mon shape (691, 2048)
118
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 26 cat exec shape (2287, 2048)
119
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 26 cat all_nl shape (6479, 2048)
120
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 26 cat non_nl shape (3000, 2048)
121
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 27 cat plan shape (259, 2048)
122
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 27 cat mon shape (691, 2048)
123
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 27 cat exec shape (2287, 2048)
124
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 27 cat all_nl shape (6479, 2048)
125
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 27 cat non_nl shape (3000, 2048)
126
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 28 cat plan shape (259, 2048)
127
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 28 cat mon shape (691, 2048)
128
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 28 cat exec shape (2287, 2048)
129
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 28 cat all_nl shape (6479, 2048)
130
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 28 cat non_nl shape (3000, 2048)
131
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 29 cat plan shape (259, 2048)
132
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 29 cat mon shape (691, 2048)
133
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 29 cat exec shape (2287, 2048)
134
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 29 cat all_nl shape (6479, 2048)
135
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 29 cat non_nl shape (3000, 2048)
136
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 30 cat plan shape (259, 2048)
137
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 30 cat mon shape (691, 2048)
138
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 30 cat exec shape (2287, 2048)
139
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 30 cat all_nl shape (6479, 2048)
140
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 30 cat non_nl shape (3000, 2048)
141
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 31 cat plan shape (259, 2048)
142
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 31 cat mon shape (691, 2048)
143
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 31 cat exec shape (2287, 2048)
144
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 31 cat all_nl shape (6479, 2048)
145
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 31 cat non_nl shape (3000, 2048)
146
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 32 cat plan shape (259, 2048)
147
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 32 cat mon shape (691, 2048)
148
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 32 cat exec shape (2287, 2048)
149
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 32 cat all_nl shape (6479, 2048)
150
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 32 cat non_nl shape (3000, 2048)
151
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 33 cat plan shape (259, 2048)
152
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 33 cat mon shape (691, 2048)
153
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 33 cat exec shape (2287, 2048)
154
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 33 cat all_nl shape (6479, 2048)
155
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 33 cat non_nl shape (3000, 2048)
156
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 34 cat plan shape (259, 2048)
157
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 34 cat mon shape (691, 2048)
158
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 34 cat exec shape (2287, 2048)
159
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 34 cat all_nl shape (6479, 2048)
160
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 34 cat non_nl shape (3000, 2048)
161
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 35 cat plan shape (259, 2048)
162
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 35 cat mon shape (691, 2048)
163
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 35 cat exec shape (2287, 2048)
164
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 35 cat all_nl shape (6479, 2048)
165
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 35 cat non_nl shape (3000, 2048)
166
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 36 cat plan shape (259, 2048)
167
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 36 cat mon shape (691, 2048)
168
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 36 cat exec shape (2287, 2048)
169
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 36 cat all_nl shape (6479, 2048)
170
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 36 cat non_nl shape (3000, 2048)
171
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 37 cat plan shape (259, 2048)
172
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 37 cat mon shape (691, 2048)
173
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 37 cat exec shape (2287, 2048)
174
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 37 cat all_nl shape (6479, 2048)
175
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 37 cat non_nl shape (3000, 2048)
176
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 38 cat plan shape (259, 2048)
177
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 38 cat mon shape (691, 2048)
178
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 38 cat exec shape (2287, 2048)
179
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 38 cat all_nl shape (6479, 2048)
180
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 38 cat non_nl shape (3000, 2048)
181
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 39 cat plan shape (259, 2048)
182
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 39 cat mon shape (691, 2048)
183
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 39 cat exec shape (2287, 2048)
184
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 39 cat all_nl shape (6479, 2048)
185
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 39 cat non_nl shape (3000, 2048)
186
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 42 cat plan shape (259, 2048)
187
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 42 cat mon shape (691, 2048)
188
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 42 cat exec shape (2287, 2048)
189
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 42 cat all_nl shape (6479, 2048)
190
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 42 cat non_nl shape (3000, 2048)
191
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 45 cat plan shape (259, 2048)
192
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 45 cat mon shape (691, 2048)
193
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 45 cat exec shape (2287, 2048)
194
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 45 cat all_nl shape (6479, 2048)
195
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 45 cat non_nl shape (3000, 2048)
196
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 47 cat plan shape (259, 2048)
197
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 47 cat mon shape (691, 2048)
198
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 47 cat exec shape (2287, 2048)
199
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 47 cat all_nl shape (6479, 2048)
200
+ [2026-04-28 19:58:16] INFO 07_residuals: layer 47 cat non_nl shape (3000, 2048)
201
+ [2026-04-28 19:58:17] INFO 07_residuals: Saved plan/mon/exec residuals: /data/home/user-test/jrhu/studentstimulation/data/activations/decision_point_residuals.pt
202
+ [2026-04-28 19:58:18] INFO 07_residuals: Saved general residuals: /data/home/user-test/jrhu/studentstimulation/data/activations/general_residuals.pt
203
+ [2026-04-29 09:47:02] INFO 07_residuals: Residuals already saved. Skipping.
logs/08_directions.log ADDED
@@ -0,0 +1,272 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [2026-04-28 19:58:20] INFO 08_directions: Loading /data/home/user-test/jrhu/studentstimulation/data/activations/decision_point_residuals.pt
2
+ [2026-04-28 19:58:20] INFO 08_directions: Loading /data/home/user-test/jrhu/studentstimulation/data/activations/general_residuals.pt
3
+ [2026-04-28 19:58:21] INFO 08_directions: Target layers: [0, 2, 3, 4, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 17, 18, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 42, 45, 47]
4
+ [2026-04-28 19:58:21] INFO 08_directions: ============================================================
5
+ [2026-04-28 19:58:21] INFO 08_directions: Step 1: Raw mean-diff (v1)
6
+ [2026-04-28 19:58:21] INFO 08_directions: L 0: ||w_plan||=0.55, ||w_mon||=1.02
7
+ [2026-04-28 19:58:21] INFO 08_directions: L 2: ||w_plan||=1.22, ||w_mon||=1.51
8
+ [2026-04-28 19:58:21] INFO 08_directions: L 3: ||w_plan||=1.89, ||w_mon||=1.99
9
+ [2026-04-28 19:58:21] INFO 08_directions: L 4: ||w_plan||=2.06, ||w_mon||=2.24
10
+ [2026-04-28 19:58:21] INFO 08_directions: L 6: ||w_plan||=2.62, ||w_mon||=2.72
11
+ [2026-04-28 19:58:21] INFO 08_directions: L 7: ||w_plan||=2.75, ||w_mon||=2.76
12
+ [2026-04-28 19:58:21] INFO 08_directions: L 8: ||w_plan||=3.45, ||w_mon||=3.65
13
+ [2026-04-28 19:58:21] INFO 08_directions: L 9: ||w_plan||=3.67, ||w_mon||=3.84
14
+ [2026-04-28 19:58:21] INFO 08_directions: L10: ||w_plan||=4.13, ||w_mon||=4.06
15
+ [2026-04-28 19:58:21] INFO 08_directions: L11: ||w_plan||=4.35, ||w_mon||=4.20
16
+ [2026-04-28 19:58:21] INFO 08_directions: L12: ||w_plan||=4.57, ||w_mon||=4.36
17
+ [2026-04-28 19:58:21] INFO 08_directions: L13: ||w_plan||=4.97, ||w_mon||=4.53
18
+ [2026-04-28 19:58:21] INFO 08_directions: L14: ||w_plan||=5.67, ||w_mon||=5.00
19
+ [2026-04-28 19:58:21] INFO 08_directions: L15: ||w_plan||=6.29, ||w_mon||=5.31
20
+ [2026-04-28 19:58:21] INFO 08_directions: L17: ||w_plan||=6.66, ||w_mon||=5.67
21
+ [2026-04-28 19:58:21] INFO 08_directions: L18: ||w_plan||=6.59, ||w_mon||=5.90
22
+ [2026-04-28 19:58:21] INFO 08_directions: L20: ||w_plan||=7.18, ||w_mon||=6.49
23
+ [2026-04-28 19:58:21] INFO 08_directions: L21: ||w_plan||=7.14, ||w_mon||=6.45
24
+ [2026-04-28 19:58:21] INFO 08_directions: L22: ||w_plan||=7.41, ||w_mon||=6.65
25
+ [2026-04-28 19:58:21] INFO 08_directions: L23: ||w_plan||=7.63, ||w_mon||=6.82
26
+ [2026-04-28 19:58:21] INFO 08_directions: L24: ||w_plan||=7.81, ||w_mon||=7.00
27
+ [2026-04-28 19:58:21] INFO 08_directions: L25: ||w_plan||=8.62, ||w_mon||=7.21
28
+ [2026-04-28 19:58:21] INFO 08_directions: L26: ||w_plan||=9.20, ||w_mon||=7.55
29
+ [2026-04-28 19:58:21] INFO 08_directions: L27: ||w_plan||=9.68, ||w_mon||=7.85
30
+ [2026-04-28 19:58:21] INFO 08_directions: L28: ||w_plan||=10.03, ||w_mon||=8.21
31
+ [2026-04-28 19:58:21] INFO 08_directions: L29: ||w_plan||=10.10, ||w_mon||=8.40
32
+ [2026-04-28 19:58:21] INFO 08_directions: L30: ||w_plan||=10.23, ||w_mon||=8.59
33
+ [2026-04-28 19:58:21] INFO 08_directions: L31: ||w_plan||=10.57, ||w_mon||=8.79
34
+ [2026-04-28 19:58:21] INFO 08_directions: L32: ||w_plan||=11.77, ||w_mon||=10.51
35
+ [2026-04-28 19:58:21] INFO 08_directions: L33: ||w_plan||=11.99, ||w_mon||=10.78
36
+ [2026-04-28 19:58:21] INFO 08_directions: L34: ||w_plan||=12.52, ||w_mon||=11.34
37
+ [2026-04-28 19:58:21] INFO 08_directions: L35: ||w_plan||=13.34, ||w_mon||=11.87
38
+ [2026-04-28 19:58:21] INFO 08_directions: L36: ||w_plan||=14.05, ||w_mon||=13.03
39
+ [2026-04-28 19:58:21] INFO 08_directions: L37: ||w_plan||=16.09, ||w_mon||=14.79
40
+ [2026-04-28 19:58:21] INFO 08_directions: L38: ||w_plan||=18.38, ||w_mon||=18.03
41
+ [2026-04-28 19:58:21] INFO 08_directions: L39: ||w_plan||=20.47, ||w_mon||=20.59
42
+ [2026-04-28 19:58:21] INFO 08_directions: L42: ||w_plan||=28.06, ||w_mon||=35.59
43
+ [2026-04-28 19:58:21] INFO 08_directions: L45: ||w_plan||=36.41, ||w_mon||=45.07
44
+ [2026-04-28 19:58:21] INFO 08_directions: L47: ||w_plan||=48.77, ||w_mon||=50.60
45
+ [2026-04-28 19:58:21] INFO 08_directions: ============================================================
46
+ [2026-04-28 19:58:21] INFO 08_directions: Step 2: General decision-point direction
47
+ [2026-04-28 19:58:22] INFO 08_directions: L 0: ||w_general||=0.50
48
+ [2026-04-28 19:58:22] INFO 08_directions: L 2: ||w_general||=0.98
49
+ [2026-04-28 19:58:22] INFO 08_directions: L 3: ||w_general||=1.46
50
+ [2026-04-28 19:58:22] INFO 08_directions: L 4: ||w_general||=1.61
51
+ [2026-04-28 19:58:22] INFO 08_directions: L 6: ||w_general||=1.90
52
+ [2026-04-28 19:58:22] INFO 08_directions: L 7: ||w_general||=2.00
53
+ [2026-04-28 19:58:22] INFO 08_directions: L 8: ||w_general||=2.30
54
+ [2026-04-28 19:58:22] INFO 08_directions: L 9: ||w_general||=2.46
55
+ [2026-04-28 19:58:22] INFO 08_directions: L10: ||w_general||=2.60
56
+ [2026-04-28 19:58:22] INFO 08_directions: L11: ||w_general||=2.75
57
+ [2026-04-28 19:58:22] INFO 08_directions: L12: ||w_general||=2.87
58
+ [2026-04-28 19:58:22] INFO 08_directions: L13: ||w_general||=3.06
59
+ [2026-04-28 19:58:22] INFO 08_directions: L14: ||w_general||=3.33
60
+ [2026-04-28 19:58:22] INFO 08_directions: L15: ||w_general||=3.44
61
+ [2026-04-28 19:58:22] INFO 08_directions: L17: ||w_general||=3.63
62
+ [2026-04-28 19:58:22] INFO 08_directions: L18: ||w_general||=3.75
63
+ [2026-04-28 19:58:22] INFO 08_directions: L20: ||w_general||=4.02
64
+ [2026-04-28 19:58:22] INFO 08_directions: L21: ||w_general||=4.11
65
+ [2026-04-28 19:58:22] INFO 08_directions: L22: ||w_general||=4.27
66
+ [2026-04-28 19:58:22] INFO 08_directions: L23: ||w_general||=4.38
67
+ [2026-04-28 19:58:22] INFO 08_directions: L24: ||w_general||=4.51
68
+ [2026-04-28 19:58:22] INFO 08_directions: L25: ||w_general||=4.80
69
+ [2026-04-28 19:58:22] INFO 08_directions: L26: ||w_general||=5.06
70
+ [2026-04-28 19:58:22] INFO 08_directions: L27: ||w_general||=5.04
71
+ [2026-04-28 19:58:22] INFO 08_directions: L28: ||w_general||=5.22
72
+ [2026-04-28 19:58:22] INFO 08_directions: L29: ||w_general||=5.33
73
+ [2026-04-28 19:58:22] INFO 08_directions: L30: ||w_general||=5.44
74
+ [2026-04-28 19:58:22] INFO 08_directions: L31: ||w_general||=5.56
75
+ [2026-04-28 19:58:22] INFO 08_directions: L32: ||w_general||=6.18
76
+ [2026-04-28 19:58:22] INFO 08_directions: L33: ||w_general||=6.56
77
+ [2026-04-28 19:58:22] INFO 08_directions: L34: ||w_general||=6.77
78
+ [2026-04-28 19:58:22] INFO 08_directions: L35: ||w_general||=7.30
79
+ [2026-04-28 19:58:22] INFO 08_directions: L36: ||w_general||=7.67
80
+ [2026-04-28 19:58:22] INFO 08_directions: L37: ||w_general||=8.69
81
+ [2026-04-28 19:58:22] INFO 08_directions: L38: ||w_general||=9.69
82
+ [2026-04-28 19:58:22] INFO 08_directions: L39: ||w_general||=10.41
83
+ [2026-04-28 19:58:22] INFO 08_directions: L42: ||w_general||=15.81
84
+ [2026-04-28 19:58:22] INFO 08_directions: L45: ||w_general||=17.63
85
+ [2026-04-28 19:58:22] INFO 08_directions: L47: ||w_general||=26.21
86
+ [2026-04-28 19:58:22] INFO 08_directions: ============================================================
87
+ [2026-04-28 19:58:22] INFO 08_directions: Step 3: Ortho-general (v2)
88
+ [2026-04-28 19:58:22] INFO 08_directions: ============================================================
89
+ [2026-04-28 19:58:22] INFO 08_directions: Step 4: Ortho-crossdim (v3) — plan ⊥ mon and vice versa
90
+ [2026-04-28 19:58:22] INFO 08_directions: ============================================================
91
+ [2026-04-28 19:58:22] INFO 08_directions: Step 5: PCA basis + projection (v4)
92
+ [2026-04-28 19:59:09] INFO 08_directions: L 0: PCA basis shape (128, 2048)
93
+ [2026-04-28 19:59:09] INFO 08_directions: L 2: PCA basis shape (128, 2048)
94
+ [2026-04-28 19:59:09] INFO 08_directions: L 3: PCA basis shape (128, 2048)
95
+ [2026-04-28 19:59:09] INFO 08_directions: L 4: PCA basis shape (128, 2048)
96
+ [2026-04-28 19:59:09] INFO 08_directions: L 6: PCA basis shape (128, 2048)
97
+ [2026-04-28 19:59:09] INFO 08_directions: L 7: PCA basis shape (128, 2048)
98
+ [2026-04-28 19:59:09] INFO 08_directions: L 8: PCA basis shape (128, 2048)
99
+ [2026-04-28 19:59:09] INFO 08_directions: L 9: PCA basis shape (128, 2048)
100
+ [2026-04-28 19:59:09] INFO 08_directions: L10: PCA basis shape (128, 2048)
101
+ [2026-04-28 19:59:09] INFO 08_directions: L11: PCA basis shape (128, 2048)
102
+ [2026-04-28 19:59:09] INFO 08_directions: L12: PCA basis shape (128, 2048)
103
+ [2026-04-28 19:59:09] INFO 08_directions: L13: PCA basis shape (128, 2048)
104
+ [2026-04-28 19:59:09] INFO 08_directions: L14: PCA basis shape (128, 2048)
105
+ [2026-04-28 19:59:09] INFO 08_directions: L15: PCA basis shape (128, 2048)
106
+ [2026-04-28 19:59:09] INFO 08_directions: L17: PCA basis shape (128, 2048)
107
+ [2026-04-28 19:59:09] INFO 08_directions: L18: PCA basis shape (128, 2048)
108
+ [2026-04-28 19:59:09] INFO 08_directions: L20: PCA basis shape (128, 2048)
109
+ [2026-04-28 19:59:09] INFO 08_directions: L21: PCA basis shape (128, 2048)
110
+ [2026-04-28 19:59:09] INFO 08_directions: L22: PCA basis shape (128, 2048)
111
+ [2026-04-28 19:59:09] INFO 08_directions: L23: PCA basis shape (128, 2048)
112
+ [2026-04-28 19:59:09] INFO 08_directions: L24: PCA basis shape (128, 2048)
113
+ [2026-04-28 19:59:09] INFO 08_directions: L25: PCA basis shape (128, 2048)
114
+ [2026-04-28 19:59:09] INFO 08_directions: L26: PCA basis shape (128, 2048)
115
+ [2026-04-28 19:59:09] INFO 08_directions: L27: PCA basis shape (128, 2048)
116
+ [2026-04-28 19:59:09] INFO 08_directions: L28: PCA basis shape (128, 2048)
117
+ [2026-04-28 19:59:09] INFO 08_directions: L29: PCA basis shape (128, 2048)
118
+ [2026-04-28 19:59:09] INFO 08_directions: L30: PCA basis shape (128, 2048)
119
+ [2026-04-28 19:59:09] INFO 08_directions: L31: PCA basis shape (128, 2048)
120
+ [2026-04-28 19:59:09] INFO 08_directions: L32: PCA basis shape (128, 2048)
121
+ [2026-04-28 19:59:09] INFO 08_directions: L33: PCA basis shape (128, 2048)
122
+ [2026-04-28 19:59:09] INFO 08_directions: L34: PCA basis shape (128, 2048)
123
+ [2026-04-28 19:59:09] INFO 08_directions: L35: PCA basis shape (128, 2048)
124
+ [2026-04-28 19:59:09] INFO 08_directions: L36: PCA basis shape (128, 2048)
125
+ [2026-04-28 19:59:09] INFO 08_directions: L37: PCA basis shape (128, 2048)
126
+ [2026-04-28 19:59:09] INFO 08_directions: L38: PCA basis shape (128, 2048)
127
+ [2026-04-28 19:59:09] INFO 08_directions: L39: PCA basis shape (128, 2048)
128
+ [2026-04-28 19:59:09] INFO 08_directions: L42: PCA basis shape (128, 2048)
129
+ [2026-04-28 19:59:09] INFO 08_directions: L45: PCA basis shape (128, 2048)
130
+ [2026-04-28 19:59:09] INFO 08_directions: L47: PCA basis shape (128, 2048)
131
+ [2026-04-28 19:59:09] INFO 08_directions: ============================================================
132
+ [2026-04-28 19:59:09] INFO 08_directions: Normalizing and saving all 4 versions × 2 dimensions
133
+ [2026-04-28 19:59:09] INFO 08_directions: All directions saved.
134
+ [2026-04-28 19:59:09] INFO 08_directions: Computing cosine similarity between versions...
