Hugging Face's logo

Datasets:
SHAPE-Morph
/
benchmark-logs

Modalities:
Text
Formats:
csv
Size:
10K - 100K
Tags:
biology
microscopy
cell-segmentation
agent
benchmark
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
5xuekun commited on
Commit
552eb56
·
verified ·
1 Parent(s): 13a05e0

Publish measured Luna/Sonnet scenario-2 comparison; n=1, max failure explicit

Browse files
Files changed (12) hide show
  1. logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/MANIFEST.json +18 -0
  2. logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/PROVENANCE.json +116 -0
  3. logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/README.md +28 -0
  4. logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/comparison.csv +11 -0
  5. logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/comparison.json +480 -0
  6. logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/luna/PREPARE.json +103 -0
  7. logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/luna/SWEEP_REPORT.json +203 -0
  8. logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/sonnet/DATA_PROVENANCE.json +24 -0
  9. logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/sonnet/PREFLIGHT.json +32 -0
  10. logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/sonnet/PREPARE.json +96 -0
  11. logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/sonnet/STUDY.json +26 -0
  12. logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/sonnet/SWEEP_REPORT.json +187 -0
logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/MANIFEST.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "sha256": {
3
+ "comparison.csv": "422196211bbab1500d11342f7e9e24cd196c5c2bcd6bd979603f5e1ce1f7937d",
4
+ "comparison.json": "56a9c5c2d7b8f027892370c5d908da2ce0a9d11f1401f26c53f49a03a4bb7dc2",
5
+ "luna/PREPARE.json": "c30e35b71ced53228ec7224f76807494eb82db40ab2109e5b162d49a9ec20f38",
6
+ "luna/SWEEP_REPORT.json": "4e7ffc7b9fb5288af1fcc03d99d026d5e46aef4fbd476a798aa45507f86910f5",
7
+ "PROVENANCE.json": "88d2b9361d105a5fcd5d53f032c402f649242a0ea6db44b70e16c425ff6a5a58",
8
+ "README.md": "b5e7d922fe62cefde7126668d9802cf6a91528ec3b3e0828bddd66cc61aeb8ef",
9
+ "sonnet/DATA_PROVENANCE.json": "df21ddb48095276a76bd0347af58f30a84086a3540c25113f9e6baea7a056243",
10
+ "sonnet/PREFLIGHT.json": "31b48116305d32be6bb88603ebc09556cf17d76141af47999dba10c217753836",
11
+ "sonnet/PREPARE.json": "1bb787ea23a2e929a6fe427821d86a84590ffb4f4535a7e059e20ebafaaa72e0",
12
+ "sonnet/STUDY.json": "9fde2e03b12883f0ceeaae9e4ebcbc226bc03c7995609d958717e4645e6f4b03",
13
+ "sonnet/SWEEP_REPORT.json": "e3d63f6b82c50f507f7e23bf252663cdc81aee83684f1005ebf6d6ee3d9abdec"
14
+ },
15
+ "rows": 10,
16
+ "normal_performance_bars": 9,
17
+ "max_failed_under_fixed_budget": true
18
+ }
logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/PROVENANCE.json ADDED
@@ -0,0 +1,116 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "scenario": "scenario_2",
3
+ "n_runs_per_model_effort": 1,
4
+ "common_efforts": [
5
+ "low",
6
+ "medium",
7
+ "high",
8
+ "xhigh"
9
+ ],
10
+ "model_specific_efforts": {
11
+ "luna": [
12
+ "none"
13
+ ],
14
+ "sonnet": [
15
+ "max"
16
+ ]
17
+ },
18
+ "max_rounds": 8,
19
+ "max_tokens": 32000,
20
+ "timeout_seconds": 900,
21
+ "cores": 16,
22
+ "input_feature_sha256": "65366e8bbcef188def7801e5ef2a54fc7e740a2eedc41a010cfb2647b723604a",
23
+ "input_dimensions": {
24
+ "wells": 14194,
25
+ "plates": 37,
26
+ "features": 1027
27
+ },
28
+ "compound_counts": {
29
+ "calibration": 146,
30
+ "test": 147
31
+ },
32
+ "metadata_parquet_sha256": {
33
+ "luna": "e1c131da3e6c76b7974ab4246964662f2b842baf897b8484acfd2c25f921196f",
34
+ "sonnet": "fa7984a396f23dc2d0b5049e220ddd03d015eeec525d9c5f3d44053ae87e303b"
35
+ },
36
+ "sources": {
37
+ "luna": {
38
+ "path": "/home/xuekun/shape/code/SHAPE-Revision-batch-sonnet55-effort-20261031/benchmarks/batch_correction_JUMP/results/scenario2_effort_e3b9e47b_20261013T013018Z_47584915",