135
+ [2026-04-28 19:59:09] INFO 08_directions: Saved direction_cosines.json
136
+ [2026-04-28 19:59:12] INFO 08_directions: Saved /data/home/user-test/jrhu/studentstimulation/data/results/direction_cosine_matrix.png
137
+ [2026-04-29 09:47:04] INFO 08_directions: Loading /data/home/user-test/jrhu/studentstimulation/data/activations/decision_point_residuals.pt
138
+ [2026-04-29 09:47:04] INFO 08_directions: Loading /data/home/user-test/jrhu/studentstimulation/data/activations/general_residuals.pt
139
+ [2026-04-29 09:47:05] INFO 08_directions: Target layers: [0, 2, 3, 4, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 17, 18, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 42, 45, 47]
140
+ [2026-04-29 09:47:05] INFO 08_directions: ============================================================
141
+ [2026-04-29 09:47:05] INFO 08_directions: Step 1: Raw mean-diff (v1)
142
+ [2026-04-29 09:47:05] INFO 08_directions: L 0: ||w_plan||=0.55, ||w_mon||=1.02
143
+ [2026-04-29 09:47:05] INFO 08_directions: L 2: ||w_plan||=1.22, ||w_mon||=1.51
144
+ [2026-04-29 09:47:05] INFO 08_directions: L 3: ||w_plan||=1.89, ||w_mon||=1.99
145
+ [2026-04-29 09:47:05] INFO 08_directions: L 4: ||w_plan||=2.06, ||w_mon||=2.24
146
+ [2026-04-29 09:47:05] INFO 08_directions: L 6: ||w_plan||=2.62, ||w_mon||=2.72
147
+ [2026-04-29 09:47:05] INFO 08_directions: L 7: ||w_plan||=2.75, ||w_mon||=2.76
148
+ [2026-04-29 09:47:05] INFO 08_directions: L 8: ||w_plan||=3.45, ||w_mon||=3.65
149
+ [2026-04-29 09:47:05] INFO 08_directions: L 9: ||w_plan||=3.67, ||w_mon||=3.84
150
+ [2026-04-29 09:47:05] INFO 08_directions: L10: ||w_plan||=4.13, ||w_mon||=4.06
151
+ [2026-04-29 09:47:05] INFO 08_directions: L11: ||w_plan||=4.35, ||w_mon||=4.20
152
+ [2026-04-29 09:47:05] INFO 08_directions: L12: ||w_plan||=4.57, ||w_mon||=4.36
153
+ [2026-04-29 09:47:05] INFO 08_directions: L13: ||w_plan||=4.97, ||w_mon||=4.53
154
+ [2026-04-29 09:47:05] INFO 08_directions: L14: ||w_plan||=5.67, ||w_mon||=5.00
155
+ [2026-04-29 09:47:05] INFO 08_directions: L15: ||w_plan||=6.29, ||w_mon||=5.31
156
+ [2026-04-29 09:47:05] INFO 08_directions: L17: ||w_plan||=6.66, ||w_mon||=5.67
157
+ [2026-04-29 09:47:05] INFO 08_directions: L18: ||w_plan||=6.59, ||w_mon||=5.90
158
+ [2026-04-29 09:47:05] INFO 08_directions: L20: ||w_plan||=7.18, ||w_mon||=6.49
159
+ [2026-04-29 09:47:05] INFO 08_directions: L21: ||w_plan||=7.14, ||w_mon||=6.45
160
+ [2026-04-29 09:47:05] INFO 08_directions: L22: ||w_plan||=7.41, ||w_mon||=6.65
161
+ [2026-04-29 09:47:05] INFO 08_directions: L23: ||w_plan||=7.63, ||w_mon||=6.82
162
+ [2026-04-29 09:47:05] INFO 08_directions: L24: ||w_plan||=7.81, ||w_mon||=7.00
163
+ [2026-04-29 09:47:05] INFO 08_directions: L25: ||w_plan||=8.62, ||w_mon||=7.21
164
+ [2026-04-29 09:47:05] INFO 08_directions: L26: ||w_plan||=9.20, ||w_mon||=7.55
165
+ [2026-04-29 09:47:05] INFO 08_directions: L27: ||w_plan||=9.68, ||w_mon||=7.85
166
+ [2026-04-29 09:47:05] INFO 08_directions: L28: ||w_plan||=10.03, ||w_mon||=8.21
167
+ [2026-04-29 09:47:05] INFO 08_directions: L29: ||w_plan||=10.10, ||w_mon||=8.40
168
+ [2026-04-29 09:47:05] INFO 08_directions: L30: ||w_plan||=10.23, ||w_mon||=8.59
169
+ [2026-04-29 09:47:05] INFO 08_directions: L31: ||w_plan||=10.57, ||w_mon||=8.79
170
+ [2026-04-29 09:47:05] INFO 08_directions: L32: ||w_plan||=11.77, ||w_mon||=10.51
171
+ [2026-04-29 09:47:05] INFO 08_directions: L33: ||w_plan||=11.99, ||w_mon||=10.78
172
+ [2026-04-29 09:47:05] INFO 08_directions: L34: ||w_plan||=12.52, ||w_mon||=11.34
173
+ [2026-04-29 09:47:05] INFO 08_directions: L35: ||w_plan||=13.34, ||w_mon||=11.87
174
+ [2026-04-29 09:47:05] INFO 08_directions: L36: ||w_plan||=14.05, ||w_mon||=13.03
175
+ [2026-04-29 09:47:05] INFO 08_directions: L37: ||w_plan||=16.09, ||w_mon||=14.79
176
+ [2026-04-29 09:47:05] INFO 08_directions: L38: ||w_plan||=18.38, ||w_mon||=18.03
177
+ [2026-04-29 09:47:05] INFO 08_directions: L39: ||w_plan||=20.47, ||w_mon||=20.59
178
+ [2026-04-29 09:47:05] INFO 08_directions: L42: ||w_plan||=28.06, ||w_mon||=35.59
179
+ [2026-04-29 09:47:05] INFO 08_directions: L45: ||w_plan||=36.41, ||w_mon||=45.07
180
+ [2026-04-29 09:47:05] INFO 08_directions: L47: ||w_plan||=48.77, ||w_mon||=50.60
181
+ [2026-04-29 09:47:05] INFO 08_directions: ============================================================
182
+ [2026-04-29 09:47:05] INFO 08_directions: Step 2: General decision-point direction
183
+ [2026-04-29 09:47:05] INFO 08_directions: L 0: ||w_general||=0.50
184
+ [2026-04-29 09:47:05] INFO 08_directions: L 2: ||w_general||=0.98
185
+ [2026-04-29 09:47:05] INFO 08_directions: L 3: ||w_general||=1.46
186
+ [2026-04-29 09:47:05] INFO 08_directions: L 4: ||w_general||=1.61
187
+ [2026-04-29 09:47:05] INFO 08_directions: L 6: ||w_general||=1.90
188
+ [2026-04-29 09:47:05] INFO 08_directions: L 7: ||w_general||=2.00
189
+ [2026-04-29 09:47:05] INFO 08_directions: L 8: ||w_general||=2.30
190
+ [2026-04-29 09:47:05] INFO 08_directions: L 9: ||w_general||=2.46
191
+ [2026-04-29 09:47:05] INFO 08_directions: L10: ||w_general||=2.60
192
+ [2026-04-29 09:47:05] INFO 08_directions: L11: ||w_general||=2.75
193
+ [2026-04-29 09:47:05] INFO 08_directions: L12: ||w_general||=2.87
194
+ [2026-04-29 09:47:05] INFO 08_directions: L13: ||w_general||=3.06
195
+ [2026-04-29 09:47:05] INFO 08_directions: L14: ||w_general||=3.33
196
+ [2026-04-29 09:47:05] INFO 08_directions: L15: ||w_general||=3.44
197
+ [2026-04-29 09:47:05] INFO 08_directions: L17: ||w_general||=3.63
198
+ [2026-04-29 09:47:05] INFO 08_directions: L18: ||w_general||=3.75
199
+ [2026-04-29 09:47:05] INFO 08_directions: L20: ||w_general||=4.02
200
+ [2026-04-29 09:47:05] INFO 08_directions: L21: ||w_general||=4.11
201
+ [2026-04-29 09:47:05] INFO 08_directions: L22: ||w_general||=4.27
202
+ [2026-04-29 09:47:05] INFO 08_directions: L23: ||w_general||=4.38
203
+ [2026-04-29 09:47:05] INFO 08_directions: L24: ||w_general||=4.51
204
+ [2026-04-29 09:47:05] INFO 08_directions: L25: ||w_general||=4.80
205
+ [2026-04-29 09:47:05] INFO 08_directions: L26: ||w_general||=5.06
206
+ [2026-04-29 09:47:05] INFO 08_directions: L27: ||w_general||=5.04
207
+ [2026-04-29 09:47:05] INFO 08_directions: L28: ||w_general||=5.22
208
+ [2026-04-29 09:47:05] INFO 08_directions: L29: ||w_general||=5.33
209
+ [2026-04-29 09:47:05] INFO 08_directions: L30: ||w_general||=5.44
210
+ [2026-04-29 09:47:05] INFO 08_directions: L31: ||w_general||=5.56
211
+ [2026-04-29 09:47:05] INFO 08_directions: L32: ||w_general||=6.18
212
+ [2026-04-29 09:47:05] INFO 08_directions: L33: ||w_general||=6.56
213
+ [2026-04-29 09:47:05] INFO 08_directions: L34: ||w_general||=6.77
214
+ [2026-04-29 09:47:05] INFO 08_directions: L35: ||w_general||=7.30
215
+ [2026-04-29 09:47:05] INFO 08_directions: L36: ||w_general||=7.67
216
+ [2026-04-29 09:47:05] INFO 08_directions: L37: ||w_general||=8.69
217
+ [2026-04-29 09:47:05] INFO 08_directions: L38: ||w_general||=9.69
218
+ [2026-04-29 09:47:05] INFO 08_directions: L39: ||w_general||=10.41
219
+ [2026-04-29 09:47:05] INFO 08_directions: L42: ||w_general||=15.81
220
+ [2026-04-29 09:47:05] INFO 08_directions: L45: ||w_general||=17.63
221
+ [2026-04-29 09:47:05] INFO 08_directions: L47: ||w_general||=26.21
222
+ [2026-04-29 09:47:05] INFO 08_directions: ============================================================
223
+ [2026-04-29 09:47:05] INFO 08_directions: Step 3: Ortho-general (v2)
224
+ [2026-04-29 09:47:05] INFO 08_directions: ============================================================
225
+ [2026-04-29 09:47:05] INFO 08_directions: Step 4: Ortho-crossdim (v3) — plan ⊥ mon and vice versa
226
+ [2026-04-29 09:47:05] INFO 08_directions: ============================================================
227
+ [2026-04-29 09:47:05] INFO 08_directions: Step 5: PCA basis + projection (v4)
228
+ [2026-04-29 09:47:45] INFO 08_directions: L 0: PCA basis shape (128, 2048)
229
+ [2026-04-29 09:47:45] INFO 08_directions: L 2: PCA basis shape (128, 2048)
230
+ [2026-04-29 09:47:45] INFO 08_directions: L 3: PCA basis shape (128, 2048)
231
+ [2026-04-29 09:47:45] INFO 08_directions: L 4: PCA basis shape (128, 2048)
232
+ [2026-04-29 09:47:45] INFO 08_directions: L 6: PCA basis shape (128, 2048)
233
+ [2026-04-29 09:47:45] INFO 08_directions: L 7: PCA basis shape (128, 2048)
234
+ [2026-04-29 09:47:45] INFO 08_directions: L 8: PCA basis shape (128, 2048)
235
+ [2026-04-29 09:47:45] INFO 08_directions: L 9: PCA basis shape (128, 2048)
236
+ [2026-04-29 09:47:45] INFO 08_directions: L10: PCA basis shape (128, 2048)
237
+ [2026-04-29 09:47:45] INFO 08_directions: L11: PCA basis shape (128, 2048)
238
+ [2026-04-29 09:47:45] INFO 08_directions: L12: PCA basis shape (128, 2048)
239
+ [2026-04-29 09:47:45] INFO 08_directions: L13: PCA basis shape (128, 2048)
240
+ [2026-04-29 09:47:45] INFO 08_directions: L14: PCA basis shape (128, 2048)
241
+ [2026-04-29 09:47:45] INFO 08_directions: L15: PCA basis shape (128, 2048)
242
+ [2026-04-29 09:47:45] INFO 08_directions: L17: PCA basis shape (128, 2048)
243
+ [2026-04-29 09:47:45] INFO 08_directions: L18: PCA basis shape (128, 2048)
244
+ [2026-04-29 09:47:45] INFO 08_directions: L20: PCA basis shape (128, 2048)
245
+ [2026-04-29 09:47:45] INFO 08_directions: L21: PCA basis shape (128, 2048)
246
+ [2026-04-29 09:47:45] INFO 08_directions: L22: PCA basis shape (128, 2048)
247
+ [2026-04-29 09:47:45] INFO 08_directions: L23: PCA basis shape (128, 2048)
248
+ [2026-04-29 09:47:45] INFO 08_directions: L24: PCA basis shape (128, 2048)
249
+ [2026-04-29 09:47:45] INFO 08_directions: L25: PCA basis shape (128, 2048)
250
+ [2026-04-29 09:47:45] INFO 08_directions: L26: PCA basis shape (128, 2048)
251
+ [2026-04-29 09:47:45] INFO 08_directions: L27: PCA basis shape (128, 2048)
252
+ [2026-04-29 09:47:45] INFO 08_directions: L28: PCA basis shape (128, 2048)
253
+ [2026-04-29 09:47:45] INFO 08_directions: L29: PCA basis shape (128, 2048)
254
+ [2026-04-29 09:47:45] INFO 08_directions: L30: PCA basis shape (128, 2048)
255
+ [2026-04-29 09:47:45] INFO 08_directions: L31: PCA basis shape (128, 2048)
256
+ [2026-04-29 09:47:45] INFO 08_directions: L32: PCA basis shape (128, 2048)
257
+ [2026-04-29 09:47:45] INFO 08_directions: L33: PCA basis shape (128, 2048)
258
+ [2026-04-29 09:47:45] INFO 08_directions: L34: PCA basis shape (128, 2048)
259
+ [2026-04-29 09:47:45] INFO 08_directions: L35: PCA basis shape (128, 2048)
260
+ [2026-04-29 09:47:45] INFO 08_directions: L36: PCA basis shape (128, 2048)
261
+ [2026-04-29 09:47:45] INFO 08_directions: L37: PCA basis shape (128, 2048)
262
+ [2026-04-29 09:47:45] INFO 08_directions: L38: PCA basis shape (128, 2048)
263
+ [2026-04-29 09:47:45] INFO 08_directions: L39: PCA basis shape (128, 2048)
264
+ [2026-04-29 09:47:45] INFO 08_directions: L42: PCA basis shape (128, 2048)
265
+ [2026-04-29 09:47:45] INFO 08_directions: L45: PCA basis shape (128, 2048)
266
+ [2026-04-29 09:47:45] INFO 08_directions: L47: PCA basis shape (128, 2048)
267
+ [2026-04-29 09:47:46] INFO 08_directions: ============================================================
268
+ [2026-04-29 09:47:46] INFO 08_directions: Normalizing and saving all 4 versions × 2 dimensions
269
+ [2026-04-29 09:47:46] INFO 08_directions: All directions saved.
270
+ [2026-04-29 09:47:46] INFO 08_directions: Computing cosine similarity between versions...
271
+ [2026-04-29 09:47:46] INFO 08_directions: Saved direction_cosines.json
272
+ [2026-04-29 09:47:49] INFO 08_directions: Saved /data/home/user-test/jrhu/studentstimulation/data/results/direction_cosine_matrix.png
logs/09_sweep.log ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [2026-04-28 19:59:18] INFO 09_sweep: Test problems: 30
2
+ [2026-04-28 19:59:18] INFO 09_sweep: Loading model...
3
+ [2026-04-28 19:59:38] INFO 09_sweep: Loaded planning/v1_raw: 39 layers, 39 non-zero
4
+ [2026-04-28 19:59:38] INFO 09_sweep: Loaded planning/v2_ortho_general: 39 layers, 39 non-zero
5
+ [2026-04-28 19:59:38] INFO 09_sweep: Loaded planning/v3_ortho_crossdim: 39 layers, 39 non-zero
6
+ [2026-04-28 19:59:38] INFO 09_sweep: Loaded planning/v4_pca: 39 layers, 39 non-zero
7
+ [2026-04-28 19:59:38] INFO 09_sweep: Loaded monitoring/v1_raw: 39 layers, 39 non-zero
8
+ [2026-04-28 19:59:38] INFO 09_sweep: Loaded monitoring/v2_ortho_general: 39 layers, 39 non-zero
9
+ [2026-04-28 19:59:38] INFO 09_sweep: Loaded monitoring/v3_ortho_crossdim: 39 layers, 39 non-zero
10
+ [2026-04-28 19:59:38] INFO 09_sweep: Loaded monitoring/v4_pca: 39 layers, 39 non-zero
11
+ [2026-04-28 19:59:38] INFO 09_sweep: Resume: 0 experiments already logged
12
+ [2026-04-28 19:59:38] INFO 09_sweep: Computing baselines (alpha=1, NEW semantics: no steering)...
13
+ [2026-04-28 20:21:04] INFO 09_sweep: Total sweep runs: 1920
14
+ [2026-04-29 09:47:55] INFO 09_sweep: Test problems: 30
15
+ [2026-04-29 09:47:55] INFO 09_sweep: Loading model...
16
+ [2026-04-29 09:48:16] INFO 09_sweep: Loaded planning/v1_raw: 39 layers, 39 non-zero
17
+ [2026-04-29 09:48:16] INFO 09_sweep: Loaded planning/v2_ortho_general: 39 layers, 39 non-zero
18
+ [2026-04-29 09:48:16] INFO 09_sweep: Loaded planning/v3_ortho_crossdim: 39 layers, 39 non-zero
19
+ [2026-04-29 09:48:16] INFO 09_sweep: Loaded planning/v4_pca: 39 layers, 39 non-zero
20
+ [2026-04-29 09:48:16] INFO 09_sweep: Loaded monitoring/v1_raw: 39 layers, 39 non-zero
21
+ [2026-04-29 09:48:16] INFO 09_sweep: Loaded monitoring/v2_ortho_general: 39 layers, 39 non-zero
22
+ [2026-04-29 09:48:16] INFO 09_sweep: Loaded monitoring/v3_ortho_crossdim: 39 layers, 39 non-zero
23
+ [2026-04-29 09:48:16] INFO 09_sweep: Loaded monitoring/v4_pca: 39 layers, 39 non-zero
24
+ [2026-04-29 09:48:16] INFO 09_sweep: Resume: 1062 experiments already logged
25
+ [2026-04-29 09:48:16] INFO 09_sweep: Computing baselines (alpha=1, NEW semantics: no steering)...
26
+ [2026-04-29 10:09:51] INFO 09_sweep: Total sweep runs: 1920
27
+ [2026-04-29 19:37:22] INFO 09_sweep: Force-prompt baselines...
28
+ [2026-04-29 21:00:21] INFO 09_sweep: Sweep complete. Log: /data/home/user-test/jrhu/studentstimulation/data/results/sweep_log.jsonl
logs/12_downstream.log ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [2026-04-29 21:00:25] INFO 12_downstream: Test sets: ['MATH-500-holdout', 'AIME-24']
2
+ [2026-04-29 21:00:25] INFO 12_downstream: MATH-500-holdout: 50 problems
3
+ [2026-04-29 21:00:25] INFO 12_downstream: AIME-24: 30 problems
4
+ [2026-04-29 21:00:25] INFO 12_downstream: Loading model...
5
+ [2026-04-29 21:00:38] INFO 12_downstream: === Config: baseline ===
6
+ [2026-04-29 22:05:06] INFO 12_downstream: MATH-500-holdout: 15/50 = 0.300
7
+ [2026-04-29 22:47:56] INFO 12_downstream: AIME-24: 1/30 = 0.033
8
+ [2026-04-29 22:47:56] INFO 12_downstream: === Config: plan_alpha_0 ===
9
+ [2026-04-30 00:04:14] INFO 12_downstream: MATH-500-holdout: 2/50 = 0.040
10
+ [2026-04-30 00:50:06] INFO 12_downstream: AIME-24: 0/30 = 0.000
11
+ [2026-04-30 00:50:06] INFO 12_downstream: === Config: mon_alpha_0 ===
12
+ [2026-04-30 02:00:53] INFO 12_downstream: MATH-500-holdout: 16/50 = 0.320
13
+ [2026-04-30 02:46:44] INFO 12_downstream: AIME-24: 0/30 = 0.000
14
+ [2026-04-30 02:46:44] INFO 12_downstream: ============================================================
15
+ [2026-04-30 02:46:44] INFO 12_downstream: Computing per-config accuracy drop vs baseline...
16
+ [2026-04-30 02:46:44] INFO 12_downstream: plan_alpha_0/MATH-500-holdout: acc 0.300 -> 0.040 (Δ=+0.260, rel=+86.7%) regressions=13 recoveries=0 McNemar p=0.000244
17
+ [2026-04-30 02:46:44] INFO 12_downstream: plan_alpha_0/AIME-24: acc 0.033 -> 0.000 (Δ=+0.033, rel=+100.0%) regressions=1 recoveries=0 McNemar p=1
18
+ [2026-04-30 02:46:44] INFO 12_downstream: mon_alpha_0/MATH-500-holdout: acc 0.300 -> 0.320 (Δ=-0.020, rel=-6.7%) regressions=4 recoveries=5 McNemar p=1
19
+ [2026-04-30 02:46:44] INFO 12_downstream: mon_alpha_0/AIME-24: acc 0.033 -> 0.000 (Δ=+0.033, rel=+100.0%) regressions=1 recoveries=0 McNemar p=1
20
+ [2026-04-30 02:46:44] INFO 12_downstream: Saved: /data/home/user-test/jrhu/studentstimulation/data/results/downstream_accuracy.json
logs/13_analyze.log ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ [2026-04-30 02:46:45] INFO 13_analyze: Loaded 2040 sweep records
2
+ [2026-04-30 02:46:45] INFO 13_analyze: Saved aggregated sweep: /data/home/user-test/jrhu/studentstimulation/data/results/sweep_rr.json
3
+ [2026-04-30 02:46:46] INFO 13_analyze: Saved curves: /data/home/user-test/jrhu/studentstimulation/data/results/sweep_curves.png
4
+ [2026-04-30 02:46:47] INFO 13_analyze: Saved final report: /data/home/user-test/jrhu/studentstimulation/data/results/final_report.md
logs/runall.log ADDED
The diff for this file is too large to render. See raw diff
 
results/direction_cosines.json ADDED
@@ -0,0 +1,992 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "within_planning": {
3
+ "v1_raw__VS__v1_raw": {
4
+ "0": 1.0000001192092896,
5
+ "2": 0.9999999403953552,
6
+ "3": 1.0,
7
+ "4": 1.0000001192092896,
8
+ "6": 0.9999998807907104,
9
+ "7": 0.9999998807907104,
10
+ "8": 1.0,
11
+ "9": 1.0000001192092896,
12
+ "10": 1.0000001192092896,
13
+ "11": 1.0000001192092896,
14
+ "12": 1.0,
15
+ "13": 1.0000001192092896,
16
+ "14": 1.0,
17
+ "15": 1.0,
18
+ "17": 1.0,
19
+ "18": 1.0,
20
+ "20": 1.0000001192092896,
21
+ "21": 1.0000001192092896,
22
+ "22": 1.0,
23
+ "23": 1.0,
24
+ "24": 1.0000001192092896,
25
+ "25": 1.0,
26
+ "26": 1.0000001192092896,
27
+ "27": 1.0000001192092896,
28
+ "28": 1.000000238418579,
29
+ "29": 1.0000001192092896,
30
+ "30": 1.0,
31
+ "31": 1.0000001192092896,
32
+ "32": 1.0,
33
+ "33": 0.9999999403953552,
34
+ "34": 1.0000001192092896,
35
+ "35": 1.0000001192092896,
36
+ "36": 1.0,
37
+ "37": 1.0000001192092896,
38
+ "38": 1.0,
39
+ "39": 1.0,
40
+ "42": 1.0,
41
+ "45": 1.0,
42
+ "47": 1.000000238418579
43
+ },
44
+ "v1_raw__VS__v2_ortho_general": {
45
+ "0": 0.997159481048584,
46
+ "2": 0.9649484157562256,
47
+ "3": 0.9730725884437561,
48
+ "4": 0.9631573557853699,
49
+ "6": 0.9305754899978638,
50
+ "7": 0.9249453544616699,
51
+ "8": 0.9292974472045898,
52
+ "9": 0.9355981945991516,
53
+ "10": 0.9296989440917969,
54
+ "11": 0.9247703552246094,
55
+ "12": 0.918307363986969,
56
+ "13": 0.9102700352668762,
57
+ "14": 0.8945496678352356,
58
+ "15": 0.892246425151825,
59
+ "17": 0.8907243013381958,
60
+ "18": 0.8900290727615356,
61
+ "20": 0.8870463371276855,
62
+ "21": 0.8981274962425232,
63
+ "22": 0.8881056308746338,
64
+ "23": 0.880789041519165,
65
+ "24": 0.8788645267486572,
66
+ "25": 0.8652130365371704,
67
+ "26": 0.8568968176841736,
68
+ "27": 0.8815737962722778,
69
+ "28": 0.8722836971282959,
70
+ "29": 0.8788827061653137,
71
+ "30": 0.8719693422317505,
72
+ "31": 0.8683976531028748,
73
+ "32": 0.8877090811729431,
74
+ "33": 0.9013210535049438,
75
+ "34": 0.898939847946167,
76
+ "35": 0.9028238654136658,
77
+ "36": 0.9098505973815918,
78
+ "37": 0.9053845405578613,
79
+ "38": 0.8991852402687073,
80
+ "39": 0.9117066264152527,
81
+ "42": 0.9325175285339355,
82
+ "45": 0.9361408352851868,
83
+ "47": 0.8353425860404968
84
+ },
85
+ "v1_raw__VS__v3_ortho_crossdim": {
86
+ "0": 0.9965951442718506,
87
+ "2": 0.9544905424118042,
88
+ "3": 0.9619656205177307,
89
+ "4": 0.9553236961364746,
90
+ "6": 0.9247667789459229,
91
+ "7": 0.9205638766288757,
92
+ "8": 0.9246719479560852,
93
+ "9": 0.9316375255584717,
94
+ "10": 0.9280927777290344,
95
+ "11": 0.9240896105766296,
96
+ "12": 0.9181762933731079,
97
+ "13": 0.9098619818687439,
98
+ "14": 0.8934292793273926,
99
+ "15": 0.8918048143386841,
100
+ "17": 0.890674352645874,
101
+ "18": 0.8899945020675659,
102
+ "20": 0.8867461085319519,
103
+ "21": 0.8965198993682861,
104
+ "22": 0.8870662450790405,
105
+ "23": 0.880109429359436,
106
+ "24": 0.878195583820343,
107
+ "25": 0.8619124889373779,
108
+ "26": 0.8508340716362,
109
+ "27": 0.8784091472625732,
110
+ "28": 0.8678410053253174,
111
+ "29": 0.8744179010391235,
112
+ "30": 0.8696860074996948,
113
+ "31": 0.8663952946662903,
114
+ "32": 0.8835557699203491,
115
+ "33": 0.8973360061645508,
116
+ "34": 0.8957799077033997,
117
+ "35": 0.8999952077865601,
118
+ "36": 0.9078393578529358,
119
+ "37": 0.9007821083068848,
120
+ "38": 0.8952031135559082,
121
+ "39": 0.9077118039131165,
122
+ "42": 0.9248653650283813,
123
+ "45": 0.928388237953186,
124
+ "47": 0.8110612034797668
125
+ },
126
+ "v1_raw__VS__v4_pca": {
127
+ "0": 0.9979102611541748,
128
+ "2": 0.9776266813278198,
129
+ "3": 0.9773249626159668,
130
+ "4": 0.9748995900154114,
131
+ "6": 0.9557616710662842,
132
+ "7": 0.9508732557296753,
133
+ "8": 0.9520366787910461,
134
+ "9": 0.9555205702781677,
135
+ "10": 0.9531899094581604,
136
+ "11": 0.9500094652175903,
137
+ "12": 0.9449301958084106,
138
+ "13": 0.9393899440765381,
139
+ "14": 0.9264705181121826,
140
+ "15": 0.9215720295906067,
141
+ "17": 0.9202317595481873,
142
+ "18": 0.9182098507881165,
143
+ "20": 0.9163202047348022,
144
+ "21": 0.9244973063468933,
145
+ "22": 0.9171287417411804,
146
+ "23": 0.9124293923377991,
147
+ "24": 0.910251259803772,
148
+ "25": 0.899188756942749,
149
+ "26": 0.8906795978546143,
150
+ "27": 0.9127213358879089,
151
+ "28": 0.9036138653755188,
152
+ "29": 0.9093997478485107,
153
+ "30": 0.9047744870185852,
154
+ "31": 0.9016937613487244,
155
+ "32": 0.9178523421287537,
156
+ "33": 0.929762601852417,
157
+ "34": 0.9288420081138611,
158
+ "35": 0.9322060346603394,
159
+ "36": 0.9391618967056274,
160
+ "37": 0.933300793170929,
161
+ "38": 0.9279406070709229,
162
+ "39": 0.9364868998527527,
163
+ "42": 0.9523126482963562,
164
+ "45": 0.9541194438934326,
165
+ "47": 0.8399131298065186
166
+ },
167
+ "v2_ortho_general__VS__v2_ortho_general": {
168
+ "0": 1.0,
169
+ "2": 1.000000238418579,
170
+ "3": 1.0,
171
+ "4": 1.0,
172
+ "6": 0.9999998807907104,
173
+ "7": 0.9999999403953552,
174
+ "8": 1.0,
175
+ "9": 1.0,
176
+ "10": 1.0,
177
+ "11": 1.000000238418579,
178
+ "12": 0.9999999403953552,
179
+ "13": 1.0000001192092896,
180
+ "14": 0.9999998807907104,
181
+ "15": 1.0000001192092896,
182
+ "17": 1.0000001192092896,
183
+ "18": 1.0,
184
+ "20": 1.0000001192092896,
185
+ "21": 1.0,
186
+ "22": 1.0000001192092896,
187
+ "23": 1.0,
188
+ "24": 1.0000001192092896,
189
+ "25": 1.000000238418579,
190
+ "26": 0.9999999403953552,
191
+ "27": 1.0,
192
+ "28": 0.9999999403953552,
193
+ "29": 1.0000001192092896,
194
+ "30": 1.0,
195
+ "31": 0.9999999403953552,
196
+ "32": 1.0,
197
+ "33": 1.0000001192092896,
198
+ "34": 0.9999998807907104,
199
+ "35": 1.0,
200
+ "36": 1.000000238418579,
201
+ "37": 1.0,
202
+ "38": 1.0000001192092896,
203
+ "39": 1.0000001192092896,
204
+ "42": 1.0000001192092896,
205
+ "45": 0.9999999403953552,
206
+ "47": 1.0000001192092896
207
+ },
208
+ "v2_ortho_general__VS__v3_ortho_crossdim": {
209
+ "0": 0.9994341135025024,
210
+ "2": 0.9891624450683594,
211
+ "3": 0.9885856509208679,
212
+ "4": 0.9918666481971741,
213
+ "6": 0.993757963180542,
214
+ "7": 0.9952629804611206,
215
+ "8": 0.9950225949287415,
216
+ "9": 0.995766818523407,
217
+ "10": 0.9982724785804749,
218
+ "11": 0.999264121055603,
219
+ "12": 0.9998571872711182,
220
+ "13": 0.9995517730712891,
221
+ "14": 0.9987474679946899,
222
+ "15": 0.9995051622390747,
223
+ "17": 0.9999439120292664,
224
+ "18": 0.9999611973762512,
225
+ "20": 0.9996616840362549,
226
+ "21": 0.9982101917266846,
227
+ "22": 0.9988297820091248,
228
+ "23": 0.9992284774780273,
229
+ "24": 0.9992389678955078,
230
+ "25": 0.9961855411529541,
231
+ "26": 0.9929247498512268,
232
+ "27": 0.9964101314544678,
233
+ "28": 0.9949068427085876,
234
+ "29": 0.994920015335083,
235
+ "30": 0.9973814487457275,
236
+ "31": 0.9976940751075745,
237
+ "32": 0.9953214526176453,
238
+ "33": 0.9955788850784302,
239
+ "34": 0.9964847564697266,
240
+ "35": 0.9968668818473816,
241
+ "36": 0.9977896809577942,
242
+ "37": 0.9949166774749756,
243
+ "38": 0.995571494102478,
244
+ "39": 0.9956184029579163,
245
+ "42": 0.9917941093444824,
246
+ "45": 0.9917186498641968,
247
+ "47": 0.9709324240684509
248
+ },
249
+ "v2_ortho_general__VS__v4_pca": {
250
+ "0": 0.9976089000701904,
251
+ "2": 0.9723131060600281,
252
+ "3": 0.9769964218139648,
253
+ "4": 0.9763866662979126,
254
+ "6": 0.9670107364654541,
255
+ "7": 0.9676867127418518,
256
+ "8": 0.9707514643669128,
257
+ "9": 0.9743568897247314,
258
+ "10": 0.974492609500885,
259
+ "11": 0.9735289216041565,
260
+ "12": 0.972087562084198,
261
+ "13": 0.969159722328186,
262
+ "14": 0.9649456739425659,
263
+ "15": 0.9656915664672852,
264
+ "17": 0.9672946333885193,
265
+ "18": 0.9687793850898743,
266
+ "20": 0.968298077583313,
267
+ "21": 0.9699345231056213,
268
+ "22": 0.967728853225708,
269
+ "23": 0.965173602104187,
270
+ "24": 0.9653860926628113,
271
+ "25": 0.958077073097229,
272
+ "26": 0.9524403810501099,
273
+ "27": 0.9614723920822144,
274
+ "28": 0.9584618210792542,
275
+ "29": 0.9595796465873718,
276
+ "30": 0.9609693884849548,
277
+ "31": 0.960560142993927,
278
+ "32": 0.9612539410591125,
279
+ "33": 0.9637250900268555,
280
+ "34": 0.9636179208755493,
281
+ "35": 0.9646608233451843,
282
+ "36": 0.9664263725280762,
283
+ "37": 0.963169515132904,
284
+ "38": 0.9634197950363159,
285
+ "39": 0.9677342176437378,
286
+ "42": 0.9684389233589172,
287
+ "45": 0.9696422815322876,
288
+ "47": 0.9374622702598572
289
+ },
290
+ "v3_ortho_crossdim__VS__v3_ortho_crossdim": {
291
+ "0": 0.9999998807907104,
292
+ "2": 0.9999999403953552,
293
+ "3": 1.0,
294
+ "4": 1.0000001192092896,
295
+ "6": 0.9999998807907104,
296
+ "7": 1.000000238418579,
297
+ "8": 0.9999998807907104,
298
+ "9": 1.0,
299
+ "10": 1.0,
300
+ "11": 1.0,
301
+ "12": 1.0000001192092896,
302
+ "13": 1.000000238418579,
303
+ "14": 1.0,
304
+ "15": 0.9999998211860657,
305
+ "17": 1.000000238418579,
306
+ "18": 1.0000001192092896,
307
+ "20": 0.9999999403953552,
308
+ "21": 1.0000001192092896,
309
+ "22": 1.000000238418579,
310
+ "23": 1.0000001192092896,
311
+ "24": 1.0000001192092896,
312
+ "25": 1.000000238418579,
313
+ "26": 1.0,
314
+ "27": 1.0,
315
+ "28": 1.000000238418579,
316
+ "29": 0.9999999403953552,
317
+ "30": 1.0000001192092896,
318
+ "31": 1.0,
319
+ "32": 0.9999997615814209,
320
+ "33": 1.0,
321
+ "34": 1.0,
322
+ "35": 0.9999998807907104,
323
+ "36": 1.0000001192092896,
324
+ "37": 1.0000001192092896,
325
+ "38": 0.9999998807907104,
326
+ "39": 0.9999999403953552,
327
+ "42": 1.0000001192092896,
328
+ "45": 0.9999998807907104,
329
+ "47": 1.0000001192092896
330
+ },
331
+ "v3_ortho_crossdim__VS__v4_pca": {
332
+ "0": 0.997680127620697,
333
+ "2": 0.9758776426315308,
334
+ "3": 0.9839513897895813,
335
+ "4": 0.9794027805328369,
336
+ "6": 0.967278242111206,
337
+ "7": 0.9676596522331238,
338
+ "8": 0.9711106419563293,
339
+ "9": 0.9747692346572876,
340
+ "10": 0.9735124707221985,
341
+ "11": 0.9724835753440857,
342
+ "12": 0.9713910818099976,
343
+ "13": 0.9682139158248901,
344
+ "14": 0.9639472365379333,
345
+ "15": 0.9674946665763855,
346
+ "17": 0.9677612781524658,
347
+ "18": 0.96918123960495,
348
+ "20": 0.9675778746604919,
349
+ "21": 0.9694366455078125,
350
+ "22": 0.966861367225647,
351
+ "23": 0.9642056226730347,
352
+ "24": 0.9644226431846619,
353
+ "25": 0.9582943320274353,
354
+ "26": 0.9550814032554626,
355
+ "27": 0.9621031880378723,
356
+ "28": 0.9600833058357239,
357
+ "29": 0.9612242579460144,
358
+ "30": 0.9609367251396179,
359
+ "31": 0.960490345954895,
360
+ "32": 0.9622399210929871,
361
+ "33": 0.9645735025405884,
362
+ "34": 0.9638328552246094,
363
+ "35": 0.9648866057395935,
364
+ "36": 0.9660252928733826,
365
+ "37": 0.9645795822143555,
366
+ "38": 0.9639973044395447,
367
+ "39": 0.9686235785484314,
368
+ "42": 0.9700583815574646,
369
+ "45": 0.9717600345611572,
370
+ "47": 0.9643595218658447
371
+ },
372
+ "v4_pca__VS__v4_pca": {
373
+ "0": 1.0,
374
+ "2": 1.0000001192092896,
375
+ "3": 1.000000238418579,
376
+ "4": 1.0,
377
+ "6": 0.9999998807907104,
378
+ "7": 1.0000001192092896,
379
+ "8": 1.0000001192092896,
380
+ "9": 1.0,
381
+ "10": 1.0,
382
+ "11": 1.0000001192092896,
383
+ "12": 1.0,
384
+ "13": 1.0000001192092896,
385
+ "14": 1.0000001192092896,
386
+ "15": 1.0,
387
+ "17": 1.0,
388
+ "18": 1.0,
389
+ "20": 1.0,
390
+ "21": 1.0,
391
+ "22": 1.0000001192092896,
392
+ "23": 1.0,
393
+ "24": 0.9999997615814209,
394
+ "25": 1.000000238418579,
395
+ "26": 1.000000238418579,
396
+ "27": 1.0,
397
+ "28": 0.9999998807907104,
398
+ "29": 1.0000001192092896,
399
+ "30": 1.0000001192092896,
400
+ "31": 1.000000238418579,
401
+ "32": 1.0,
402
+ "33": 0.9999998807907104,
403
+ "34": 1.0,
404
+ "35": 1.0,
405
+ "36": 1.0,
406
+ "37": 0.9999999403953552,
407
+ "38": 0.9999998807907104,
408
+ "39": 1.0,
409
+ "42": 1.0000001192092896,
410
+ "45": 1.0000001192092896,
411
+ "47": 1.000000238418579
412
+ }
413
+ },
414
+ "within_monitoring": {
415
+ "v1_raw__VS__v1_raw": {
416
+ "0": 1.0000001192092896,
417
+ "2": 1.0,
418
+ "3": 1.0000001192092896,
419
+ "4": 1.000000238418579,
420
+ "6": 0.9999998807907104,
421
+ "7": 1.0,
422
+ "8": 1.0,
423
+ "9": 0.9999998807907104,
424
+ "10": 1.0,
425
+ "11": 1.0,
426
+ "12": 0.9999999403953552,
427
+ "13": 1.000000238418579,
428
+ "14": 0.9999999403953552,
429
+ "15": 1.0,
430
+ "17": 1.000000238418579,
431
+ "18": 1.0000001192092896,
432
+ "20": 0.9999999403953552,
433
+ "21": 0.9999998807907104,
434
+ "22": 1.0000001192092896,
435
+ "23": 1.0000001192092896,
436
+ "24": 1.0,
437
+ "25": 1.000000238418579,
438
+ "26": 1.0000001192092896,
439
+ "27": 1.0,
440
+ "28": 1.000000238418579,
441
+ "29": 1.0,
442
+ "30": 1.0,
443
+ "31": 1.0,
444
+ "32": 1.0000001192092896,
445
+ "33": 1.000000238418579,
446
+ "34": 0.9999997615814209,
447
+ "35": 1.000000238418579,
448
+ "36": 1.0000001192092896,
449
+ "37": 0.9999999403953552,
450
+ "38": 1.0000003576278687,
451
+ "39": 1.0000001192092896,
452
+ "42": 0.9999998807907104,
453
+ "45": 1.0,
454
+ "47": 1.0000001192092896
455
+ },
456
+ "v1_raw__VS__v2_ortho_general": {
457
+ "0": 0.9438518285751343,
458
+ "2": 0.9531650543212891,
459
+ "3": 0.9712194204330444,
460
+ "4": 0.9570870399475098,
461
+ "6": 0.9471193552017212,
462
+ "7": 0.9517127871513367,
463
+ "8": 0.9490780830383301,
464
+ "9": 0.9511783719062805,
465
+ "10": 0.9431129097938538,
466
+ "11": 0.943788468837738,
467
+ "12": 0.9367179274559021,
468
+ "13": 0.9488347768783569,
469
+ "14": 0.9549822807312012,
470
+ "15": 0.9533748030662537,
471
+ "17": 0.9573957920074463,
472
+ "18": 0.9474153518676758,
473
+ "20": 0.9513649940490723,
474
+ "21": 0.9559173583984375,
475
+ "22": 0.9498403668403625,
476
+ "23": 0.9493352770805359,
477
+ "24": 0.9480846524238586,
478
+ "25": 0.9625384211540222,
479
+ "26": 0.9723770618438721,
480
+ "27": 0.9719191193580627,
481
+ "28": 0.973452627658844,
482
+ "29": 0.974128782749176,
483
+ "30": 0.9683023691177368,
484
+ "31": 0.9707658290863037,
485
+ "32": 0.968063235282898,
486
+ "33": 0.9684690833091736,
487
+ "34": 0.9666540026664734,
488