39
+ "report_sha256": "4e7ffc7b9fb5288af1fcc03d99d026d5e46aef4fbd476a798aa45507f86910f5",
40
+ "prepare_sha256": "c30e35b71ced53228ec7224f76807494eb82db40ab2109e5b162d49a9ec20f38"
41
+ },
42
+ "sonnet": {
43
+ "path": "/mnt/filesystem-e3/SHAPE-Revision/benchmarks/batch_correction_JUMP/results/scenario2_sonnet55_effort_d9a07987_20261031",
44
+ "report_sha256": "e3d63f6b82c50f507f7e23bf252663cdc81aee83684f1005ebf6d6ee3d9abdec",
45
+ "prepare_sha256": "1bb787ea23a2e929a6fe427821d86a84590ffb4f4535a7e059e20ebafaaa72e0"
46
+ }
47
+ },
48
+ "task_commits": {
49
+ "luna": "e3b9e47b",
50
+ "sonnet": "d9a07987ce103eb4ffa35bdcd50972cb3d1baa8e"
51
+ },
52
+ "historical_luna_repository_snapshot": "d9a07987ce103eb4ffa35bdcd50972cb3d1baa8e",
53
+ "sonnet_executed_source_sha256": {
54
+ "benchmarks/llm_reasoning/test_native_effort_engine.py": "f5b3fbb17e6a15e517433b0721bd1ce592349311eec1fa7c2c61b8b60d73737c",
55
+ "benchmarks/llm_reasoning/native_effort_engine.py": "d4cfd9c1b2404d4a7cdb6378303e470ed5012fef2eab919fc27f63cd20aa5e5a",
56
+ "benchmarks/llm_reasoning/run_batch_effort.py": "1729ad70da69d9a31fb0b53e5a57c35b4a14ff4b2a5fd50016d9c183d6f03be1"
57
+ },
58
+ "builder_sha256": "415335e03cabf22251760f22f5fbce02a702d178b0b9c7b5533a400e1918be4b",
59
+ "comparison_scope": "Descriptive comparison of existing measured single runs; no repeat-run error bars or significance tests.",
60
+ "version_notes": "Task runner, scorer, control and quantification code are unchanged between the two task commits. The batch verifier difference is an improved null-parameter error message. Package environments differ; metadata Parquet hashes are retained separately. This is not an isolated causal model-effect experiment.",
61
+ "cancelled_repeat_study": {
62
+ "planned_runs": 20,
63
+ "executed_runs": 0,
64
+ "cancelled_by_user": true
65
+ },
66
+ "reply_errors_reviewed": true,
67
+ "native_usage_reconciliation": {
68
+ "low": {
69
+ "response_records": 7,
70
+ "usage_records": 7,
71
+ "thinking_tokens": 0,
72
+ "thinking_usage_coverage": 7,
73
+ "total_tokens": 96837,
74
+ "max_token_truncations": 0,
75
+ "source_sha256": "713b7b11bb121d54792c11cc3b9c8ad4524280f201963defea2537ef462a6961"
76
+ },
77
+ "medium": {
78
+ "response_records": 9,
79
+ "usage_records": 9,
80
+ "thinking_tokens": 0,
81
+ "thinking_usage_coverage": 9,
82
+ "total_tokens": 163416,
83
+ "max_token_truncations": 0,
84
+ "source_sha256": "f5f0f989ae3aee0e206104864b3d032a831cd3bba3a3e409505c92a64a3fecf0"
85
+ },
86
+ "high": {
87
+ "response_records": 8,
88
+ "usage_records": 8,
89
+ "thinking_tokens": 2041,
90
+ "thinking_usage_coverage": 8,
91
+ "total_tokens": 200781,
92
+ "max_token_truncations": 0,
93
+ "source_sha256": "04f04de9d6c82c484b1b02d34e06fc457cbe0a2097ba9bcb9caba4943e694186"
94
+ },
95
+ "xhigh": {
96
+ "response_records": 9,
97
+ "usage_records": 9,
98
+ "thinking_tokens": 10731,
99
+ "thinking_usage_coverage": 9,
100
+ "total_tokens": 359276,
101
+ "max_token_truncations": 0,
102
+ "source_sha256": "fec73d03704e648a0f8ccb5f6c7a48f356f2087e59b7c4b5f4ca9e936b5df106"
103
+ },
104
+ "max": {
105
+ "response_records": 4,
106
+ "usage_records": 4,
107
+ "thinking_tokens": 96500,
108
+ "thinking_usage_coverage": 4,
109
+ "total_tokens": 116361,
110
+ "max_token_truncations": 3,
111
+ "source_sha256": "28de4f069de010c7f4258a6a6f7deee015b0afc58e458ee45d523df1e6203572"
112
+ }
113
+ },
114
+ "usage_note": "Original canonical reports remain unchanged. The initial adapter omitted the provider nested thinking-token field; this derived comparison recovers exact reported values from raw native traces and verifies total tokens. No reasoning-token value is estimated.",
115
+ "failure_policy": "Keep recorded replies, truncations and harness fallback scores in the data. A condition with reply failures and no tool search is marked failed and omitted from normal performance bars; never fabricate a replacement."