+ "35": 0.9706434607505798,
489
+ "36": 0.9668235778808594,
490
+ "37": 0.9724254012107849,
491
+ "38": 0.9634373784065247,
492
+ "39": 0.9649926424026489,
493
+ "42": 0.9400622844696045,
494
+ "45": 0.9442551136016846,
495
+ "47": 0.9990965127944946
496
+ },
497
+ "v1_raw__VS__v3_ortho_crossdim": {
498
+ "0": 0.9433177709579468,
499
+ "2": 0.9428349137306213,
500
+ "3": 0.9601336717605591,
501
+ "4": 0.9493026733398438,
502
+ "6": 0.9412074685096741,
503
+ "7": 0.9472044110298157,
504
+ "8": 0.9443544149398804,
505
+ "9": 0.9471518397331238,
506
+ "10": 0.9414835572242737,
507
+ "11": 0.9430938959121704,
508
+ "12": 0.9365841746330261,
509
+ "13": 0.9484094977378845,
510
+ "14": 0.9537861943244934,
511
+ "15": 0.9529029130935669,
512
+ "17": 0.9573419094085693,
513
+ "18": 0.9473787546157837,
514
+ "20": 0.9510432481765747,
515
+ "21": 0.9542065262794495,
516
+ "22": 0.9487288594245911,
517
+ "23": 0.9486027359962463,
518
+ "24": 0.9473629593849182,
519
+ "25": 0.9588667750358582,
520
+ "26": 0.9654972553253174,
521
+ "27": 0.9684301018714905,
522
+ "28": 0.968494713306427,
523
+ "29": 0.9691802263259888,
524
+ "30": 0.9657666683197021,
525
+ "31": 0.9685274362564087,
526
+ "32": 0.9635341763496399,
527
+ "33": 0.9641873836517334,
528
+ "34": 0.9632558822631836,
529
+ "35": 0.9676024317741394,
530
+ "36": 0.9646866321563721,
531
+ "37": 0.9674820899963379,
532
+ "38": 0.9591707587242126,
533
+ "39": 0.9607644081115723,
534
+ "42": 0.9323481917381287,
535
+ "45": 0.936435341835022,
536
+ "47": 0.9700551629066467
537
+ },
538
+ "v1_raw__VS__v4_pca": {
539
+ "0": 0.9718185663223267,
540
+ "2": 0.9720360040664673,
541
+ "3": 0.9763880968093872,
542
+ "4": 0.9716101884841919,
543
+ "6": 0.9658738970756531,
544
+ "7": 0.9678991436958313,
545
+ "8": 0.9650120735168457,
546
+ "9": 0.965965986251831,
547
+ "10": 0.9620813131332397,
548
+ "11": 0.9625723958015442,
549
+ "12": 0.9571533203125,
550
+ "13": 0.9651802182197571,
551
+ "14": 0.9683049917221069,
552
+ "15": 0.9642326831817627,
553
+ "17": 0.9678332805633545,
554
+ "18": 0.9599108695983887,
555
+ "20": 0.9635767340660095,
556
+ "21": 0.9669395685195923,
557
+ "22": 0.9625208973884583,
558
+ "23": 0.9624016880989075,
559
+ "24": 0.9611483812332153,
560
+ "25": 0.9708266258239746,
561
+ "26": 0.9759269952774048,
562
+ "27": 0.9780619740486145,
563
+ "28": 0.9779258966445923,
564
+ "29": 0.9786235690116882,
565
+ "30": 0.9755846261978149,
566
+ "31": 0.9773844480514526,
567
+ "32": 0.9751286506652832,
568
+ "33": 0.9762512445449829,
569
+ "34": 0.9755951762199402,
570
+ "35": 0.9786450862884521,
571
+ "36": 0.9771538972854614,
572
+ "37": 0.9788910150527954,
573
+ "38": 0.972733736038208,
574
+ "39": 0.9736830592155457,
575
+ "42": 0.9576069116592407,
576
+ "45": 0.959835410118103,
577
+ "47": 0.9728759527206421
578
+ },
579
+ "v2_ortho_general__VS__v2_ortho_general": {
580
+ "0": 1.0,
581
+ "2": 1.0000001192092896,
582
+ "3": 1.0000001192092896,
583
+ "4": 1.0,
584
+ "6": 1.000000238418579,
585
+ "7": 1.0,
586
+ "8": 0.9999996423721313,
587
+ "9": 1.0000001192092896,
588
+ "10": 1.0000001192092896,
589
+ "11": 1.0000001192092896,
590
+ "12": 1.0000001192092896,
591
+ "13": 1.0,
592
+ "14": 1.0000001192092896,
593
+ "15": 1.0000001192092896,
594
+ "17": 1.000000238418579,
595
+ "18": 0.9999999403953552,
596
+ "20": 0.9999998807907104,
597
+ "21": 0.9999997615814209,
598
+ "22": 0.9999999403953552,
599
+ "23": 1.000000238418579,
600
+ "24": 1.0000001192092896,
601
+ "25": 1.0,
602
+ "26": 1.0,
603
+ "27": 0.9999998807907104,
604
+ "28": 1.0,
605
+ "29": 0.9999998807907104,
606
+ "30": 1.0,
607
+ "31": 1.0,
608
+ "32": 1.0,
609
+ "33": 0.9999999403953552,
610
+ "34": 1.0000001192092896,
611
+ "35": 1.0000001192092896,
612
+ "36": 0.9999997615814209,
613
+ "37": 1.0000001192092896,
614
+ "38": 1.0000001192092896,
615
+ "39": 1.0,
616
+ "42": 1.0,
617
+ "45": 1.0,
618
+ "47": 1.0000001192092896
619
+ },
620
+ "v2_ortho_general__VS__v3_ortho_crossdim": {
621
+ "0": 0.999434232711792,
622
+ "2": 0.9891625046730042,
623
+ "3": 0.9885857105255127,
624
+ "4": 0.9918666481971741,
625
+ "6": 0.9937582612037659,
626
+ "7": 0.9952627420425415,
627
+ "8": 0.9950225353240967,
628
+ "9": 0.9957668781280518,
629
+ "10": 0.9982725381851196,
630
+ "11": 0.9992640614509583,
631
+ "12": 0.9998573064804077,
632
+ "13": 0.9995517134666443,
633
+ "14": 0.9987476468086243,
634
+ "15": 0.999505341053009,
635
+ "17": 0.9999439716339111,
636
+ "18": 0.9999611973762512,
637
+ "20": 0.9996616244316101,
638
+ "21": 0.9982099533081055,
639
+ "22": 0.99882972240448,
640
+ "23": 0.9992284774780273,
641
+ "24": 0.999238908290863,
642
+ "25": 0.9961854219436646,
643
+ "26": 0.992924690246582,
644
+ "27": 0.9964101910591125,
645
+ "28": 0.9949069023132324,
646
+ "29": 0.9949198961257935,
647
+ "30": 0.9973813891410828,
648
+ "31": 0.9976941347122192,
649
+ "32": 0.9953215718269348,
650
+ "33": 0.9955788850784302,
651
+ "34": 0.9964847564697266,
652
+ "35": 0.9968671202659607,
653
+ "36": 0.9977893829345703,
654
+ "37": 0.9949166774749756,
655
+ "38": 0.9955715537071228,
656
+ "39": 0.995618462562561,
657
+ "42": 0.9917940497398376,
658
+ "45": 0.9917187094688416,
659
+ "47": 0.9709324836730957
660
+ },
661
+ "v2_ortho_general__VS__v4_pca": {
662
+ "0": 0.9704648852348328,
663
+ "2": 0.9660898447036743,
664
+ "3": 0.976309061050415,
665
+ "4": 0.973873496055603,
666
+ "6": 0.9742251634597778,
667
+ "7": 0.9786409139633179,
668
+ "8": 0.9782382249832153,
669
+ "9": 0.9800677299499512,
670
+ "10": 0.9794178009033203,
671
+ "11": 0.9806268811225891,
672
+ "12": 0.97905433177948,
673
+ "13": 0.9834149479866028,
674
+ "14": 0.9858850240707397,
675
+ "15": 0.986358106136322,
676
+ "17": 0.9886331558227539,
677
+ "18": 0.9865451455116272,
678
+ "20": 0.9876081347465515,
679
+ "21": 0.9872385263442993,
680
+ "22": 0.9864363074302673,
681
+ "23": 0.9864980578422546,
682
+ "24": 0.9864625334739685,
683
+ "25": 0.9875662922859192,
684
+ "26": 0.9870969653129578,
685
+ "27": 0.9895696043968201,
686
+ "28": 0.9889065623283386,
687
+ "29": 0.9888129234313965,
688
+ "30": 0.989903450012207,
689
+ "31": 0.9909472465515137,
690
+ "32": 0.9871730804443359,
691
+ "33": 0.9868060350418091,
692
+ "34": 0.9871065616607666,
693
+ "35": 0.9884214997291565,
694
+ "36": 0.9875127077102661,
695
+ "37": 0.9869857430458069,
696
+ "38": 0.985481321811676,
697
+ "39": 0.9858270883560181,
698
+ "42": 0.9718344211578369,
699
+ "45": 0.9732871651649475,
700
+ "47": 0.9724766612052917
701
+ },
702
+ "v3_ortho_crossdim__VS__v3_ortho_crossdim": {
703
+ "0": 1.0,
704
+ "2": 1.0000001192092896,
705
+ "3": 1.0,
706
+ "4": 1.0,
707
+ "6": 1.0,
708
+ "7": 0.9999998807907104,
709
+ "8": 1.0000001192092896,
710
+ "9": 1.0,
711
+ "10": 1.0,
712
+ "11": 1.0000001192092896,
713
+ "12": 1.0000001192092896,
714
+ "13": 1.0000001192092896,
715
+ "14": 1.000000238418579,
716
+ "15": 1.0,
717
+ "17": 1.0,
718
+ "18": 1.000000238418579,
719
+ "20": 1.0000001192092896,
720
+ "21": 1.0000001192092896,
721
+ "22": 1.000000238418579,
722
+ "23": 0.9999998807907104,
723
+ "24": 1.0,
724
+ "25": 1.0000001192092896,
725
+ "26": 1.0,
726
+ "27": 0.9999999403953552,
727
+ "28": 1.0000001192092896,
728
+ "29": 1.0000001192092896,
729
+ "30": 1.0,
730
+ "31": 1.0,
731
+ "32": 1.0000001192092896,
732
+ "33": 1.0000001192092896,
733
+ "34": 0.9999998807907104,
734
+ "35": 1.0000001192092896,
735
+ "36": 1.0,
736
+ "37": 1.000000238418579,
737
+ "38": 1.0,
738
+ "39": 1.0,
739
+ "42": 1.0000001192092896,
740
+ "45": 1.0,
741
+ "47": 1.000000238418579
742
+ },
743
+ "v3_ortho_crossdim__VS__v4_pca": {
744
+ "0": 0.9705691337585449,
745
+ "2": 0.9698101282119751,
746
+ "3": 0.9832841157913208,
747
+ "4": 0.9769375324249268,
748
+ "6": 0.9744015336036682,
749
+ "7": 0.9785113334655762,
750
+ "8": 0.9785210490226746,
751
+ "9": 0.9804296493530273,
752
+ "10": 0.9784255027770996,
753
+ "11": 0.9795787334442139,
754
+ "12": 0.9783607125282288,
755
+ "13": 0.9824701547622681,
756
+ "14": 0.9848558306694031,
757
+ "15": 0.988103449344635,
758
+ "17": 0.9890871644020081,
759
+ "18": 0.9869385361671448,
760
+ "20": 0.9868893027305603,
761
+ "21": 0.9866880774497986,
762
+ "22": 0.9855419993400574,
763
+ "23": 0.9855183959007263,
764
+ "24": 0.9854879379272461,
765
+ "25": 0.9875563979148865,
766
+ "26": 0.9891694784164429,
767
+ "27": 0.9899845719337463,
768
+ "28": 0.990175187587738,
769
+ "29": 0.9901189208030701,
770
+ "30": 0.9897226691246033,
771
+ "31": 0.9907418489456177,
772
+ "32": 0.9878955483436584,
773
+ "33": 0.987434446811676,
774
+ "34": 0.9871537089347839,
775
+ "35": 0.988495409488678,
776
+ "36": 0.9870281219482422,
777
+ "37": 0.9881243705749512,
778
+ "38": 0.9858537912368774,
779
+ "39": 0.9865440130233765,
780
+ "42": 0.9733933210372925,
781
+ "45": 0.9753375053405762,
782
+ "47": 0.9966042637825012
783
+ },
784
+ "v4_pca__VS__v4_pca": {
785
+ "0": 0.9999999403953552,
786
+ "2": 1.0000001192092896,
787
+ "3": 1.0,
788
+ "4": 1.0000001192092896,
789
+ "6": 1.0000001192092896,
790
+ "7": 0.9999998807907104,
791
+ "8": 1.0000001192092896,
792
+ "9": 1.0,
793
+ "10": 0.9999998807907104,
794
+ "11": 1.0,
795
+ "12": 1.0,
796
+ "13": 1.0,
797
+ "14": 1.0,
798
+ "15": 1.0,
799
+ "17": 1.0,
800
+ "18": 0.9999999403953552,
801
+ "20": 1.0,
802
+ "21": 1.0000001192092896,
803
+ "22": 1.0,
804
+ "23": 0.9999999403953552,
805
+ "24": 1.0,
806
+ "25": 1.0000001192092896,
807
+ "26": 1.0,
808
+ "27": 1.0000001192092896,
809
+ "28": 1.0,
810
+ "29": 1.0,
811
+ "30": 1.000000238418579,
812
+ "31": 0.9999998807907104,
813
+ "32": 1.0,
814
+ "33": 1.0,
815
+ "34": 1.0,
816
+ "35": 1.0000001192092896,
817
+ "36": 1.0,
818
+ "37": 1.0000001192092896,
819
+ "38": 1.0,
820
+ "39": 1.0,
821
+ "42": 0.9999999403953552,
822
+ "45": 1.0,
823
+ "47": 1.0
824
+ }
825
+ },
826
+ "cross_dim_per_version": {
827
+ "plan_v1_raw__VS__mon_v1_raw": {
828
+ "0": -0.0067728618159890175,
829
+ "2": -0.055668748915195465,
830
+ "3": -0.08748188614845276,
831
+ "4": -0.03939281776547432,
832
+ "6": 0.019152695313096046,
833
+ "7": 0.031105361878871918,
834
+ "8": 0.028466008603572845,
835
+ "9": 0.027173638343811035,
836
+ "10": 0.0709397941827774,
837
+ "11": 0.09230419248342514,
838
+ "12": 0.1240522637963295,
839
+ "13": 0.10487423837184906,
840
+ "14": 0.08985582739114761,
841
+ "15": 0.1630288064479828,
842
+ "17": 0.1403084546327591,
843
+ "18": 0.15332505106925964,
844
+ "20": 0.1202765479683876,
845
+ "21": 0.07777776569128036,
846
+ "22": 0.10294556617736816,
847
+ "23": 0.11596696823835373,
848
+ "24": 0.11921444535255432,
849
+ "25": 0.0632816031575203,
850
+ "26": 0.021381724625825882,
851
+ "27": 0.038544606417417526,
852
+ "28": 0.026335537433624268,
853
+ "29": 0.021619966253638268,
854
+ "30": 0.06121957302093506,
855
+ "31": 0.061807166785001755,
856
+ "32": 0.03239646181464195,
857
+ "33": 0.025921350345015526,
858
+ "34": 0.03938739001750946,
859
+ "35": 0.03411402925848961,
860
+ "36": 0.04753681272268295,
861
+ "37": 0.010360519401729107,
862
+ "38": 0.03580017760396004,
863
+ "39": 0.025486305356025696,
864
+ "42": 0.011071556247770786,
865
+ "45": 0.002234235405921936,
866
+ "47": -0.17639747262001038
867
+ },
868
+ "plan_v2_ortho_general__VS__mon_v2_ortho_general": {
869
+ "0": -0.033635057508945465,
870
+ "2": -0.14682605862617493,
871
+ "3": -0.1506599485874176,
872
+ "4": -0.12728174030780792,
873
+ "6": -0.1115565076470375,
874
+ "7": -0.09722047299146652,
875
+ "8": -0.09964963793754578,
876
+ "9": -0.09191551059484482,
877
+ "10": -0.05875474214553833,
878
+ "11": -0.03835861757397652,
879
+ "12": -0.016897650435566902,
880
+ "13": -0.029941793531179428,
881
+ "14": -0.05003418028354645,
882
+ "15": 0.03145408630371094,
883
+ "17": 0.010606815107166767,
884
+ "18": 0.0088149169459939,
885
+ "20": -0.02601168490946293,
886
+ "21": -0.05980429798364639,
887
+ "22": -0.048366472125053406,
888
+ "23": -0.039275217801332474,
889
+ "24": -0.03900831192731857,
890
+ "25": -0.08726224303245544,
891
+ "26": -0.11874494701623917,
892
+ "27": -0.0846562385559082,
893
+ "28": -0.10079923272132874,
894
+ "29": -0.10066935420036316,
895
+ "30": -0.0723218321800232,
896
+ "31": -0.0678703561425209,
897
+ "32": -0.09661855548620224,
898
+ "33": -0.09392993897199631,
899
+ "34": -0.0837741568684578,
900
+ "35": -0.07909621298313141,
901
+ "36": -0.06645289063453674,
902
+ "37": -0.10070281475782394,
903
+ "38": -0.09400738030672073,
904
+ "39": -0.0935090109705925,
905
+ "42": -0.1278461068868637,
906
+ "45": -0.12842951714992523,
907
+ "47": -0.23935407400131226
908
+ },
909
+ "plan_v3_ortho_crossdim__VS__mon_v3_ortho_crossdim": {
910
+ "0": 0.03363505005836487,
911
+ "2": 0.14682608842849731,
912
+ "3": 0.15065988898277283,
913
+ "4": 0.1272818148136139,
914
+ "6": 0.1115565225481987,
915
+ "7": 0.09722045063972473,
916
+ "8": 0.09964966028928757,
917
+ "9": 0.09191549569368362,
918
+ "10": 0.05875474587082863,
919
+ "11": 0.03835863620042801,
920
+ "12": 0.016897615045309067,
921
+ "13": 0.029941823333501816,
922
+ "14": 0.05003422498703003,
923
+ "15": -0.03145407512784004,
924
+ "17": -0.01060680951923132,
925
+ "18": -0.008814888074994087,
926
+ "20": 0.026011662557721138,
927
+ "21": 0.05980434641242027,
928
+ "22": 0.04836644232273102,
929
+ "23": 0.03927519544959068,
930
+ "24": 0.03900832682847977,
931
+ "25": 0.08726222068071365,
932
+ "26": 0.11874491721391678,
933
+ "27": 0.0846562311053276,
934
+ "28": 0.10079924017190933,
935
+ "29": 0.10066938400268555,
936
+ "30": 0.072321817278862,
937
+ "31": 0.06787031888961792,
938
+ "32": 0.09661854058504105,
939
+ "33": 0.09392993152141571,
940
+ "34": 0.083774134516716,
941
+ "35": 0.07909620553255081,
942
+ "36": 0.06645289808511734,
943
+ "37": 0.10070281475782394,
944
+ "38": 0.09400739520788193,
945
+ "39": 0.0935090109705925,
946
+ "42": 0.1278461217880249,
947
+ "45": 0.12842954695224762,
948
+ "47": 0.23935411870479584
949
+ },
950
+ "plan_v4_pca__VS__mon_v4_pca": {
951
+ "0": 0.01947270706295967,
952
+ "2": 0.09903767704963684,
953
+ "3": 0.12222529947757721,
954
+ "4": 0.08811970800161362,
955
+ "6": 0.05799045041203499,
956
+ "7": 0.04790189862251282,
957
+ "8": 0.05323654040694237,
958
+ "9": 0.050345081835985184,
959
+ "10": 0.012233899906277657,
960
+ "11": -0.008752653375267982,
961
+ "12": -0.0337398424744606,
962
+ "13": -0.01738101802766323,
963
+ "14": 0.004263312090188265,
964
+ "15": -0.07339227944612503,
965
+ "17": -0.04966166615486145,
966
+ "18": -0.05051732808351517,
967
+ "20": -0.01528917346149683,
968
+ "21": 0.020983245223760605,
969
+ "22": 0.00556010939180851,
970
+ "23": -0.00576805742457509,
971
+ "24": -0.005948912352323532,
972
+ "25": 0.04493141546845436,
973
+ "26": 0.07985581457614899,
974
+ "27": 0.0487092062830925,
975
+ "28": 0.06515415757894516,
976
+ "29": 0.06540689617395401,
977
+ "30": 0.03469874709844589,
978
+ "31": 0.031902190297842026,
979
+ "32": 0.05744742229580879,
980
+ "33": 0.05508880317211151,
981
+ "34": 0.04355907812714577,
982
+ "35": 0.04154374077916145,
983
+ "36": 0.026454683393239975,
984
+ "37": 0.06337542831897736,
985
+ "38": 0.05226850137114525,
986
+ "39": 0.05560529977083206,
987
+ "42": 0.07687295973300934,
988
+ "45": 0.08101249486207962,
989
+ "47": 0.22699248790740967
990
+ }
991
+ }
992
+ }
results/downstream_accuracy.json ADDED
@@ -0,0 +1,1534 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "baseline": {
3
+ "MATH-500-holdout": {
4
+ "accuracy": 0.3,
5
+ "correct": 15,
6
+ "n": 50,
7
+ "per_sample": [
8
+ {
9
+ "idx": 450,
10
+ "pred": "48",
11
+ "gold": "64",
12
+ "correct": false
13