116
+ }
logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/README.md ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # JUMP Scenario 2: Luna and Sonnet reasoning effort
2
+
3
+ Existing GPT-5.6-luna results are reused; Claude Sonnet 5.5 was run once at each native effort. **n=1 attempt per model and effort. No error bars across LLM runs are available.**
4
+
5
+ | model | effort | status | test overall | gain vs Harmony | rounds | tool calls | tokens | wall (min) |
6
+ |---|---|---|---|---|---|---|---|---|
7
+ | GPT-5.6-luna | none | completed | 0.5589 | +0.0079 | 3 | 13 | 67473 | 5.6 |
8
+ | GPT-5.6-luna | low | completed | 0.5631 | +0.0121 | 4 | 27 | 89710 | 11.9 |
9
+ | Claude Sonnet 5.5 | low | completed | 0.5641 | +0.0131 | 4 | 19 | 96837 | 5.1 |
10
+ | GPT-5.6-luna | medium | completed | 0.5679 | +0.0169 | 3 | 35 | 109831 | 11.3 |
11
+ | Claude Sonnet 5.5 | medium | completed | 0.5789 | +0.0279 | 6 | 26 | 163416 | 22.6 |
12
+ | GPT-5.6-luna | high | completed | 0.5717 | +0.0207 | 3 | 36 | 128376 | 111.7 |
13
+ | Claude Sonnet 5.5 | high | completed | 0.5902 | +0.0392 | 5 | 40 | 200781 | 43.8 |
14
+ | GPT-5.6-luna | xhigh | completed | 0.5685 | +0.0175 | 5 | 83 | 270770 | 41.3 |
15
+ | Claude Sonnet 5.5 | xhigh | completed | 0.5898 | +0.0388 | 6 | 63 | 359276 | 108.1 |
16
+ | Claude Sonnet 5.5 | max | failed_no_tool_search | 0.5098 | -0.0412 | 0 | 0 | 116361 | 13.6 |
17
+
18
+ The prepared Scenario-2 feature matrix is byte-identical: 14,194 wells, 37 TARGET2 plates, 1,027 features, sources 2/6/10. Calibration uses 146 compounds and evaluation uses 147 held-out compounds plus controls. Budgets are at most 8 tuning rounds, 32,000 tokens per reply, 900-second request timeout and 16 CPU cores. A round may contain several tool calls.
19
+
20
+ Fig1e uses `overall_gain_vs_harmony` directly from each canonical sweep report, with no error bars. Failed no-search conditions are explicitly marked; their harness fallback score is retained in the table but is not plotted as normal model performance. Original paired-plate ±2SE values remain in the data for provenance and are not interpreted as uncertainty across independent LLM repeats. Unsupported levels are omitted rather than plotted as zero.
21
+
22
+ Task runner, scorer, control and quantification code are unchanged between the two task commits. The batch verifier difference is an improved null-parameter error message. Package environments differ; metadata Parquet hashes are retained separately. This is not an isolated causal model-effect experiment.
23
+
24
+ This descriptive comparison reuses historical Luna runs rather than rerunning them. The planned 20-condition repeat experiment was cancelled before execution; none of its preflight calls are counted as experimental runs. Native effort labels do not imply equal reasoning compute across providers. Wall time is the canonical AGENT_RUN duration and excludes the separate final held-out scoring stage. Native thinking-token totals are recovered exactly from `usage.output_tokens_details.thinking_tokens` in the raw responses; the original canonical reports are retained unchanged and total tokens are reconciled. Missing provider counts remain missing, rather than being recorded as zero.