+ },
14
+ {
15
+ "idx": 451,
16
+ "pred": "71",
17
+ "gold": "\\frac{639}{40}",
18
+ "correct": false
19
+ },
20
+ {
21
+ "idx": 452,
22
+ "pred": "143",
23
+ "gold": "143",
24
+ "correct": true
25
+ },
26
+ {
27
+ "idx": 453,
28
+ "pred": "10",
29
+ "gold": "10",
30
+ "correct": true
31
+ },
32
+ {
33
+ "idx": 454,
34
+ "pred": "3",
35
+ "gold": "y = 2x + 3",
36
+ "correct": false
37
+ },
38
+ {
39
+ "idx": 455,
40
+ "pred": "6",
41
+ "gold": "6",
42
+ "correct": true
43
+ },
44
+ {
45
+ "idx": 456,
46
+ "pred": "7",
47
+ "gold": "-2,1",
48
+ "correct": false
49
+ },
50
+ {
51
+ "idx": 457,
52
+ "pred": "0",
53
+ "gold": "0",
54
+ "correct": true
55
+ },
56
+ {
57
+ "idx": 458,
58
+ "pred": "2",
59
+ "gold": "41",
60
+ "correct": false
61
+ },
62
+ {
63
+ "idx": 459,
64
+ "pred": "5",
65
+ "gold": "\\$18.90",
66
+ "correct": false
67
+ },
68
+ {
69
+ "idx": 460,
70
+ "pred": "35",
71
+ "gold": "75^\\circ",
72
+ "correct": false
73
+ },
74
+ {
75
+ "idx": 461,
76
+ "pred": "12.",
77
+ "gold": "12",
78
+ "correct": true
79
+ },
80
+ {
81
+ "idx": 462,
82
+ "pred": "2",
83
+ "gold": "59",
84
+ "correct": false
85
+ },
86
+ {
87
+ "idx": 463,
88
+ "pred": "2",
89
+ "gold": "2",
90
+ "correct": true
91
+ },
92
+ {
93
+ "idx": 464,
94
+ "pred": "1",
95
+ "gold": "9",
96
+ "correct": false
97
+ },
98
+ {
99
+ "idx": 465,
100
+ "pred": "3",
101
+ "gold": "121",
102
+ "correct": false
103
+ },
104
+ {
105
+ "idx": 466,
106
+ "pred": "1",
107
+ "gold": "8 \\pi",
108
+ "correct": false
109
+ },
110
+ {
111
+ "idx": 467,
112
+ "pred": "0",
113
+ "gold": "15\\mbox{ cm}^2",
114
+ "correct": false
115
+ },
116
+ {
117
+ "idx": 468,
118
+ "pred": "8",
119
+ "gold": "8",
120
+ "correct": true
121
+ },
122
+ {
123
+ "idx": 469,
124
+ "pred": "-1.2",
125
+ "gold": "25",
126
+ "correct": false
127
+ },
128
+ {
129
+ "idx": 470,
130
+ "pred": "3",
131
+ "gold": "27648",
132
+ "correct": false
133
+ },
134
+ {
135
+ "idx": 471,
136
+ "pred": "84",
137
+ "gold": "84",
138
+ "correct": true
139
+ },
140
+ {
141
+ "idx": 472,
142
+ "pred": "137\\frac{1",
143
+ "gold": "137 \\frac{1}{2}",
144
+ "correct": false
145
+ },
146
+ {
147
+ "idx": 473,
148
+ "pred": "2.5",
149
+ "gold": "7",
150
+ "correct": false
151
+ },
152
+ {
153
+ "idx": 474,
154
+ "pred": "6",
155
+ "gold": "3",
156
+ "correct": false
157
+ },
158
+ {
159
+ "idx": 475,
160
+ "pred": "7",
161
+ "gold": "29",
162
+ "correct": false
163
+ },
164
+ {
165
+ "idx": 476,
166
+ "pred": "200",
167
+ "gold": "200",
168
+ "correct": true
169
+ },
170
+ {
171
+ "idx": 477,
172
+ "pred": "9",
173
+ "gold": "2",
174
+ "correct": false
175
+ },
176
+ {
177
+ "idx": 478,
178
+ "pred": "4",
179
+ "gold": "\\frac{2}{1005}",
180
+ "correct": false
181
+ },
182
+ {
183
+ "idx": 479,
184
+ "pred": "2.",
185
+ "gold": "2",
186
+ "correct": true
187
+ },
188
+ {
189
+ "idx": 480,
190
+ "pred": "1.",
191
+ "gold": "1",
192
+ "correct": true
193
+ },
194
+ {
195
+ "idx": 481,
196
+ "pred": "2",
197
+ "gold": "8n^2 + 4n + 1",
198
+ "correct": false
199
+ },
200
+ {
201
+ "idx": 482,
202
+ "pred": "20",
203
+ "gold": "202",
204
+ "correct": false
205
+ },
206
+ {
207
+ "idx": 483,
208
+ "pred": "180",
209
+ "gold": "\\frac{8}{15}",
210
+ "correct": false
211
+ },
212
+ {
213
+ "idx": 484,
214
+ "pred": "6",
215
+ "gold": "10",
216
+ "correct": false
217
+ },
218
+ {
219
+ "idx": 485,
220
+ "pred": "6",
221
+ "gold": "13",
222
+ "correct": false
223
+ },
224
+ {
225
+ "idx": 486,
226
+ "pred": "216.",
227
+ "gold": "216",
228
+ "correct": true
229
+ },
230
+ {
231
+ "idx": 487,
232
+ "pred": "2",
233
+ "gold": "2",
234
+ "correct": true
235
+ },
236
+ {
237
+ "idx": 488,
238
+ "pred": "36",
239
+ "gold": "\\$36",
240
+ "correct": false
241
+ },
242
+ {
243
+ "idx": 489,
244
+ "pred": "15",
245
+ "gold": "15",
246
+ "correct": true
247
+ },
248
+ {
249
+ "idx": 490,
250
+ "pred": "-11",
251
+ "gold": "14",
252
+ "correct": false
253
+ },
254
+ {
255
+ "idx": 491,
256
+ "pred": "3",
257
+ "gold": "64",
258
+ "correct": false
259
+ },
260
+ {
261
+ "idx": 492,
262
+ "pred": "y = -2x",
263
+ "gold": "y = -2x",
264
+ "correct": true
265
+ },
266
+ {
267
+ "idx": 493,
268
+ "pred": "2",
269
+ "gold": "20",
270
+ "correct": false
271
+ },
272
+ {
273
+ "idx": 494,
274
+ "pred": "2",
275
+ "gold": "\\frac{1}{4}",
276
+ "correct": false
277
+ },
278
+ {
279
+ "idx": 495,
280
+ "pred": "10",
281
+ "gold": "(2,12) \\cup (12,102)",
282
+ "correct": false
283
+ },
284
+ {
285
+ "idx": 496,
286
+ "pred": "\\frac{5",
287
+ "gold": "\\frac{5}{13}",
288
+ "correct": false
289
+ },
290
+ {
291
+ "idx": 497,
292
+ "pred": "2",
293
+ "gold": "\\frac{7}{2}",
294
+ "correct": false
295
+ },
296
+ {
297
+ "idx": 498,
298
+ "pred": "0",
299
+ "gold": "-1",
300
+ "correct": false
301
+ },
302
+ {
303
+ "idx": 499,
304
+ "pred": "54",
305
+ "gold": "106^\\circ",
306
+ "correct": false
307
+ }
308
+ ]
309
+ },
310
+ "AIME-24": {
311
+ "accuracy": 0.03333333333333333,
312
+ "correct": 1,
313
+ "n": 30,
314
+ "per_sample": [
315
+ {
316
+ "idx": "aime24_0",
317
+ "pred": "204",
318
+ "gold": "204",
319
+ "correct": true
320
+ },
321
+ {
322
+ "idx": "aime24_1",
323
+ "pred": "2",
324
+ "gold": "113",
325
+ "correct": false
326
+ },
327
+ {
328
+ "idx": "aime24_2",
329
+ "pred": "1",
330
+ "gold": "371",
331
+ "correct": false
332
+ },
333
+ {
334
+ "idx": "aime24_3",
335
+ "pred": "1",
336
+ "gold": "385",
337
+ "correct": false
338
+ },
339
+ {
340
+ "idx": "aime24_4",
341
+ "pred": "3",
342
+ "gold": "110",
343
+ "correct": false
344
+ },
345
+ {
346
+ "idx": "aime24_5",
347
+ "pred": "756.",
348
+ "gold": "104",
349
+ "correct": false
350
+ },
351
+ {
352
+ "idx": "aime24_6",
353
+ "pred": "2",
354
+ "gold": "721",
355
+ "correct": false
356
+ },
357
+ {
358
+ "idx": "aime24_7",
359
+ "pred": "10",
360
+ "gold": "025",
361
+ "correct": false
362
+ },
363
+ {
364
+ "idx": "aime24_8",
365
+ "pred": "8",
366
+ "gold": "809",
367
+ "correct": false
368
+ },
369
+ {
370
+ "idx": "aime24_9",
371
+ "pred": "115",
372
+ "gold": "116",
373
+ "correct": false
374
+ },
375
+ {
376
+ "idx": "aime24_10",
377
+ "pred": "107.",
378
+ "gold": "104",
379
+ "correct": false
380
+ },
381
+ {
382
+ "idx": "aime24_11",
383
+ "pred": "2",
384
+ "gold": "294",
385
+ "correct": false
386
+ },
387
+ {
388
+ "idx": "aime24_12",
389
+ "pred": "16",
390
+ "gold": "540",
391
+ "correct": false
392
+ },
393
+ {
394
+ "idx": "aime24_13",
395
+ "pred": "2",
396
+ "gold": "197",
397
+ "correct": false
398
+ },
399
+ {
400
+ "idx": "aime24_14",
401
+ "pred": "24.",
402
+ "gold": "480",
403
+ "correct": false
404
+ },
405
+ {
406
+ "idx": "aime24_15",
407
+ "pred": "3",
408
+ "gold": "073",
409
+ "correct": false
410
+ },
411
+ {
412
+ "idx": "aime24_16",
413
+ "pred": "3",
414
+ "gold": "468",
415
+ "correct": false
416
+ },
417
+ {
418
+ "idx": "aime24_17",
419
+ "pred": "30",
420
+ "gold": "601",
421
+ "correct": false
422
+ },
423
+ {
424
+ "idx": "aime24_18",
425
+ "pred": "0.",
426
+ "gold": "023",
427
+ "correct": false
428
+ },
429
+ {
430
+ "idx": "aime24_19",
431
+ "pred": "1",
432
+ "gold": "321",
433
+ "correct": false
434
+ },
435
+ {
436
+ "idx": "aime24_20",
437
+ "pred": "8",
438
+ "gold": "211",
439
+ "correct": false
440
+ },
441
+ {
442
+ "idx": "aime24_21",
443
+ "pred": "0.",
444
+ "gold": "315",
445
+ "correct": false
446
+ },
447
+ {
448
+ "idx": "aime24_22",
449
+ "pred": "5",
450
+ "gold": "236",
451
+ "correct": false
452
+ },
453
+ {
454
+ "idx": "aime24_23",
455
+ "pred": "2",
456
+ "gold": "045",
457
+ "correct": false
458
+ },
459
+ {
460
+ "idx": "aime24_24",
461
+ "pred": "8.",
462
+ "gold": "033",
463
+ "correct": false
464
+ },
465
+ {
466
+ "idx": "aime24_25",
467
+ "pred": "0",
468
+ "gold": "080",
469
+ "correct": false
470
+ },
471
+ {
472
+ "idx": "aime24_26",
473
+ "pred": "487",
474
+ "gold": "055",
475
+ "correct": false
476
+ },
477
+ {
478
+ "idx": "aime24_27",
479
+ "pred": "1",
480
+ "gold": "699",
481
+ "correct": false
482
+ },
483
+ {
484
+ "idx": "aime24_28",
485
+ "pred": "0.",
486
+ "gold": "127",
487
+ "correct": false
488
+ },
489
+ {
490
+ "idx": "aime24_29",
491
+ "pred": "25",
492
+ "gold": "902",
493
+ "correct": false
494
+ }
495
+ ]
496
+ }
497
+ },
498
+ "plan_alpha_0": {
499
+ "MATH-500-holdout": {
500
+ "accuracy": 0.04,
501
+ "correct": 2,
502
+ "n": 50,
503
+ "per_sample": [
504
+ {
505
+ "idx": 450,
506
+ "pred": "",
507
+ "gold": "64",
508
+ "correct": false
509
+ },
510
+ {
511
+ "idx": 451,
512
+ "pred": "",
513
+ "gold": "\\frac{639}{40}",
514
+ "correct": false
515
+ },
516
+ {
517
+ "idx": 452,
518
+ "pred": "",
519
+ "gold": "143",
520
+ "correct": false
521
+ },
522
+ {
523
+ "idx": 453,
524
+ "pred": "",
525
+ "gold": "10",
526
+ "correct": false
527
+ },
528
+ {
529
+ "idx": 454,
530
+ "pred": "",
531
+ "gold": "y = 2x + 3",
532
+ "correct": false
533
+ },
534
+ {
535
+ "idx": 455,
536
+ "pred": "",
537
+ "gold": "6",
538
+ "correct": false
539
+ },
540
+ {
541
+ "idx": 456,
542
+ "pred": "",
543
+ "gold": "-2,1",
544
+ "correct": false
545
+ },
546
+ {
547
+ "idx": 457,
548
+ "pred": "",
549
+ "gold": "0",
550
+ "correct": false
551
+ },
552
+ {
553
+ "idx": 458,
554
+ "pred": "",
555
+ "gold": "41",
556
+ "correct": false
557
+ },
558
+ {
559
+ "idx": 459,
560
+ "pred": "18.90",
561
+ "gold": "\\$18.90",
562
+ "correct": false
563
+ },
564
+ {
565
+ "idx": 460,
566
+ "pred": "0",
567
+ "gold": "75^\\circ",
568
+ "correct": false
569
+ },
570
+ {
571
+ "idx": 461,
572
+ "pred": "",
573
+ "gold": "12",
574
+ "correct": false
575
+ },
576
+ {
577
+ "idx": 462,
578
+ "pred": "",
579
+ "gold": "59",
580
+ "correct": false
581
+ },
582
+ {
583
+ "idx": 463,
584
+ "pred": "2",
585
+ "gold": "2",
586
+ "correct": true
587
+ },
588
+ {
589
+ "idx": 464,
590
+ "pred": "",
591
+ "gold": "9",
592
+ "correct": false
593
+ },
594
+ {
595
+ "idx": 465,
596
+ "pred": "3",
597
+ "gold": "121",
598
+ "correct": false
599
+ },
600
+ {
601
+ "idx": 466,
602
+ "pred": "",
603
+ "gold": "8 \\pi",
604
+ "correct": false
605
+ },
606
+ {
607
+ "idx": 467,
608
+ "pred": "",
609
+ "gold": "15\\mbox{ cm}^2",
610
+ "correct": false
611
+ },
612
+ {
613
+ "idx": 468,
614
+ "pred": "",
615
+ "gold": "8",
616
+ "correct": false
617
+ },
618
+ {
619
+ "idx": 469,
620
+ "pred": "",
621
+ "gold": "25",
622
+ "correct": false
623
+ },
624
+ {
625
+ "idx": 470,
626
+ "pred": "",
627
+ "gold": "27648",
628
+ "correct": false
629
+ },
630
+ {
631
+ "idx": 471,
632
+ "pred": "2520",
633
+ "gold": "84",
634
+ "correct": false
635
+ },
636
+ {
637
+ "idx": 472,
638
+ "pred": "",
639
+ "gold": "137 \\frac{1}{2}",
640
+ "correct": false
641
+ },
642
+ {
643
+ "idx": 473,
644
+ "pred": "2",
645
+ "gold": "7",
646
+ "correct": false
647
+ },
648
+ {
649
+ "idx": 474,
650
+ "pred": "",
651
+ "gold": "3",
652
+ "correct": false
653
+ },
654
+ {
655
+ "idx": 475,
656
+ "pred": "",
657
+ "gold": "29",
658
+ "correct": false
659
+ },
660
+ {
661
+ "idx": 476,
662
+ "pred": "",
663
+ "gold": "200",
664
+ "correct": false
665
+ },
666
+ {
667
+ "idx": 477,
668
+ "pred": "",
669
+ "gold": "2",
670
+ "correct": false
671
+ },
672
+ {
673
+ "idx": 478,
674
+ "pred": "",
675
+ "gold": "\\frac{2}{1005}",
676
+ "correct": false
677
+ },
678
+ {
679
+ "idx": 479,
680
+ "pred": "2",
681
+ "gold": "2",
682
+ "correct": true
683
+ },
684
+ {
685
+ "idx": 480,
686
+ "pred": "",
687
+ "gold": "1",
688
+ "correct": false
689
+ },
690
+ {
691
+ "idx": 481,
692
+ "pred": "",
693
+ "gold": "8n^2 + 4n + 1",
694
+ "correct": false
695
+ },
696
+ {
697
+ "idx": 482,
698
+ "pred": "",
699
+ "gold": "202",
700
+ "correct": false
701
+ },
702
+ {
703
+ "idx": 483,
704
+ "pred": "",
705
+ "gold": "\\frac{8}{15}",
706
+ "correct": false
707
+ },
708
+ {
709
+ "idx": 484,
710
+ "pred": "",
711
+ "gold": "10",
712
+ "correct": false
713
+ },
714
+ {
715
+ "idx": 485,
716
+ "pred": "",
717
+ "gold": "13",
718
+ "correct": false
719
+ },
720
+ {
721
+ "idx": 486,
722
+ "pred": "",
723
+ "gold": "216",
724
+ "correct": false
725
+ },
726
+ {
727
+ "idx": 487,
728
+ "pred": "",
729
+ "gold": "2",
730
+ "correct": false
731
+ },
732
+ {
733
+ "idx": 488,
734
+ "pred": "",
735
+ "gold": "\\$36",
736
+ "correct": false
737
+ },
738
+ {
739
+ "idx": 489,
740
+ "pred": "",
741
+ "gold": "15",
742
+ "correct": false
743
+ },
744
+ {
745
+ "idx": 490,
746
+ "pred": "",
747
+ "gold": "14",
748
+ "correct": false
749
+ },
750
+ {
751
+ "idx": 491,
752
+ "pred": "",
753
+ "gold": "64",
754
+ "correct": false
755
+ },
756
+ {
757
+ "idx": 492,
758
+ "pred": "",
759
+ "gold": "y = -2x",
760
+ "correct": false
761
+ },
762
+ {
763
+ "idx": 493,
764
+ "pred": "3.",
765
+ "gold": "20",
766
+ "correct": false
767
+ },
768
+ {
769
+ "idx": 494,
770
+ "pred": "",
771
+ "gold": "\\frac{1}{4}",
772
+ "correct": false
773
+ },
774
+ {
775
+ "idx": 495,
776
+ "pred": "",
777
+ "gold": "(2,12) \\cup (12,102)",
778
+ "correct": false
779
+ },
780
+ {
781
+ "idx": 496,
782
+ "pred": "",
783
+ "gold": "\\frac{5}{13}",
784
+ "correct": false
785
+ },
786
+ {
787
+ "idx": 497,
788
+ "pred": "",
789
+ "gold": "\\frac{7}{2}",
790
+ "correct": false
791
+ },
792
+ {
793
+ "idx": 498,
794
+ "pred": "",
795
+ "gold": "-1",
796
+ "correct": false
797
+ },
798
+ {
799
+ "idx": 499,
800
+ "pred": "",
801
+ "gold": "106^\\circ",
802
+ "correct": false
803
+ }
804
+ ],
805
+ "vs_baseline": {
806
+ "baseline_accuracy": 0.3,
807
+ "steered_accuracy": 0.04,
808
+ "absolute_drop": 0.26,
809
+ "relative_drop": 0.8666666666666667,
810
+ "n_common": 50,
811
+ "n_regressions": 13,
812
+ "n_recoveries": 0,
813
+ "mcnemar_p_value": 0.000244140625,
814
+ "significant_at_0_05": true
815
+ }
816
+ },
817
+ "AIME-24": {
818
+ "accuracy": 0.0,
819
+ "correct": 0,
820
+ "n": 30,
821
+ "per_sample": [
822
+ {
823
+ "idx": "aime24_0",
824
+ "pred": "9",
825
+ "gold": "204",
826
+ "correct": false
827
+ },
828
+ {
829
+ "idx": "aime24_1",
830
+ "pred": "",
831
+ "gold": "113",
832
+ "correct": false
833
+ },
834
+ {
835
+ "idx": "aime24_2",
836
+ "pred": "",
837
+ "gold": "371",
838
+ "correct": false
839
+ },
840
+ {
841
+ "idx": "aime24_3",
842
+ "pred": "",
843
+ "gold": "385",
844
+ "correct": false
845
+ },
846
+ {
847
+ "idx": "aime24_4",
848
+ "pred": "",
849
+ "gold": "110",
850
+ "correct": false
851
+ },
852
+ {
853
+ "idx": "aime24_5",
854
+ "pred": "",
855
+ "gold": "104",
856
+ "correct": false
857
+ },
858
+ {
859
+ "idx": "aime24_6",
860
+ "pred": "",
861
+ "gold": "721",
862
+ "correct": false
863
+ },
864
+ {
865
+ "idx": "aime24_7",
866
+ "pred": "",
867
+ "gold": "025",
868
+ "correct": false
869
+ },
870
+ {
871
+ "idx": "aime24_8",
872
+ "pred": "",
873
+ "gold": "809",
874
+ "correct": false
875
+ },
876
+ {
877
+ "idx": "aime24_9",
878
+ "pred": "",
879
+ "gold": "116",
880
+ "correct": false
881
+ },
882
+ {
883
+ "idx": "aime24_10",
884
+ "pred": "",
885
+ "gold": "104",
886
+ "correct": false
887
+ },
888
+ {
889
+ "idx": "aime24_11",
890
+ "pred": "",
891
+ "gold": "294",
892
+ "correct": false
893
+ },
894
+ {
895
+ "idx": "aime24_12",
896
+ "pred": "",
897
+ "gold": "540",
898
+ "correct": false
899
+ },
900
+ {
901
+ "idx": "aime24_13",
902
+ "pred": "",
903
+ "gold": "197",
904
+ "correct": false
905
+ },
906
+ {
907
+ "idx": "aime24_14",
908
+ "pred": "",
909
+ "gold": "480",
910
+ "correct": false
911
+ },
912
+ {
913
+ "idx": "aime24_15",
914
+ "pred": "",
915
+ "gold": "073",
916
+ "correct": false
917
+ },
918
+ {
919
+ "idx": "aime24_16",
920
+ "pred": "",
921
+ "gold": "468",
922
+ "correct": false
923
+ },