25
+
26
+ Full traces and per-file hashes: [GPT-5.6-luna](https://huggingface.co/datasets/SHAPE-Morph/benchmark-logs/tree/main/logs/reasoning_effort/scenario2_effort_e3b9e47b_20261013T013018Z_47584915), [Claude Sonnet 5.5](https://huggingface.co/datasets/SHAPE-Morph/benchmark-logs/tree/main/logs/reasoning_effort/scenario2_sonnet55_effort_d9a07987_20261031).
27
+
28
+ Source tables and provenance: `comparison.csv`, `comparison.json`, `PROVENANCE.json`, and the canonical per-model `SWEEP_REPORT.json` and `PREPARE.json` snapshots.
logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/comparison.csv ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model_key,model,model_label,effort,n_runs,test_overall,overall_gain_vs_harmony,paired_plate_band_2se_not_plotted,test_batch_mean,test_bio_mean,test_nonrep_mean_map,rounds,tool_calls,total_tokens,reasoning_tokens,canonical_reasoning_tokens,reasoning_token_source,reply_errors,truncated_replies,outcome_status,plot_included,wall_seconds,from_call,decision,stopped_because,non_default_parameters,study
2
+ luna,gpt-5.6-luna,GPT-5.6-luna,none,1,0.5589373506038878,0.007939896512330069,0.0028078987308382108,0.5089071446119035,0.5922908212652108,0.5652392744438554,3,13,67473,0,0,canonical API usage,0,0,completed,True,335.9,agent_13,RUN,satisfied,"{""input_pca"": 100, ""input_scale"": ""zscore"", ""lambda_value"": 0.2, ""max_iter_harmony"": 30, ""theta"": 4.0}",scenario2_effort_e3b9e47b_20261013T013018Z_47584915
3
+ luna,gpt-5.6-luna,GPT-5.6-luna,low,1,0.5631142086462773,0.0121167545547195,0.0025752097311447313,0.5212980495643695,0.5909916480342158,0.5543929767588093,4,27,89710,848,848,canonical API usage,0,0,completed,True,714.6,agent_22,RUN,plateau,"{""block_size"": 0.1, ""input_pca"": 400, ""input_scale"": ""zscore"", ""lambda_value"": 1.0, ""max_iter_harmony"": 35, ""nclust"": 300, ""theta"": 16.0}",scenario2_effort_e3b9e47b_20261013T013018Z_47584915
4
+ sonnet,claude-sonnet-5-5,Claude Sonnet 5.5,low,1,0.5641117776830196,0.013107532376362085,0.0033610760681294432,0.5166596999139929,0.5957464961957041,0.5719303606516389,4,19,96837,0,,native usage.output_tokens_details.thinking_tokens,0,0,completed,True,308.9,agent_16,RUN,satisfied,"{""epsilon_harmony"": 0.001, ""input_pca"": 50, ""input_scale"": ""zscore"", ""lambda_value"": 0.3, ""max_iter_harmony"": 30, ""theta"": 8.0}",scenario2_sonnet55_effort_d9a07987_20261031
5
+ luna,gpt-5.6-luna,GPT-5.6-luna,medium,1,0.5678742039333458,0.016876749841787997,0.002743804854260152,0.5315159095120535,0.592113066880874,0.5650862454436287,3,35,109831,1621,1621,canonical API usage,0,0,completed,True,678.7,agent_22,RUN,satisfied,"{""epsilon_harmony"": 0.005, ""input_pca"": 128, ""input_scale"": ""zscore"", ""lambda_value"": 0.5, ""max_iter_harmony"": 20, ""nclust"": 200, ""theta"": 6.0}",scenario2_effort_e3b9e47b_20261013T013018Z_47584915
6
+ sonnet,claude-sonnet-5-5,Claude Sonnet 5.5,medium,1,0.5789328652909089,0.027928619984251446,0.0035681044753392097,0.548963537300199,0.5989124172847157,0.569990239246832,6,26,163416,0,,native usage.output_tokens_details.thinking_tokens,0,0,completed,True,1354.2,agent_23,RUN,satisfied,"{""epsilon_harmony"": 0.001, ""input_pca"": 100, ""nclust"": 2000, ""theta"": 3.0}",scenario2_sonnet55_effort_d9a07987_20261031
7
+ luna,gpt-5.6-luna,GPT-5.6-luna,high,1,0.5717093741326299,0.02071192004107214,0.004071871255351384,0.549205162791825,0.5867121816931665,0.5503436432783145,3,36,128376,10082,10082,canonical API usage,0,0,completed,True,6704.7,agent_31,RUN,satisfied,"{""epsilon_harmony"": 0.0001, ""input_pca"": 800, ""input_scale"": ""zscore"", ""lambda_value"": 0.5, ""max_iter_harmony"": 40, ""nclust"": 500, ""sigma"": 0.085, ""theta"": 20.0}",scenario2_effort_e3b9e47b_20261013T013018Z_47584915
8
+ sonnet,claude-sonnet-5-5,Claude Sonnet 5.5,high,1,0.5902377036116497,0.03923345830499225,0.003724520645030581,0.5863485084238774,0.592830500403498,0.5755813568696371,5,40,200781,2041,,native usage.output_tokens_details.thinking_tokens,0,0,completed,True,2625.8,agent_35,RUN,satisfied,"{""input_pca"": 50, ""input_scale"": ""zscore"", ""max_iter_kmeans"": 20, ""nclust"": 1500, ""sigma"": 0.07, ""theta"": 12.0}",scenario2_sonnet55_effort_d9a07987_20261031
9
+ luna,gpt-5.6-luna,GPT-5.6-luna,xhigh,1,0.5684772870340074,0.01747983294244959,0.0028938841405390884,0.5317450306175225,0.5929654579783307,0.5612766108681524,5,83,270770,27823,27823,canonical API usage,0,0,completed,True,2477.3,agent_27,RUN,plateau,"{""block_size"": 0.1, ""epsilon_cluster"": 0.0001, ""epsilon_harmony"": 0.001, ""input_pca"": 100, ""lambda_value"": 5.0, ""max_iter_harmony"": 30, ""max_iter_kmeans"": 10, ""nclust"": 200, ""sigma"": 0.05, ""tau"": 50.0, ""theta"": 6.0}",scenario2_effort_e3b9e47b_20261013T013018Z_47584915
10
+ sonnet,claude-sonnet-5-5,Claude Sonnet 5.5,xhigh,1,0.5897806350012761,0.03877638969461861,0.003844277805579813,0.5851601200331521,0.5928609783133587,0.5755300812664508,6,63,359276,10731,,native usage.output_tokens_details.thinking_tokens,0,0,completed,True,6486.9,agent_49,RUN,satisfied,"{""epsilon_harmony"": 0.0001, ""input_pca"": 50, ""input_scale"": ""zscore"", ""max_iter_harmony"": 40, ""nclust"": 800, ""sigma"": 0.06, ""theta"": 5.0}",scenario2_sonnet55_effort_d9a07987_20261031
11
+ sonnet,claude-sonnet-5-5,Claude Sonnet 5.5,max,1,0.5098476612243077,-0.04115658408234979,0.002083827092896988,0.4096857507996082,0.5766222681741072,0.5094418711383188,0,0,116361,96500,,native usage.output_tokens_details.thinking_tokens,3,3,failed_no_tool_search,False,815.3,,ABSTAIN,execution_error,,scenario2_sonnet55_effort_d9a07987_20261031
logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/comparison.json ADDED
@@ -0,0 +1,480 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "rows": [
3
+ {
4
+ "model_key": "luna",
5
+ "model": "gpt-5.6-luna",
6
+ "model_label": "GPT-5.6-luna",
7
+ "effort": "none",
8
+ "n_runs": 1,
9
+ "test_overall": 0.5589373506038878,
10
+ "overall_gain_vs_harmony": 0.007939896512330069,
11
+ "paired_plate_band_2se_not_plotted": 0.0028078987308382108,
12
+ "test_batch_mean": 0.5089071446119035,
13
+ "test_bio_mean": 0.5922908212652108,
14
+ "test_nonrep_mean_map": 0.5652392744438554,
15
+ "rounds": 3,
16
+ "tool_calls": 13,
17
+ "total_tokens": 67473,
18
+ "reasoning_tokens": 0,
19
+ "canonical_reasoning_tokens": 0,
20
+ "reasoning_token_source": "canonical API usage",
21
+ "reply_errors": 0,
22
+ "truncated_replies": 0,
23
+ "outcome_status": "completed",
24
+ "plot_included": true,
25
+ "wall_seconds": 335.9,
26
+ "from_call": "agent_13",
27
+ "decision": "RUN",
28
+ "stopped_because": "satisfied",
29
+ "non_default_parameters": {