924
+ {
925
+ "idx": "aime24_17",
926
+ "pred": "",
927
+ "gold": "601",
928
+ "correct": false
929
+ },
930
+ {
931
+ "idx": "aime24_18",
932
+ "pred": "",
933
+ "gold": "023",
934
+ "correct": false
935
+ },
936
+ {
937
+ "idx": "aime24_19",
938
+ "pred": "",
939
+ "gold": "321",
940
+ "correct": false
941
+ },
942
+ {
943
+ "idx": "aime24_20",
944
+ "pred": "",
945
+ "gold": "211",
946
+ "correct": false
947
+ },
948
+ {
949
+ "idx": "aime24_21",
950
+ "pred": "",
951
+ "gold": "315",
952
+ "correct": false
953
+ },
954
+ {
955
+ "idx": "aime24_22",
956
+ "pred": "",
957
+ "gold": "236",
958
+ "correct": false
959
+ },
960
+ {
961
+ "idx": "aime24_23",
962
+ "pred": "",
963
+ "gold": "045",
964
+ "correct": false
965
+ },
966
+ {
967
+ "idx": "aime24_24",
968
+ "pred": "",
969
+ "gold": "033",
970
+ "correct": false
971
+ },
972
+ {
973
+ "idx": "aime24_25",
974
+ "pred": "",
975
+ "gold": "080",
976
+ "correct": false
977
+ },
978
+ {
979
+ "idx": "aime24_26",
980
+ "pred": "",
981
+ "gold": "055",
982
+ "correct": false
983
+ },
984
+ {
985
+ "idx": "aime24_27",
986
+ "pred": "",
987
+ "gold": "699",
988
+ "correct": false
989
+ },
990
+ {
991
+ "idx": "aime24_28",
992
+ "pred": "",
993
+ "gold": "127",
994
+ "correct": false
995
+ },
996
+ {
997
+ "idx": "aime24_29",
998
+ "pred": "",
999
+ "gold": "902",
1000
+ "correct": false
1001
+ }
1002
+ ],
1003
+ "vs_baseline": {
1004
+ "baseline_accuracy": 0.03333333333333333,
1005
+ "steered_accuracy": 0.0,
1006
+ "absolute_drop": 0.03333333333333333,
1007
+ "relative_drop": 1.0,
1008
+ "n_common": 30,
1009
+ "n_regressions": 1,
1010
+ "n_recoveries": 0,
1011
+ "mcnemar_p_value": 1.0,
1012
+ "significant_at_0_05": false
1013
+ }
1014
+ }
1015
+ },
1016
+ "mon_alpha_0": {
1017
+ "MATH-500-holdout": {
1018
+ "accuracy": 0.32,
1019
+ "correct": 16,
1020
+ "n": 50,
1021
+ "per_sample": [
1022
+ {
1023
+ "idx": 450,
1024
+ "pred": "50",
1025
+ "gold": "64",
1026
+ "correct": false
1027
+ },
1028
+ {
1029
+ "idx": 451,
1030
+ "pred": "629",
1031
+ "gold": "\\frac{639}{40}",
1032
+ "correct": false
1033
+ },
1034
+ {
1035
+ "idx": 452,
1036
+ "pred": "85",
1037
+ "gold": "143",
1038
+ "correct": false
1039
+ },
1040
+ {
1041
+ "idx": 453,
1042
+ "pred": "10",
1043
+ "gold": "10",
1044
+ "correct": true
1045
+ },
1046
+ {
1047
+ "idx": 454,
1048
+ "pred": "3",
1049
+ "gold": "y = 2x + 3",
1050
+ "correct": false
1051
+ },
1052
+ {
1053
+ "idx": 455,
1054
+ "pred": "6",
1055
+ "gold": "6",
1056
+ "correct": true
1057
+ },
1058
+ {
1059
+ "idx": 456,
1060
+ "pred": "7",
1061
+ "gold": "-2,1",
1062
+ "correct": false
1063
+ },
1064
+ {
1065
+ "idx": 457,
1066
+ "pred": "0",
1067
+ "gold": "0",
1068
+ "correct": true
1069
+ },
1070
+ {
1071
+ "idx": 458,
1072
+ "pred": "41",
1073
+ "gold": "41",
1074
+ "correct": true
1075
+ },
1076
+ {
1077
+ "idx": 459,
1078
+ "pred": "6.80",
1079
+ "gold": "\\$18.90",
1080
+ "correct": false
1081
+ },
1082
+ {
1083
+ "idx": 460,
1084
+ "pred": "",
1085
+ "gold": "75^\\circ",
1086
+ "correct": false
1087
+ },
1088
+ {
1089
+ "idx": 461,
1090
+ "pred": "3",
1091
+ "gold": "12",
1092
+ "correct": false
1093
+ },
1094
+ {
1095
+ "idx": 462,
1096
+ "pred": "59",
1097
+ "gold": "59",
1098
+ "correct": true
1099
+ },
1100
+ {
1101
+ "idx": 463,
1102
+ "pred": "2",
1103
+ "gold": "2",
1104
+ "correct": true
1105
+ },
1106
+ {
1107
+ "idx": 464,
1108
+ "pred": "-4",
1109
+ "gold": "9",
1110
+ "correct": false
1111
+ },
1112
+ {
1113
+ "idx": 465,
1114
+ "pred": "2.",
1115
+ "gold": "121",
1116
+ "correct": false
1117
+ },
1118
+ {
1119
+ "idx": 466,
1120
+ "pred": "2",
1121
+ "gold": "8 \\pi",
1122
+ "correct": false
1123
+ },
1124
+ {
1125
+ "idx": 467,
1126
+ "pred": "8",
1127
+ "gold": "15\\mbox{ cm}^2",
1128
+ "correct": false
1129
+ },
1130
+ {
1131
+ "idx": 468,
1132
+ "pred": "8",
1133
+ "gold": "8",
1134
+ "correct": true
1135
+ },
1136
+ {
1137
+ "idx": 469,
1138
+ "pred": "25",
1139
+ "gold": "25",
1140
+ "correct": true
1141
+ },
1142
+ {
1143
+ "idx": 470,
1144
+ "pred": "1",
1145
+ "gold": "27648",
1146
+ "correct": false
1147
+ },
1148
+ {
1149
+ "idx": 471,
1150
+ "pred": "84",
1151
+ "gold": "84",
1152
+ "correct": true
1153
+ },
1154
+ {
1155
+ "idx": 472,
1156
+ "pred": "137\\frac{1",
1157
+ "gold": "137 \\frac{1}{2}",
1158
+ "correct": false
1159
+ },
1160
+ {
1161
+ "idx": 473,
1162
+ "pred": "1",
1163
+ "gold": "7",
1164
+ "correct": false
1165
+ },
1166
+ {
1167
+ "idx": 474,
1168
+ "pred": "2.",
1169
+ "gold": "3",
1170
+ "correct": false
1171
+ },
1172
+ {
1173
+ "idx": 475,
1174
+ "pred": "2",
1175
+ "gold": "29",
1176
+ "correct": false
1177
+ },
1178
+ {
1179
+ "idx": 476,
1180
+ "pred": "200",
1181
+ "gold": "200",
1182
+ "correct": true
1183
+ },
1184
+ {
1185
+ "idx": 477,
1186
+ "pred": "1",
1187
+ "gold": "2",
1188
+ "correct": false
1189
+ },
1190
+ {
1191
+ "idx": 478,
1192
+ "pred": "3",
1193
+ "gold": "\\frac{2}{1005}",
1194
+ "correct": false
1195
+ },
1196
+ {
1197
+ "idx": 479,
1198
+ "pred": "2",
1199
+ "gold": "2",
1200
+ "correct": true
1201
+ },
1202
+ {
1203
+ "idx": 480,
1204
+ "pred": "0",
1205
+ "gold": "1",
1206
+ "correct": false
1207
+ },
1208
+ {
1209
+ "idx": 481,
1210
+ "pred": "1",
1211
+ "gold": "8n^2 + 4n + 1",
1212
+ "correct": false
1213
+ },
1214
+ {
1215
+ "idx": 482,
1216
+ "pred": "2",
1217
+ "gold": "202",
1218
+ "correct": false
1219
+ },
1220
+ {
1221
+ "idx": 483,
1222
+ "pred": "16.",
1223
+ "gold": "\\frac{8}{15}",
1224
+ "correct": false
1225
+ },
1226
+ {
1227
+ "idx": 484,
1228
+ "pred": "54",
1229
+ "gold": "10",
1230
+ "correct": false
1231
+ },
1232
+ {
1233
+ "idx": 485,
1234
+ "pred": "5",
1235
+ "gold": "13",
1236
+ "correct": false
1237
+ },
1238
+ {
1239
+ "idx": 486,
1240
+ "pred": "216",
1241
+ "gold": "216",
1242
+ "correct": true
1243
+ },
1244
+ {
1245
+ "idx": 487,
1246
+ "pred": "2.",
1247
+ "gold": "2",
1248
+ "correct": true
1249
+ },
1250
+ {
1251
+ "idx": 488,
1252
+ "pred": "36",
1253
+ "gold": "\\$36",
1254
+ "correct": false
1255
+ },
1256
+ {
1257
+ "idx": 489,
1258
+ "pred": "15",
1259
+ "gold": "15",
1260
+ "correct": true
1261
+ },
1262
+ {
1263
+ "idx": 490,
1264
+ "pred": "2",
1265
+ "gold": "14",
1266
+ "correct": false
1267
+ },
1268
+ {
1269
+ "idx": 491,
1270
+ "pred": "64",
1271
+ "gold": "64",
1272
+ "correct": true
1273
+ },
1274
+ {
1275
+ "idx": 492,
1276
+ "pred": "0",
1277
+ "gold": "y = -2x",
1278
+ "correct": false
1279
+ },
1280
+ {
1281
+ "idx": 493,
1282
+ "pred": "20",
1283
+ "gold": "20",
1284
+ "correct": true
1285
+ },
1286
+ {
1287
+ "idx": 494,
1288
+ "pred": "-2",
1289
+ "gold": "\\frac{1}{4}",
1290
+ "correct": false
1291
+ },
1292
+ {
1293
+ "idx": 495,
1294
+ "pred": "3",
1295
+ "gold": "(2,12) \\cup (12,102)",
1296
+ "correct": false
1297
+ },
1298
+ {
1299
+ "idx": 496,
1300
+ "pred": "1",
1301
+ "gold": "\\frac{5}{13}",
1302
+ "correct": false
1303
+ },
1304
+ {
1305
+ "idx": 497,
1306
+ "pred": "0",
1307
+ "gold": "\\frac{7}{2}",
1308
+ "correct": false
1309
+ },
1310
+ {
1311
+ "idx": 498,
1312
+ "pred": "2",
1313
+ "gold": "-1",
1314
+ "correct": false
1315
+ },
1316
+ {
1317
+ "idx": 499,
1318
+ "pred": "106.",
1319
+ "gold": "106^\\circ",
1320
+ "correct": false
1321
+ }
1322
+ ],
1323
+ "vs_baseline": {
1324
+ "baseline_accuracy": 0.3,
1325
+ "steered_accuracy": 0.32,
1326
+ "absolute_drop": -0.020000000000000018,
1327
+ "relative_drop": -0.06666666666666674,
1328
+ "n_common": 50,
1329
+ "n_regressions": 4,
1330
+ "n_recoveries": 5,
1331
+ "mcnemar_p_value": 1.0,
1332
+ "significant_at_0_05": false
1333
+ }
1334
+ },
1335
+ "AIME-24": {
1336
+ "accuracy": 0.0,
1337
+ "correct": 0,
1338
+ "n": 30,
1339
+ "per_sample": [
1340
+ {
1341
+ "idx": "aime24_0",
1342
+ "pred": "60",
1343
+ "gold": "204",
1344
+ "correct": false
1345
+ },
1346
+ {
1347
+ "idx": "aime24_1",
1348
+ "pred": "9.",
1349
+ "gold": "113",
1350
+ "correct": false
1351
+ },
1352
+ {
1353
+ "idx": "aime24_2",
1354
+ "pred": "",
1355
+ "gold": "371",
1356
+ "correct": false
1357
+ },
1358
+ {
1359
+ "idx": "aime24_3",
1360
+ "pred": "2",
1361
+ "gold": "385",
1362
+ "correct": false
1363
+ },
1364
+ {
1365
+ "idx": "aime24_4",
1366
+ "pred": "3",
1367
+ "gold": "110",
1368
+ "correct": false
1369
+ },
1370
+ {
1371
+ "idx": "aime24_5",
1372
+ "pred": "4",
1373
+ "gold": "104",
1374
+ "correct": false
1375
+ },
1376
+ {
1377
+ "idx": "aime24_6",
1378
+ "pred": "3",
1379
+ "gold": "721",
1380
+ "correct": false
1381
+ },
1382
+ {
1383
+ "idx": "aime24_7",
1384
+ "pred": "5",
1385
+ "gold": "025",
1386
+ "correct": false
1387
+ },
1388
+ {
1389
+ "idx": "aime24_8",
1390
+ "pred": "1",
1391
+ "gold": "809",
1392
+ "correct": false
1393
+ },
1394
+ {
1395
+ "idx": "aime24_9",
1396
+ "pred": "1.",
1397
+ "gold": "116",
1398
+ "correct": false
1399
+ },
1400
+ {
1401
+ "idx": "aime24_10",
1402
+ "pred": "107",
1403
+ "gold": "104",
1404
+ "correct": false
1405
+ },
1406
+ {
1407
+ "idx": "aime24_11",
1408
+ "pred": "8",
1409
+ "gold": "294",
1410
+ "correct": false
1411
+ },
1412
+ {
1413
+ "idx": "aime24_12",
1414
+ "pred": "117",
1415
+ "gold": "540",
1416
+ "correct": false
1417
+ },
1418
+ {
1419
+ "idx": "aime24_13",
1420
+ "pred": "0",
1421
+ "gold": "197",
1422
+ "correct": false
1423
+ },
1424
+ {
1425
+ "idx": "aime24_14",
1426
+ "pred": "5",
1427
+ "gold": "480",
1428
+ "correct": false
1429
+ },
1430
+ {
1431
+ "idx": "aime24_15",
1432
+ "pred": "43",
1433
+ "gold": "073",
1434
+ "correct": false
1435
+ },
1436
+ {
1437
+ "idx": "aime24_16",
1438
+ "pred": "13",
1439
+ "gold": "468",
1440
+ "correct": false
1441
+ },
1442
+ {
1443
+ "idx": "aime24_17",
1444
+ "pred": "300.",
1445
+ "gold": "601",
1446
+ "correct": false
1447
+ },
1448
+ {
1449
+ "idx": "aime24_18",
1450
+ "pred": "1",
1451
+ "gold": "023",
1452
+ "correct": false
1453
+ },
1454
+ {
1455
+ "idx": "aime24_19",
1456
+ "pred": "-2",
1457
+ "gold": "321",
1458
+ "correct": false
1459
+ },
1460
+ {
1461
+ "idx": "aime24_20",
1462
+ "pred": "2",
1463
+ "gold": "211",
1464
+ "correct": false
1465
+ },
1466
+ {
1467
+ "idx": "aime24_21",
1468
+ "pred": "12",
1469
+ "gold": "315",
1470
+ "correct": false
1471
+ },
1472
+ {
1473
+ "idx": "aime24_22",
1474
+ "pred": "9",
1475
+ "gold": "236",
1476
+ "correct": false
1477
+ },
1478
+ {
1479
+ "idx": "aime24_23",
1480
+ "pred": "2",
1481
+ "gold": "045",
1482
+ "correct": false
1483
+ },
1484
+ {
1485
+ "idx": "aime24_24",
1486
+ "pred": "is log base 2 too), this substitution should simplify things because all the logarithmic expressions in the equations will become linear combinations of a, b, c. Let's check each equation one by one",
1487
+ "gold": "033",
1488
+ "correct": false
1489
+ },
1490
+ {
1491
+ "idx": "aime24_25",
1492
+ "pred": "2",
1493
+ "gold": "080",
1494
+ "correct": false
1495
+ },
1496
+ {
1497
+ "idx": "aime24_26",
1498
+ "pred": "2",
1499
+ "gold": "055",
1500
+ "correct": false
1501
+ },
1502
+ {
1503
+ "idx": "aime24_27",
1504
+ "pred": "3",
1505
+ "gold": "699",
1506
+ "correct": false
1507
+ },
1508
+ {
1509
+ "idx": "aime24_28",
1510
+ "pred": "9",
1511
+ "gold": "127",
1512
+ "correct": false
1513
+ },
1514
+ {
1515
+ "idx": "aime24_29",
1516
+ "pred": "",
1517
+ "gold": "902",
1518
+ "correct": false
1519
+ }
1520
+ ],
1521
+ "vs_baseline": {
1522
+ "baseline_accuracy": 0.03333333333333333,
1523
+ "steered_accuracy": 0.0,
1524
+ "absolute_drop": 0.03333333333333333,
1525
+ "relative_drop": 1.0,
1526
+ "n_common": 30,
1527
+ "n_regressions": 1,
1528
+ "n_recoveries": 0,
1529
+ "mcnemar_p_value": 1.0,
1530
+ "significant_at_0_05": false
1531
+ }
1532
+ }
1533
+ }
1534
+ }
results/final_report.md ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Student Simulation — Final Report
2
+
3
+ ## 1. Overview
4
+
5
+ Model: Qwen3-30B-A3B-Thinking-2507
6
+
7
+ ## 2. Dimension Interaction
8
+
9
+ - Jaccard overlap of top-K experts: **0.000**
10
+ - # PMI pairs (same-layer plan-mon): 19
11
+ - PMI mean: -4.670 (positive = experts tend to co-activate)
12
+
13
+ ## 3. Steering Sweep Results
14
+
15
+ > **NEW SEMANTICS**: α=1.0 is baseline (no steering). Lower α = stronger suppression. α>1 = amplification.
16
+
17
+ > A WORKING steering shows: max RR at LOW α (e.g. α=0), and Spearman ρ(α, RR) **negative** (≤ -0.5).
18
+
19
+ ### 3.2 Monitoring
20
+
21
+ | version | best α (max RR) | max RR | ΔPQS at best α | Spearman ρ(α, RR) | monotonic? | collapse% |
22
+ |---|---|---|---|---|---|---|
23
+ | v1_raw | +1.50 | +1.000 | -0.058 | +0.563 | no | 0.0% |
24
+ | v2_ortho_general | +1.50 | +1.000 | -0.058 | +0.323 | no | 100.0% |
25
+ | v3_ortho_crossdim | +1.50 | +1.000 | -0.058 | +0.323 | no | 100.0% |
26
+ | v4_pca | +1.50 | +1.000 | -0.058 | +0.635 | no | 83.3% |
27
+
28
+ **Force-prompt baseline**:
29
+ - force_suppress: RR=+0.060, ΔPQS=+0.007
30
+ - force_enhance: RR=-0.302, ΔPQS=-0.009
31
+
32
+ ### 3.1 Planning
33
+
34
+ | version | best α (max RR) | max RR | ΔPQS at best α | Spearman ρ(α, RR) | monotonic? | collapse% |
35
+ |---|---|---|---|---|---|---|
36
+ | v1_raw | +1.50 | +0.833 | -0.039 | +0.252 | no | 100.0% |
37
+ | v2_ortho_general | +1.50 | +0.833 | -0.039 | +0.371 | no | 100.0% |
38
+ | v3_ortho_crossdim | +1.50 | +0.833 | -0.039 | +0.204 | no | 100.0% |
39
+ | v4_pca | +1.50 | +0.833 | -0.039 | +0.275 | no | 100.0% |
40
+
41
+ **Force-prompt baseline**:
42
+ - force_suppress: RR=+0.004, ΔPQS=+0.011
43
+ - force_enhance: RR=-0.177, ΔPQS=+0.008
44
+
45
+ ## 4. Diagnosis: RR vs PQS (Planning, v3_ortho_crossdim)
46
+
47
+ > Under NEW semantics: α=1 is baseline. Steering effects should appear as α decreases below 1.
48
+
49
+ | α | meaning | RR | ΔPQS | Interpretation |
50
+ |---|---|---|---|---|
51
+ | -0.50 | over-suppression | +0.239 | -0.060 | ✅ capability suppression |
52
+ | +0.00 | 100% suppression | +0.191 | -0.061 | ? |
53
+ | +0.25 | 75% suppression | +0.012 | -0.059 | — no effect |
54
+ | +0.50 | 50% suppression | +0.096 | -0.037 | — no effect |
55
+ | +0.75 | 25% suppression | -0.048 | -0.020 | — no effect |
56
+ | +1.00 | baseline | +0.000 | +0.000 | (reference) |
57
+ | +1.50 | amplification | +0.833 | -0.039 | ⚠️ surface-only (RR drops but PQS unchanged) |
58
+ | +2.00 | amplification | +0.833 | -0.039 | ⚠️ surface-only (RR drops but PQS unchanged) |
59
+
60
+ ## 5. Downstream Accuracy
61
+
62
+ ### 5.1 Raw accuracies
63
+
64
+ | config | MATH-500-holdout | AIME-24 |
65
+ |---|---|---|
66
+ | baseline | 0.300 | 0.033 |
67
+ | plan_alpha_0 | 0.040 | 0.000 |
68
+ | mon_alpha_0 | 0.320 | 0.000 |
69
+
70
+ ### 5.2 Accuracy drop vs baseline
71
+
72
+ > **absolute drop** = baseline_acc − steered_acc (positive = WORSE under steering)
73
+ > **relative drop** = absolute_drop / baseline_acc
74
+ > **McNemar p** = paired-test p-value on per-problem correctness
75
+ > **regr/rec** = #problems where baseline was right→steered wrong / vice versa
76
+
77
+ | config | testset | baseline | steered | Δ abs | Δ rel | regr/rec | McNemar p | sig p<0.05 |
78
+ |---|---|---|---|---|---|---|---|---|
79
+ | plan_alpha_0 | MATH-500-holdout | 0.300 | 0.040 | +0.260 | +86.7% | 13/0 | 0.000244 | ✅ |
80
+ | plan_alpha_0 | AIME-24 | 0.033 | 0.000 | +0.033 | +100.0% | 1/0 | 1 | — |
81
+ | mon_alpha_0 | MATH-500-holdout | 0.300 | 0.320 | -0.020 | -6.7% | 4/5 | 1 | — |
82
+ | mon_alpha_0 | AIME-24 | 0.033 | 0.000 | +0.033 | +100.0% | 1/0 | 1 | — |
83
+
84
+ ## 6. Go/No-Go Decision
85
+
86
+ > Decision criteria under NEW semantics:
87
+ > - WORKING: max RR > 0.3 AT α < 1 AND Spearman ρ(α, RR) ≤ -0.5
88
+ > - For planning, additionally need ΔPQS < -0.05 at the best α.