30
+ "theta": 4.0,
31
+ "lambda_value": 0.2,
32
+ "max_iter_harmony": 30,
33
+ "input_scale": "zscore",
34
+ "input_pca": 100
35
+ },
36
+ "study": "scenario2_effort_e3b9e47b_20261013T013018Z_47584915"
37
+ },
38
+ {
39
+ "model_key": "luna",
40
+ "model": "gpt-5.6-luna",
41
+ "model_label": "GPT-5.6-luna",
42
+ "effort": "low",
43
+ "n_runs": 1,
44
+ "test_overall": 0.5631142086462773,
45
+ "overall_gain_vs_harmony": 0.0121167545547195,
46
+ "paired_plate_band_2se_not_plotted": 0.0025752097311447313,
47
+ "test_batch_mean": 0.5212980495643695,
48
+ "test_bio_mean": 0.5909916480342158,
49
+ "test_nonrep_mean_map": 0.5543929767588093,
50
+ "rounds": 4,
51
+ "tool_calls": 27,
52
+ "total_tokens": 89710,
53
+ "reasoning_tokens": 848,
54
+ "canonical_reasoning_tokens": 848,
55
+ "reasoning_token_source": "canonical API usage",
56
+ "reply_errors": 0,
57
+ "truncated_replies": 0,
58
+ "outcome_status": "completed",
59
+ "plot_included": true,
60
+ "wall_seconds": 714.6,
61
+ "from_call": "agent_22",
62
+ "decision": "RUN",
63
+ "stopped_because": "plateau",
64
+ "non_default_parameters": {
65
+ "theta": 16.0,
66
+ "lambda_value": 1.0,
67
+ "nclust": 300,
68
+ "max_iter_harmony": 35,
69
+ "block_size": 0.1,
70
+ "input_scale": "zscore",
71
+ "input_pca": 400
72
+ },
73
+ "study": "scenario2_effort_e3b9e47b_20261013T013018Z_47584915"
74
+ },
75
+ {
76
+ "model_key": "sonnet",
77
+ "model": "claude-sonnet-5-5",
78
+ "model_label": "Claude Sonnet 5.5",
79
+ "effort": "low",
80
+ "n_runs": 1,
81
+ "test_overall": 0.5641117776830196,
82
+ "overall_gain_vs_harmony": 0.013107532376362085,
83
+ "paired_plate_band_2se_not_plotted": 0.0033610760681294432,
84
+ "test_batch_mean": 0.5166596999139929,
85
+ "test_bio_mean": 0.5957464961957041,
86
+ "test_nonrep_mean_map": 0.5719303606516389,
87
+ "rounds": 4,
88
+ "tool_calls": 19,
89
+ "total_tokens": 96837,
90
+ "reasoning_tokens": 0,
91
+ "canonical_reasoning_tokens": null,
92
+ "reasoning_token_source": "native usage.output_tokens_details.thinking_tokens",
93
+ "reply_errors": 0,
94
+ "truncated_replies": 0,
95
+ "outcome_status": "completed",
96
+ "plot_included": true,
97
+ "wall_seconds": 308.9,
98
+ "from_call": "agent_16",
99
+ "decision": "RUN",
100
+ "stopped_because": "satisfied",
101
+ "non_default_parameters": {
102
+ "theta": 8.0,
103
+ "lambda_value": 0.3,
104
+ "max_iter_harmony": 30,
105
+ "epsilon_harmony": 0.001,
106
+ "input_scale": "zscore",
107
+ "input_pca": 50
108
+ },
109
+ "study": "scenario2_sonnet55_effort_d9a07987_20261031"
110
+ },
111
+ {
112
+ "model_key": "luna",
113
+ "model": "gpt-5.6-luna",
114
+ "model_label": "GPT-5.6-luna",
115
+ "effort": "medium",
116
+ "n_runs": 1,
117
+ "test_overall": 0.5678742039333458,
118
+ "overall_gain_vs_harmony": 0.016876749841787997,
119
+ "paired_plate_band_2se_not_plotted": 0.002743804854260152,
120
+ "test_batch_mean": 0.5315159095120535,
121
+ "test_bio_mean": 0.592113066880874,
122
+ "test_nonrep_mean_map": 0.5650862454436287,
123
+ "rounds": 3,
124
+ "tool_calls": 35,
125
+ "total_tokens": 109831,
126
+ "reasoning_tokens": 1621,
127
+ "canonical_reasoning_tokens": 1621,
128
+ "reasoning_token_source": "canonical API usage",
129
+ "reply_errors": 0,
130
+ "truncated_replies": 0,
131
+ "outcome_status": "completed",
132
+ "plot_included": true,