89
+
90
+ **[NO-GO]** Neither dimension shows clean steering. Reconsider methodology or scale.
results/interaction_summary.json ADDED
@@ -0,0 +1,205 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "jaccard_overlap": 0.0,
3
+ "n_plan_experts": 32,
4
+ "n_mon_experts": 32,
5
+ "intersection": [],
6
+ "n_pmi_pairs": 19,
7
+ "pmi_pairs": [
8
+ {
9
+ "plan_layer": 20,
10
+ "plan_expert": 101,
11
+ "mon_layer": 20,
12
+ "mon_expert": 118,
13
+ "pmi": 0.1955346800284106,
14
+ "P_plan": 0.8223948223938224,
15
+ "P_mon": 1e-06,
16
+ "P_joint": 1e-06
17
+ },
18
+ {
19
+ "plan_layer": 24,
20
+ "plan_expert": 21,
21
+ "mon_layer": 24,
22
+ "mon_expert": 63,
23
+ "pmi": -8.924518573908871,
24
+ "P_plan": 0.6486496486486487,
25
+ "P_mon": 0.011584011583011582,
26
+ "P_joint": 1e-06
27
+ },
28
+ {
29
+ "plan_layer": 37,
30
+ "plan_expert": 78,
31
+ "mon_layer": 37,
32
+ "mon_expert": 84,
33
+ "pmi": -8.976704248689728,
34
+ "P_plan": 0.6833986833976834,
35
+ "P_mon": 0.011584011583011582,
36
+ "P_joint": 1e-06
37
+ },
38
+ {
39
+ "plan_layer": 37,
40
+ "plan_expert": 78,
41
+ "mon_layer": 37,
42
+ "mon_expert": 73,
43
+ "pmi": -9.48749534150732,
44
+ "P_plan": 0.6833986833976834,
45
+ "P_mon": 0.019306019305019306,
46
+ "P_joint": 1e-06
47
+ },
48
+ {
49
+ "plan_layer": 8,
50
+ "plan_expert": 111,
51
+ "mon_layer": 8,
52
+ "mon_expert": 118,
53
+ "pmi": -8.438557894626335,
54
+ "P_plan": 0.5984565984555985,
55
+ "P_mon": 0.007723007722007722,
56
+ "P_joint": 1e-06
57
+ },
58
+ {
59
+ "plan_layer": 13,
60
+ "plan_expert": 58,
61
+ "mon_layer": 13,
62
+ "mon_expert": 84,
63
+ "pmi": -8.965340506762553,
64
+ "P_plan": 0.6756766756756757,
65
+ "P_mon": 0.011584011583011582,
66
+ "P_joint": 1e-06
67
+ },
68
+ {
69
+ "plan_layer": 13,
70
+ "plan_expert": 58,
71
+ "mon_layer": 13,
72
+ "mon_expert": 101,
73
+ "pmi": -8.965340506762553,
74
+ "P_plan": 0.6756766756756757,
75
+ "P_mon": 0.011584011583011582,
76
+ "P_joint": 1e-06
77
+ },
78
+ {
79
+ "plan_layer": 10,
80
+ "plan_expert": 58,
81
+ "mon_layer": 10,
82
+ "mon_expert": 67,
83
+ "pmi": 0.6080663344436239,
84
+ "P_plan": 0.5444025444015445,
85
+ "P_mon": 1e-06,
86
+ "P_joint": 1e-06
87
+ },
88
+ {
89
+ "plan_layer": 36,
90
+ "plan_expert": 65,
91
+ "mon_layer": 36,
92
+ "mon_expert": 75,
93
+ "pmi": -10.745533672853218,
94
+ "P_plan": 0.667954667953668,
95
+ "P_mon": 0.0694990694980695,
96
+ "P_joint": 1e-06
97
+ },
98
+ {
99
+ "plan_layer": 3,
100
+ "plan_expert": 57,
101
+ "mon_layer": 3,
102
+ "mon_expert": 59,
103
+ "pmi": -11.072875128037278,
104
+ "P_plan": 0.5212365212355212,
105
+ "P_mon": 0.12355312355212356,
106
+ "P_joint": 1e-06
107
+ },
108
+ {
109
+ "plan_layer": 38,
110
+ "plan_expert": 117,
111
+ "mon_layer": 38,
112
+ "mon_expert": 22,
113
+ "pmi": 0.38634259507155955,
114
+ "P_plan": 0.6795376795366795,
115
+ "P_mon": 1e-06,
116
+ "P_joint": 1e-06
117
+ },
118
+ {
119
+ "plan_layer": 27,
120
+ "plan_expert": 49,
121
+ "mon_layer": 27,
122
+ "mon_expert": 51,
123
+ "pmi": 0.5069703941949687,
124
+ "P_plan": 0.6023176023166024,
125
+ "P_mon": 1e-06,
126
+ "P_joint": 1e-06
127
+ },
128
+ {
129
+ "plan_layer": 0,
130
+ "plan_expert": 84,
131
+ "mon_layer": 0,
132
+ "mon_expert": 124,
133
+ "pmi": -8.971038519468111,
134
+ "P_plan": 0.6795376795366795,
135
+ "P_mon": 0.011584011583011582,
136
+ "P_joint": 1e-06
137
+ },
138
+ {
139
+ "plan_layer": 0,
140
+ "plan_expert": 84,
141
+ "mon_layer": 0,
142
+ "mon_expert": 3,
143
+ "pmi": 0.38634259507155955,
144
+ "P_plan": 0.6795376795366795,
145
+ "P_mon": 1e-06,
146
+ "P_joint": 1e-06
147
+ },
148
+ {
149
+ "plan_layer": 10,
150
+ "plan_expert": 57,
151
+ "mon_layer": 10,
152
+ "mon_expert": 67,
153
+ "pmi": 0.3977712739803695,
154
+ "P_plan": 0.6718156718146718,
155
+ "P_mon": 1e-06,
156
+ "P_joint": 1e-06
157
+ },
158
+ {
159
+ "plan_layer": 21,
160
+ "plan_expert": 87,
161
+ "mon_layer": 21,
162
+ "mon_expert": 127,
163
+ "pmi": -7.872598867658467,
164
+ "P_plan": 0.6795376795366795,
165
+ "P_mon": 0.0038620038610038612,
166
+ "P_joint": 1e-06
167
+ },
168
+ {
169
+ "plan_layer": 17,
170
+ "plan_expert": 83,
171
+ "mon_layer": 17,
172
+ "mon_expert": 54,
173
+ "pmi": 0.3045532776328849,
174
+ "P_plan": 0.7374527374517375,
175
+ "P_mon": 0.0038620038610038612,
176
+ "P_joint": 0.0038620038610038612
177
+ },
178
+ {
179
+ "plan_layer": 14,
180
+ "plan_expert": 104,
181
+ "mon_layer": 14,
182
+ "mon_expert": 22,
183
+ "pmi": 0.5069703941949687,
184
+ "P_plan": 0.6023176023166024,
185
+ "P_mon": 1e-06,
186
+ "P_joint": 1e-06
187
+ },
188
+ {
189
+ "plan_layer": 15,
190
+ "plan_expert": 49,
191
+ "mon_layer": 15,
192
+ "mon_expert": 51,
193
+ "pmi": 0.3977712739803695,
194
+ "P_plan": 0.6718156718146718,
195
+ "P_mon": 1e-06,
196
+ "P_joint": 1e-06
197
+ }
198
+ ],
199
+ "pmi_stats": {
200
+ "mean": -4.669983181140828,
201
+ "std": 4.866646746187251,
202
+ "max": 0.6080663344436239,
203
+ "min": -11.072875128037278
204
+ }
205
+ }
results/sweep_log.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
results/sweep_rr.json ADDED
@@ -0,0 +1,778 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "planning": {
3
+ "v1_raw": {
4
+ "-0.5": {
5
+ "n": 30,
6
+ "rr_mean": 0.5577777777777778,
7
+ "rr_std": 0.41942935955214317,
8
+ "pqs_steered_mean": 0.21859502553354465,
9
+ "pqs_base_mean": 0.314297770303943,
10
+ "pqs_delta_mean": -0.09570274477039836,
11
+ "collapse_rate": 1.0,
12
+ "steered_plan_mean": 0.5333333333333333,
13
+ "steered_mon_mean": 0.5
14
+ },
15
+ "0.0": {
16
+ "n": 30,
17
+ "rr_mean": 0.24666666666666667,
18
+ "rr_std": 0.5464938751308706,
19
+ "pqs_steered_mean": 0.22655125564239345,
20
+ "pqs_base_mean": 0.314297770303943,
21
+ "pqs_delta_mean": -0.08774651466154956,
22
+ "collapse_rate": 1.0,
23
+ "steered_plan_mean": 1.0666666666666667,
24
+ "steered_mon_mean": 1.8666666666666667
25
+ },
26
+ "0.25": {
27
+ "n": 30,
28
+ "rr_mean": 0.06888888888888887,
29
+ "rr_std": 0.7950277578817416,
30
+ "pqs_steered_mean": 0.2680342621642518,
31
+ "pqs_base_mean": 0.314297770303943,
32
+ "pqs_delta_mean": -0.04626350813969121,
33
+ "collapse_rate": 0.9333333333333333,
34
+ "steered_plan_mean": 1.2666666666666666,
35
+ "steered_mon_mean": 20.1
36
+ },
37
+ "0.5": {
38
+ "n": 30,
39
+ "rr_mean": 0.31611111111111106,
40
+ "rr_std": 0.5530468356921824,
41
+ "pqs_steered_mean": 0.30100228859063266,
42
+ "pqs_base_mean": 0.314297770303943,
43
+ "pqs_delta_mean": -0.013295481713310353,
44
+ "collapse_rate": 0.9666666666666667,
45
+ "steered_plan_mean": 1.0666666666666667,
46
+ "steered_mon_mean": 10.7
47
+ },
48
+ "0.75": {
49
+ "n": 30,
50
+ "rr_mean": -0.08166666666666665,
51
+ "rr_std": 0.966305967187527,
52
+ "pqs_steered_mean": 0.29591473252953365,
53
+ "pqs_base_mean": 0.314297770303943,
54
+ "pqs_delta_mean": -0.01838303777440936,
55
+ "collapse_rate": 1.0,
56
+ "steered_plan_mean": 1.6333333333333333,
57
+ "steered_mon_mean": 7.8
58
+ },
59
+ "1.0": {
60
+ "n": 30,
61
+ "rr_mean": 0.0,
62
+ "rr_std": 0.0,
63
+ "pqs_steered_mean": 0.314297770303943,
64
+ "pqs_base_mean": 0.314297770303943,
65
+ "pqs_delta_mean": 0.0,
66
+ "collapse_rate": 0.9666666666666667,
67
+ "steered_plan_mean": 1.6,
68
+ "steered_mon_mean": 7.633333333333334
69
+ },
70
+ "1.5": {
71
+ "n": 30,
72
+ "rr_mean": 0.8333333333333334,
73
+ "rr_std": 0.3726779962499649,
74
+ "pqs_steered_mean": 0.2750000000000001,
75
+ "pqs_base_mean": 0.314297770303943,
76
+ "pqs_delta_mean": -0.039297770303942936,
77
+ "collapse_rate": 1.0,
78
+ "steered_plan_mean": 0.0,
79
+ "steered_mon_mean": 0.0
80
+ },
81
+ "2.0": {
82
+ "n": 30,
83
+ "rr_mean": 0.8333333333333334,
84
+ "rr_std": 0.3726779962499649,
85
+ "pqs_steered_mean": 0.2750000000000001,
86
+ "pqs_base_mean": 0.314297770303943,
87
+ "pqs_delta_mean": -0.039297770303942936,
88
+ "collapse_rate": 1.0,
89
+ "steered_plan_mean": 0.0,
90
+ "steered_mon_mean": 0.0
91
+ }
92
+ },
93
+ "v2_ortho_general": {
94
+ "-0.5": {
95
+ "n": 30,
96
+ "rr_mean": 0.13944444444444445,
97
+ "rr_std": 0.705933625740138,
98
+ "pqs_steered_mean": 0.2738495468254436,
99
+ "pqs_base_mean": 0.314297770303943,
100
+ "pqs_delta_mean": -0.04044822347849941,
101
+ "collapse_rate": 0.9666666666666667,
102
+ "steered_plan_mean": 1.5,
103
+ "steered_mon_mean": 97.66666666666667
104
+ },
105
+ "0.0": {
106
+ "n": 30,
107
+ "rr_mean": 0.053888888888888875,
108
+ "rr_std": 0.7668077567438067,
109
+ "pqs_steered_mean": 0.23186373309709302,
110
+ "pqs_base_mean": 0.314297770303943,
111
+ "pqs_delta_mean": -0.08243403720685,
112
+ "collapse_rate": 1.0,
113
+ "steered_plan_mean": 4.0,
114
+ "steered_mon_mean": 81.8
115
+ },
116
+ "0.25": {
117
+ "n": 30,
118
+ "rr_mean": 0.15833333333333333,
119
+ "rr_std": 0.5269540355244995,
120
+ "pqs_steered_mean": 0.23891818550974483,
121
+ "pqs_base_mean": 0.314297770303943,
122
+ "pqs_delta_mean": -0.07537958479419818,
123
+ "collapse_rate": 1.0,
124
+ "steered_plan_mean": 1.3,
125
+ "steered_mon_mean": 98.06666666666666
126
+ },
127
+ "0.5": {
128
+ "n": 30,
129
+ "rr_mean": 0.145,
130
+ "rr_std": 0.6923811570707761,
131
+ "pqs_steered_mean": 0.2538471343251946,
132
+ "pqs_base_mean": 0.314297770303943,
133
+ "pqs_delta_mean": -0.06045063597874839,
134
+ "collapse_rate": 1.0,
135
+ "steered_plan_mean": 1.2666666666666666,
136
+ "steered_mon_mean": 74.0
137
+ },
138
+ "0.75": {
139
+ "n": 30,
140
+ "rr_mean": -0.10944444444444444,
141
+ "rr_std": 0.8343514151426947,
142
+ "pqs_steered_mean": 0.2919660669034674,
143
+ "pqs_base_mean": 0.314297770303943,
144
+ "pqs_delta_mean": -0.022331703400475633,
145
+ "collapse_rate": 1.0,
146
+ "steered_plan_mean": 1.5333333333333334,
147
+ "steered_mon_mean": 8.4
148
+ },
149
+ "1.0": {
150
+ "n": 30,
151
+ "rr_mean": 0.0,
152
+ "rr_std": 0.0,
153
+ "pqs_steered_mean": 0.314297770303943,
154
+ "pqs_base_mean": 0.314297770303943,
155
+ "pqs_delta_mean": 0.0,
156
+ "collapse_rate": 0.9666666666666667,
157
+ "steered_plan_mean": 1.6,
158
+ "steered_mon_mean": 7.633333333333334
159
+ },
160
+ "1.5": {
161
+ "n": 30,
162
+ "rr_mean": 0.8333333333333334,
163
+ "rr_std": 0.3726779962499649,
164
+ "pqs_steered_mean": 0.2750000000000001,
165
+ "pqs_base_mean": 0.314297770303943,
166
+ "pqs_delta_mean": -0.039297770303942936,
167
+ "collapse_rate": 1.0,
168
+ "steered_plan_mean": 0.0,
169
+ "steered_mon_mean": 0.0
170
+ },
171
+ "2.0": {
172
+ "n": 30,
173
+ "rr_mean": 0.8333333333333334,
174
+ "rr_std": 0.3726779962499649,
175
+ "pqs_steered_mean": 0.2750000000000001,
176
+ "pqs_base_mean": 0.314297770303943,
177
+ "pqs_delta_mean": -0.039297770303942936,
178
+ "collapse_rate": 0.0,
179
+ "steered_plan_mean": 0.0,
180
+ "steered_mon_mean": 0.0
181
+ }
182
+ },
183
+ "v3_ortho_crossdim": {
184
+ "-0.5": {
185
+ "n": 30,
186
+ "rr_mean": 0.23944444444444443,
187
+ "rr_std": 0.6276128101832951,
188
+ "pqs_steered_mean": 0.25385654164190363,
189
+ "pqs_base_mean": 0.314297770303943,
190
+ "pqs_delta_mean": -0.06044122866203938,
191
+ "collapse_rate": 0.9666666666666667,
192
+ "steered_plan_mean": 1.2,
193
+ "steered_mon_mean": 80.53333333333333
194
+ },
195
+ "0.0": {
196
+ "n": 30,
197
+ "rr_mean": 0.19111111111111112,
198
+ "rr_std": 0.5631875770234818,
199
+ "pqs_steered_mean": 0.25325916784857466,
200
+ "pqs_base_mean": 0.314297770303943,
201
+ "pqs_delta_mean": -0.06103860245536835,
202
+ "collapse_rate": 1.0,
203
+ "steered_plan_mean": 1.1333333333333333,
204
+ "steered_mon_mean": 101.23333333333333
205
+ },
206
+ "0.25": {
207
+ "n": 30,
208
+ "rr_mean": 0.011666666666666665,
209
+ "rr_std": 0.6236563334258878,
210
+ "pqs_steered_mean": 0.2554860654101204,
211
+ "pqs_base_mean": 0.314297770303943,
212
+ "pqs_delta_mean": -0.05881170489382259,
213
+ "collapse_rate": 1.0,
214
+ "steered_plan_mean": 1.5,
215
+ "steered_mon_mean": 73.73333333333333
216
+ },
217
+ "0.5": {
218
+ "n": 30,
219
+ "rr_mean": 0.09611111111111109,
220
+ "rr_std": 0.6531077180358725,
221
+ "pqs_steered_mean": 0.2776005368545984,
222
+ "pqs_base_mean": 0.314297770303943,
223
+ "pqs_delta_mean": -0.03669723344934461,
224
+ "collapse_rate": 1.0,
225
+ "steered_plan_mean": 1.5333333333333334,
226
+ "steered_mon_mean": 35.266666666666666
227
+ },
228
+ "0.75": {
229
+ "n": 30,
230
+ "rr_mean": -0.04833333333333333,
231
+ "rr_std": 0.6618765410222329,
232
+ "pqs_steered_mean": 0.29478959140366856,
233
+ "pqs_base_mean": 0.314297770303943,
234
+ "pqs_delta_mean": -0.019508178900274453,
235
+ "collapse_rate": 1.0,
236
+ "steered_plan_mean": 1.6333333333333333,
237
+ "steered_mon_mean": 7.733333333333333
238
+ },
239
+ "1.0": {
240
+ "n": 30,
241
+ "rr_mean": 0.0,
242
+ "rr_std": 0.0,
243
+ "pqs_steered_mean": 0.314297770303943,
244
+ "pqs_base_mean": 0.314297770303943,
245
+ "pqs_delta_mean": 0.0,
246
+ "collapse_rate": 0.9666666666666667,
247
+ "steered_plan_mean": 1.6,
248
+ "steered_mon_mean": 7.633333333333334
249
+ },
250
+ "1.5": {
251
+ "n": 30,
252
+ "rr_mean": 0.8333333333333334,
253
+ "rr_std": 0.3726779962499649,
254
+ "pqs_steered_mean": 0.2750000000000001,
255
+ "pqs_base_mean": 0.314297770303943,
256
+ "pqs_delta_mean": -0.039297770303942936,
257
+ "collapse_rate": 1.0,
258
+ "steered_plan_mean": 0.0,
259
+ "steered_mon_mean": 0.0
260
+ },
261
+ "2.0": {
262
+ "n": 30,
263
+ "rr_mean": 0.8333333333333334,
264
+ "rr_std": 0.3726779962499649,
265
+ "pqs_steered_mean": 0.2750000000000001,
266
+ "pqs_base_mean": 0.314297770303943,
267
+ "pqs_delta_mean": -0.039297770303942936,
268
+ "collapse_rate": 1.0,
269
+ "steered_plan_mean": 0.0,
270
+ "steered_mon_mean": 0.0
271
+ }
272
+ },
273
+ "v4_pca": {
274
+ "-0.5": {
275
+ "n": 30,
276
+ "rr_mean": 0.2372222222222222,
277
+ "rr_std": 0.657149897401134,
278
+ "pqs_steered_mean": 0.27847033214756994,
279
+ "pqs_base_mean": 0.314297770303943,
280
+ "pqs_delta_mean": -0.03582743815637307,
281
+ "collapse_rate": 0.8666666666666667,
282
+ "steered_plan_mean": 1.3666666666666667,
283
+ "steered_mon_mean": 62.53333333333333
284
+ },
285
+ "0.0": {
286
+ "n": 30,
287
+ "rr_mean": 0.008333333333333333,
288
+ "rr_std": 0.8282595540583211,
289
+ "pqs_steered_mean": 0.3055106488837709,
290
+ "pqs_base_mean": 0.314297770303943,
291
+ "pqs_delta_mean": -0.008787121420172128,
292
+ "collapse_rate": 0.9666666666666667,
293
+ "steered_plan_mean": 1.3666666666666667,
294
+ "steered_mon_mean": 50.3
295
+ },
296
+ "0.25": {
297
+ "n": 30,
298
+ "rr_mean": 0.13722222222222225,
299
+ "rr_std": 0.7600156349918723,
300
+ "pqs_steered_mean": 0.2986814542200189,
301
+ "pqs_base_mean": 0.314297770303943,
302
+ "pqs_delta_mean": -0.015616316083924109,
303
+ "collapse_rate": 1.0,
304
+ "steered_plan_mean": 1.2333333333333334,
305
+ "steered_mon_mean": 73.3
306
+ },
307
+ "0.5": {
308
+ "n": 30,
309
+ "rr_mean": -0.2011111111111111,
310
+ "rr_std": 1.1317058030651597,
311
+ "pqs_steered_mean": 0.29877299566269305,
312
+ "pqs_base_mean": 0.314297770303943,
313
+ "pqs_delta_mean": -0.015524774641249961,
314
+ "collapse_rate": 1.0,
315
+ "steered_plan_mean": 1.7333333333333334,
316
+ "steered_mon_mean": 39.0
317
+ },
318
+ "0.75": {
319
+ "n": 30,
320
+ "rr_mean": -0.08277777777777778,
321
+ "rr_std": 0.9574511237243329,
322
+ "pqs_steered_mean": 0.32578590443484834,
323
+ "pqs_base_mean": 0.314297770303943,
324
+ "pqs_delta_mean": 0.011488134130905325,
325
+ "collapse_rate": 1.0,
326
+ "steered_plan_mean": 1.7333333333333334,
327
+ "steered_mon_mean": 10.433333333333334
328
+ },
329
+ "1.0": {
330
+ "n": 30,
331
+ "rr_mean": 0.0,
332
+ "rr_std": 0.0,
333
+ "pqs_steered_mean": 0.314297770303943,
334
+ "pqs_base_mean": 0.314297770303943,
335
+ "pqs_delta_mean": 0.0,
336
+ "collapse_rate": 0.9666666666666667,
337
+ "steered_plan_mean": 1.6,
338
+ "steered_mon_mean": 7.633333333333334
339
+ },
340
+ "1.5": {
341
+ "n": 30,
342
+ "rr_mean": 0.8333333333333334,
343
+ "rr_std": 0.3726779962499649,
344
+ "pqs_steered_mean": 0.2750000000000001,
345
+ "pqs_base_mean": 0.314297770303943,
346
+ "pqs_delta_mean": -0.039297770303942936,
347
+ "collapse_rate": 1.0,
348
+ "steered_plan_mean": 0.0,
349
+ "steered_mon_mean": 0.0
350
+ },
351
+ "2.0": {
352
+ "n": 30,
353
+ "rr_mean": 0.8333333333333334,
354
+ "rr_std": 0.3726779962499649,
355
+ "pqs_steered_mean": 0.2750000000000001,
356
+ "pqs_base_mean": 0.314297770303943,
357
+ "pqs_delta_mean": -0.039297770303942936,
358
+ "collapse_rate": 1.0,
359
+ "steered_plan_mean": 0.0,
360
+ "steered_mon_mean": 0.0
361
+ }
362
+ },
363
+ "force_suppress": {
364
+ "None": {
365
+ "n": 30,
366
+ "rr_mean": 0.0044444444444444375,
367
+ "rr_std": 0.7359918142325252,
368
+ "pqs_steered_mean": 0.3191702739101362,
369
+ "pqs_base_mean": 0.30841244078579816,
370
+ "pqs_delta_mean": 0.010757833124338034,
371
+ "collapse_rate": 1.0,
372
+ "steered_plan_mean": 1.3,
373
+ "steered_mon_mean": 7.6
374
+ }
375
+ },
376
+ "force_enhance": {
377
+ "None": {
378
+ "n": 30,
379
+ "rr_mean": -0.17722222222222223,
380
+ "rr_std": 0.8478129088395987,
381
+ "pqs_steered_mean": 0.31601437824776085,
382
+ "pqs_base_mean": 0.30841244078579816,
383
+ "pqs_delta_mean": 0.007601937461962693,
384
+ "collapse_rate": 1.0,
385
+ "steered_plan_mean": 1.6,
386
+ "steered_mon_mean": 9.8
387
+ }
388
+ }
389
+ },
390
+ "monitoring": {
391
+ "v1_raw": {
392
+ "-0.5": {
393
+ "n": 30,
394
+ "rr_mean": 0.16279128823246466,
395
+ "rr_std": 1.465739085419085,
396
+ "pqs_steered_mean": 0.2918421726094732,
397
+ "pqs_base_mean": 0.314297770303943,
398
+ "pqs_delta_mean": -0.02245559769446981,
399
+ "collapse_rate": 0.9666666666666667,
400
+ "steered_plan_mean": 0.9666666666666667,
401
+ "steered_mon_mean": 4.633333333333334
402
+ },
403
+ "0.0": {
404
+ "n": 30,
405
+ "rr_mean": -0.7991717977012095,
406
+ "rr_std": 1.823646029909259,
407
+ "pqs_steered_mean": 0.29817680418181514,
408
+ "pqs_base_mean": 0.314297770303943,
409
+ "pqs_delta_mean": -0.016120966122127878,
410
+ "collapse_rate": 1.0,
411
+ "steered_plan_mean": 1.0333333333333334,
412
+ "steered_mon_mean": 9.633333333333333
413
+ },
414
+ "0.25": {
415
+ "n": 30,
416
+ "rr_mean": -0.5264151861210684,
417
+ "rr_std": 1.1735454740120077,
418
+ "pqs_steered_mean": 0.30254814643713535,
419
+ "pqs_base_mean": 0.314297770303943,
420
+ "pqs_delta_mean": -0.011749623866807668,
421
+ "collapse_rate": 1.0,
422
+ "steered_plan_mean": 1.1666666666666667,
423
+ "steered_mon_mean": 9.766666666666667
424
+ },
425
+ "0.5": {
426
+ "n": 30,
427
+ "rr_mean": -3.5288139329805994,
428
+ "rr_std": 14.23810114972099,
429
+ "pqs_steered_mean": 0.307789521694332,
430
+ "pqs_base_mean": 0.30679878932608795,
431
+ "pqs_delta_mean": 0.0009907323682440405,
432
+ "collapse_rate": 1.0,
433
+ "steered_plan_mean": 1.0666666666666667,
434
+ "steered_mon_mean": 33.9
435
+ },
436
+ "0.75": {
437
+ "n": 30,
438
+ "rr_mean": -0.47021524771524775,
439
+ "rr_std": 1.0970956604181485,
440
+ "pqs_steered_mean": 0.32301581512712546,
441
+ "pqs_base_mean": 0.30841244078579816,
442
+ "pqs_delta_mean": 0.014603374341327302,
443
+ "collapse_rate": 1.0,
444
+ "steered_plan_mean": 1.1666666666666667,
445
+ "steered_mon_mean": 10.4
446
+ },
447
+ "1.0": {
448
+ "n": 30,
449
+ "rr_mean": 0.0,
450
+ "rr_std": 0.0,
451
+ "pqs_steered_mean": 0.30841244078579816,
452
+ "pqs_base_mean": 0.30841244078579816,
453
+ "pqs_delta_mean": 0.0,
454
+ "collapse_rate": 1.0,
455
+ "steered_plan_mean": 1.3666666666666667,
456
+ "steered_mon_mean": 8.833333333333334
457
+ },
458
+ "1.5": {
459
+ "n": 30,
460
+ "rr_mean": 1.0,
461
+ "rr_std": 0.0,
462
+ "pqs_steered_mean": 0.25,
463
+ "pqs_base_mean": 0.30841244078579816,
464
+ "pqs_delta_mean": -0.05841244078579816,
465
+ "collapse_rate": 0.0,
466
+ "steered_plan_mean": 0.0,
467
+ "steered_mon_mean": 0.0
468
+ },
469
+ "2.0": {
470
+ "n": 30,
471
+ "rr_mean": 1.0,
472
+ "rr_std": 0.0,
473
+ "pqs_steered_mean": 0.25,
474
+ "pqs_base_mean": 0.30841244078579816,
475
+ "pqs_delta_mean": -0.05841244078579816,
476
+ "collapse_rate": 0.0,
477
+ "steered_plan_mean": 0.0,
478
+ "steered_mon_mean": 0.0
479
+ }
480
+ },
481
+ "v2_ortho_general": {
482
+ "-0.5": {
483
+ "n": 30,
484
+ "rr_mean": 0.32190476190476186,
485
+ "rr_std": 0.4103074489446344,
486
+ "pqs_steered_mean": 0.298527176889486,
487
+ "pqs_base_mean": 0.30841244078579816,
488
+ "pqs_delta_mean": -0.009885263896312169,
489
+ "collapse_rate": 1.0,
490
+ "steered_plan_mean": 1.0,
491
+ "steered_mon_mean": 4.633333333333334
492
+ },
493
+ "0.0": {
494
+ "n": 30,
495
+ "rr_mean": 0.04342352092352093,
496
+ "rr_std": 0.6257206481305819,
497
+ "pqs_steered_mean": 0.31714557069926114,
498
+ "pqs_base_mean": 0.30841244078579816,
499
+ "pqs_delta_mean": 0.008733129913462978,
500
+ "collapse_rate": 1.0,
501
+ "steered_plan_mean": 0.8666666666666667,
502
+ "steered_mon_mean": 6.533333333333333
503
+ },
504
+ "0.25": {
505
+ "n": 30,
506
+ "rr_mean": -1.8218073593073596,
507
+ "rr_std": 9.496118945701825,
508
+ "pqs_steered_mean": 0.30722077258213093,
509
+ "pqs_base_mean": 0.30841244078579816,
510
+ "pqs_delta_mean": -0.0011916682036672288,
511
+ "collapse_rate": 1.0,
512
+ "steered_plan_mean": 1.2,
513
+ "steered_mon_mean": 16.8
514
+ },
515
+ "0.5": {
516
+ "n": 30,
517
+ "rr_mean": -1.1499603174603175,
518
+ "rr_std": 4.821146472475228,
519
+ "pqs_steered_mean": 0.3201625265317396,
520
+ "pqs_base_mean": 0.30841244078579816,
521
+ "pqs_delta_mean": 0.011750085745941441,
522
+ "collapse_rate": 1.0,
523
+ "steered_plan_mean": 1.0,
524
+ "steered_mon_mean": 29.2
525
+ },
526
+ "0.75": {
527
+ "n": 30,
528
+ "rr_mean": -3.0733645983645985,
529
+ "rr_std": 9.910020378454856,
530
+ "pqs_steered_mean": 0.3045656129481819,
531
+ "pqs_base_mean": 0.30841244078579816,
532
+ "pqs_delta_mean": -0.0038468278376162512,
533
+ "collapse_rate": 1.0,
534
+ "steered_plan_mean": 1.4666666666666666,
535
+ "steered_mon_mean": 33.46666666666667
536
+ },
537
+ "1.0": {
538
+ "n": 30,
539
+ "rr_mean": 0.0,
540
+ "rr_std": 0.0,
541
+ "pqs_steered_mean": 0.30841244078579816,
542
+ "pqs_base_mean": 0.30841244078579816,
543
+ "pqs_delta_mean": 0.0,
544
+ "collapse_rate": 1.0,
545
+ "steered_plan_mean": 1.3666666666666667,
546
+ "steered_mon_mean": 8.833333333333334
547
+ },
548
+ "1.5": {
549
+ "n": 30,
550
+ "rr_mean": 1.0,
551
+ "rr_std": 0.0,
552
+ "pqs_steered_mean": 0.25,
553
+ "pqs_base_mean": 0.30841244078579816,
554
+ "pqs_delta_mean": -0.05841244078579816,
555
+ "collapse_rate": 1.0,
556
+ "steered_plan_mean": 0.0,
557
+ "steered_mon_mean": 0.0
558
+ },
559
+ "2.0": {
560
+ "n": 30,
561
+ "rr_mean": 1.0,
562
+ "rr_std": 0.0,
563
+ "pqs_steered_mean": 0.25,
564
+ "pqs_base_mean": 0.30841244078579816,
565
+ "pqs_delta_mean": -0.05841244078579816,
566
+ "collapse_rate": 0.6666666666666666,
567
+ "steered_plan_mean": 0.0,
568
+ "steered_mon_mean": 0.0
569
+ }
570
+ },
571
+ "v3_ortho_crossdim": {
572
+ "-0.5": {
573
+ "n": 30,
574
+ "rr_mean": 0.3439622414622415,
575
+ "rr_std": 0.4348968221798721,
576
+ "pqs_steered_mean": 0.2968737197819469,
577
+ "pqs_base_mean": 0.30841244078579816,
578
+ "pqs_delta_mean": -0.011538721003851282,
579
+ "collapse_rate": 1.0,
580
+ "steered_plan_mean": 0.9,
581
+ "steered_mon_mean": 4.366666666666666
582
+ },
583
+ "0.0": {
584
+ "n": 30,
585
+ "rr_mean": 0.10671356421356422,
586
+ "rr_std": 0.565627572697744,
587
+ "pqs_steered_mean": 0.30792920635082843,
588
+ "pqs_base_mean": 0.30841244078579816,
589
+ "pqs_delta_mean": -0.00048323443496972596,
590
+ "collapse_rate": 1.0,
591
+ "steered_plan_mean": 0.8,
592
+ "steered_mon_mean": 6.333333333333333
593
+ },
594
+ "0.25": {
595
+ "n": 30,
596
+ "rr_mean": -0.1077934102934103,
597
+ "rr_std": 0.8850323564307493,
598
+ "pqs_steered_mean": 0.29818491426091176,
599
+ "pqs_base_mean": 0.30841244078579816,
600
+ "pqs_delta_mean": -0.010227526524886399,
601
+ "collapse_rate": 1.0,
602
+ "steered_plan_mean": 1.0333333333333334,
603
+ "steered_mon_mean": 7.1
604
+ },
605
+ "0.5": {
606
+ "n": 30,
607
+ "rr_mean": -2.281125541125541,
608
+ "rr_std": 10.731958419498815,
609
+ "pqs_steered_mean": 0.31830665354516297,
610
+ "pqs_base_mean": 0.30841244078579816,
611
+ "pqs_delta_mean": 0.009894212759364807,
612
+ "collapse_rate": 0.9666666666666667,
613
+ "steered_plan_mean": 1.1333333333333333,
614
+ "steered_mon_mean": 27.766666666666666
615
+ },
616
+ "0.75": {
617
+ "n": 30,
618
+ "rr_mean": -2.109230399230399,
619
+ "rr_std": 10.40077593153371,
620
+ "pqs_steered_mean": 0.31695194416485756,
621
+ "pqs_base_mean": 0.30841244078579816,
622
+ "pqs_delta_mean": 0.008539503379059399,
623
+ "collapse_rate": 1.0,
624
+ "steered_plan_mean": 1.4,
625
+ "steered_mon_mean": 24.333333333333332
626
+ },
627
+ "1.0": {
628
+ "n": 30,
629
+ "rr_mean": 0.0,
630
+ "rr_std": 0.0,
631
+ "pqs_steered_mean": 0.30841244078579816,
632
+ "pqs_base_mean": 0.30841244078579816,
633
+ "pqs_delta_mean": 0.0,
634
+ "collapse_rate": 1.0,
635
+ "steered_plan_mean": 1.3666666666666667,
636
+ "steered_mon_mean": 8.833333333333334
637
+ },
638
+ "1.5": {
639
+ "n": 30,
640
+ "rr_mean": 1.0,
641
+ "rr_std": 0.0,
642
+ "pqs_steered_mean": 0.25,
643
+ "pqs_base_mean": 0.30841244078579816,
644
+ "pqs_delta_mean": -0.05841244078579816,
645
+ "collapse_rate": 1.0,
646
+ "steered_plan_mean": 0.0,
647
+ "steered_mon_mean": 0.0
648
+ },
649
+ "2.0": {
650
+ "n": 30,
651
+ "rr_mean": 1.0,
652
+ "rr_std": 0.0,
653
+ "pqs_steered_mean": 0.25,
654
+ "pqs_base_mean": 0.30841244078579816,
655
+ "pqs_delta_mean": -0.05841244078579816,
656
+ "collapse_rate": 1.0,
657
+ "steered_plan_mean": 0.0,
658
+ "steered_mon_mean": 0.0
659
+ }
660
+ },
661
+ "v4_pca": {
662
+ "-0.5": {
663
+ "n": 30,
664
+ "rr_mean": -0.11861592111592106,
665
+ "rr_std": 0.8729216189884192,
666
+ "pqs_steered_mean": 0.2820137484761912,
667
+ "pqs_base_mean": 0.30841244078579816,
668
+ "pqs_delta_mean": -0.026398692309606953,
669
+ "collapse_rate": 1.0,
670
+ "steered_plan_mean": 0.8,
671
+ "steered_mon_mean": 7.466666666666667
672
+ },
673
+ "0.0": {
674
+ "n": 30,
675
+ "rr_mean": -0.8258104858104858,
676
+ "rr_std": 1.9635574366581219,
677
+ "pqs_steered_mean": 0.2894055905301104,
678
+ "pqs_base_mean": 0.30841244078579816,
679
+ "pqs_delta_mean": -0.01900685025568777,
680
+ "collapse_rate": 1.0,
681
+ "steered_plan_mean": 0.7333333333333333,
682
+ "steered_mon_mean": 11.1
683
+ },
684
+ "0.25": {
685
+ "n": 30,
686
+ "rr_mean": -2.5814646464646462,
687
+ "rr_std": 10.876049983511551,
688
+ "pqs_steered_mean": 0.30813163250421205,
689
+ "pqs_base_mean": 0.30841244078579816,
690
+ "pqs_delta_mean": -0.00028080828158610904,
691
+ "collapse_rate": 1.0,
692
+ "steered_plan_mean": 0.9333333333333333,
693
+ "steered_mon_mean": 40.43333333333333
694
+ },
695
+ "0.5": {
696
+ "n": 30,
697
+ "rr_mean": -0.6563985088985089,
698
+ "rr_std": 0.9562967704833188,
699
+ "pqs_steered_mean": 0.31778842284296754,
700
+ "pqs_base_mean": 0.30841244078579816,
701
+ "pqs_delta_mean": 0.009375982057169385,
702
+ "collapse_rate": 1.0,
703
+ "steered_plan_mean": 1.0666666666666667,
704
+ "steered_mon_mean": 13.266666666666667
705
+ },
706
+ "0.75": {
707
+ "n": 30,
708
+ "rr_mean": -1.130643338143338,
709
+ "rr_std": 2.83763625940499,
710
+ "pqs_steered_mean": 0.3299514523532939,
711
+ "pqs_base_mean": 0.30841244078579816,
712
+ "pqs_delta_mean": 0.021539011567495714,
713
+ "collapse_rate": 1.0,
714
+ "steered_plan_mean": 1.4,
715
+ "steered_mon_mean": 16.133333333333333
716
+ },
717
+ "1.0": {
718
+ "n": 30,
719
+ "rr_mean": 0.0,
720
+ "rr_std": 0.0,
721
+ "pqs_steered_mean": 0.30841244078579816,
722
+ "pqs_base_mean": 0.30841244078579816,
723
+ "pqs_delta_mean": 0.0,
724
+ "collapse_rate": 1.0,
725
+ "steered_plan_mean": 1.3666666666666667,
726
+ "steered_mon_mean": 8.833333333333334
727
+ },
728
+ "1.5": {
729
+ "n": 30,
730
+ "rr_mean": 1.0,
731
+ "rr_std": 0.0,
732
+ "pqs_steered_mean": 0.25,
733
+ "pqs_base_mean": 0.30841244078579816,
734
+ "pqs_delta_mean": -0.05841244078579816,
735
+ "collapse_rate": 0.8333333333333334,
736
+ "steered_plan_mean": 0.0,
737
+ "steered_mon_mean": 0.0
738
+ },
739
+ "2.0": {
740
+ "n": 30,
741
+ "rr_mean": 1.0,
742
+ "rr_std": 0.0,
743
+ "pqs_steered_mean": 0.25,
744
+ "pqs_base_mean": 0.30841244078579816,
745
+ "pqs_delta_mean": -0.05841244078579816,
746
+ "collapse_rate": 1.0,
747
+ "steered_plan_mean": 0.0,
748
+ "steered_mon_mean": 0.0
749
+ }
750
+ },
751
+ "force_suppress": {
752
+ "None": {
753
+ "n": 30,
754
+ "rr_mean": 0.06018157768157769,
755
+ "rr_std": 0.5947990594271211,
756
+ "pqs_steered_mean": 0.31499010451936,
757
+ "pqs_base_mean": 0.30841244078579816,
758
+ "pqs_delta_mean": 0.006577663733561834,
759
+ "collapse_rate": 1.0,
760
+ "steered_plan_mean": 1.4,
761
+ "steered_mon_mean": 7.9
762
+ }
763
+ },
764
+ "force_enhance": {
765
+ "None": {
766
+ "n": 30,
767
+ "rr_mean": -0.3016774891774892,
768
+ "rr_std": 0.7792200188471652,
769
+ "pqs_steered_mean": 0.2996864754363997,
770
+ "pqs_base_mean": 0.30841244078579816,
771
+ "pqs_delta_mean": -0.008725965349398457,
772
+ "collapse_rate": 1.0,
773
+ "steered_plan_mean": 1.4666666666666666,
774
+ "steered_mon_mean": 9.6
775
+ }
776
+ }
777
+ }
778
+ }