133
+ "wall_seconds": 678.7,
134
+ "from_call": "agent_22",
135
+ "decision": "RUN",
136
+ "stopped_because": "satisfied",
137
+ "non_default_parameters": {
138
+ "theta": 6.0,
139
+ "lambda_value": 0.5,
140
+ "nclust": 200,
141
+ "max_iter_harmony": 20,
142
+ "epsilon_harmony": 0.005,
143
+ "input_scale": "zscore",
144
+ "input_pca": 128
145
+ },
146
+ "study": "scenario2_effort_e3b9e47b_20261013T013018Z_47584915"
147
+ },
148
+ {
149
+ "model_key": "sonnet",
150
+ "model": "claude-sonnet-5-5",
151
+ "model_label": "Claude Sonnet 5.5",
152
+ "effort": "medium",
153
+ "n_runs": 1,
154
+ "test_overall": 0.5789328652909089,
155
+ "overall_gain_vs_harmony": 0.027928619984251446,
156
+ "paired_plate_band_2se_not_plotted": 0.0035681044753392097,
157
+ "test_batch_mean": 0.548963537300199,
158
+ "test_bio_mean": 0.5989124172847157,
159
+ "test_nonrep_mean_map": 0.569990239246832,
160
+ "rounds": 6,
161
+ "tool_calls": 26,
162
+ "total_tokens": 163416,
163
+ "reasoning_tokens": 0,
164
+ "canonical_reasoning_tokens": null,
165
+ "reasoning_token_source": "native usage.output_tokens_details.thinking_tokens",
166
+ "reply_errors": 0,
167
+ "truncated_replies": 0,
168
+ "outcome_status": "completed",
169
+ "plot_included": true,
170
+ "wall_seconds": 1354.2,
171
+ "from_call": "agent_23",
172
+ "decision": "RUN",
173
+ "stopped_because": "satisfied",
174
+ "non_default_parameters": {
175
+ "theta": 3.0,
176
+ "nclust": 2000,
177
+ "epsilon_harmony": 0.001,
178
+ "input_pca": 100
179
+ },
180
+ "study": "scenario2_sonnet55_effort_d9a07987_20261031"
181
+ },
182
+ {
183
+ "model_key": "luna",
184
+ "model": "gpt-5.6-luna",
185
+ "model_label": "GPT-5.6-luna",
186
+ "effort": "high",
187
+ "n_runs": 1,
188
+ "test_overall": 0.5717093741326299,
189
+ "overall_gain_vs_harmony": 0.02071192004107214,
190
+ "paired_plate_band_2se_not_plotted": 0.004071871255351384,
191
+ "test_batch_mean": 0.549205162791825,
192
+ "test_bio_mean": 0.5867121816931665,
193
+ "test_nonrep_mean_map": 0.5503436432783145,
194
+ "rounds": 3,
195
+ "tool_calls": 36,
196
+ "total_tokens": 128376,
197
+ "reasoning_tokens": 10082,
198
+ "canonical_reasoning_tokens": 10082,
199
+ "reasoning_token_source": "canonical API usage",
200
+ "reply_errors": 0,
201
+ "truncated_replies": 0,
202
+ "outcome_status": "completed",
203
+ "plot_included": true,
204
+ "wall_seconds": 6704.7,
205
+ "from_call": "agent_31",
206
+ "decision": "RUN",
207
+ "stopped_because": "satisfied",
208
+ "non_default_parameters": {
209
+ "theta": 20.0,
210
+ "lambda_value": 0.5,
211
+ "sigma": 0.085,
212
+ "nclust": 500,
213
+ "max_iter_harmony": 40,
214
+ "epsilon_harmony": 0.0001,
215
+ "input_scale": "zscore",
216
+ "input_pca": 800
217
+ },
218
+ "study": "scenario2_effort_e3b9e47b_20261013T013018Z_47584915"
219
+ },
220
+ {
221
+ "model_key": "sonnet",
222
+ "model": "claude-sonnet-5-5",
223
+ "model_label": "Claude Sonnet 5.5",
224
+ "effort": "high",
225
+ "n_runs": 1,
226
+ "test_overall": 0.5902377036116497,
227
+ "overall_gain_vs_harmony": 0.03923345830499225,
228
+ "paired_plate_band_2se_not_plotted": 0.003724520645030581,
229
+ "test_batch_mean": 0.5863485084238774,
230
+ "test_bio_mean": 0.592830500403498,
231
+ "test_nonrep_mean_map": 0.5755813568696371,
232
+ "rounds": 5,
233
+ "tool_calls": 40,
234
+ "total_tokens": 200781,
235
+ "reasoning_tokens": 2041,
236
+ "canonical_reasoning_tokens": null,
237
+ "reasoning_token_source": "native usage.output_tokens_details.thinking_tokens",
238
+ "reply_errors": 0,
239
+ "truncated_replies": 0,
240
+ "outcome_status": "completed",
241
+ "plot_included": true,
242
+ "wall_seconds": 2625.8,
243
+ "from_call": "agent_35",
244
+ "decision": "RUN",
245
+ "stopped_because": "satisfied",
246
+ "non_default_parameters": {
247
+ "theta": 12.0,
248
+ "sigma": 0.07,
249
+ "nclust": 1500,
250
+ "max_iter_kmeans": 20,
251
+ "input_scale": "zscore",
252
+ "input_pca": 50
253
+ },
254
+ "study": "scenario2_sonnet55_effort_d9a07987_20261031"
255
+ },
256
+ {
257
+ "model_key": "luna",
258
+ "model": "gpt-5.6-luna",
259
+ "model_label": "GPT-5.6-luna",
260
+ "effort": "xhigh",
261
+ "n_runs": 1,
262
+ "test_overall": 0.5684772870340074,
263
+ "overall_gain_vs_harmony": 0.01747983294244959,
264
+ "paired_plate_band_2se_not_plotted": 0.0028938841405390884,
265
+ "test_batch_mean": 0.5317450306175225,
266
+ "test_bio_mean": 0.5929654579783307,
267
+ "test_nonrep_mean_map": 0.5612766108681524,
268
+ "rounds": 5,
269
+ "tool_calls": 83,
270
+ "total_tokens": 270770,
271
+ "reasoning_tokens": 27823,
272
+ "canonical_reasoning_tokens": 27823,
273
+ "reasoning_token_source": "canonical API usage",
274
+ "reply_errors": 0,
275
+ "truncated_replies": 0,
276
+ "outcome_status": "completed",
277
+ "plot_included": true,
278
+ "wall_seconds": 2477.3,
279
+ "from_call": "agent_27",
280
+ "decision": "RUN",
281
+ "stopped_because": "plateau",
282
+ "non_default_parameters": {
283
+ "theta": 6.0,
284
+ "lambda_value": 5.0,
285
+ "sigma": 0.05,
286
+ "nclust": 200,
287
+ "max_iter_harmony": 30,
288
+ "max_iter_kmeans": 10,
289
+ "epsilon_cluster": 0.0001,
290
+ "epsilon_harmony": 0.001,
291
+ "block_size": 0.1,
292
+ "tau": 50.0,
293
+ "input_pca": 100
294
+ },
295
+ "study": "scenario2_effort_e3b9e47b_20261013T013018Z_47584915"
296
+ },
297
+ {
298
+ "model_key": "sonnet",
299
+ "model": "claude-sonnet-5-5",
300
+ "model_label": "Claude Sonnet 5.5",
301
+ "effort": "xhigh",
302
+ "n_runs": 1,
303
+ "test_overall": 0.5897806350012761,
304
+ "overall_gain_vs_harmony": 0.03877638969461861,
305
+ "paired_plate_band_2se_not_plotted": 0.003844277805579813,
306
+ "test_batch_mean": 0.5851601200331521,
307
+ "test_bio_mean": 0.5928609783133587,
308
+ "test_nonrep_mean_map": 0.5755300812664508,
309
+ "rounds": 6,
310
+ "tool_calls": 63,
311
+ "total_tokens": 359276,
312
+ "reasoning_tokens": 10731,
313
+ "canonical_reasoning_tokens": null,
314
+ "reasoning_token_source": "native usage.output_tokens_details.thinking_tokens",
315
+ "reply_errors": 0,
316
+ "truncated_replies": 0,
317
+ "outcome_status": "completed",
318
+ "plot_included": true,
319
+ "wall_seconds": 6486.9,
320
+ "from_call": "agent_49",
321
+ "decision": "RUN",
322
+ "stopped_because": "satisfied",
323
+ "non_default_parameters": {
324
+ "theta": 5.0,
325
+ "sigma": 0.06,
326
+ "nclust": 800,
327
+ "max_iter_harmony": 40,
328
+ "epsilon_harmony": 0.0001,
329
+ "input_scale": "zscore",
330
+ "input_pca": 50
331
+ },
332
+ "study": "scenario2_sonnet55_effort_d9a07987_20261031"
333
+ },
334
+ {
335
+ "model_key": "sonnet",
336
+ "model": "claude-sonnet-5-5",
337
+ "model_label": "Claude Sonnet 5.5",
338
+ "effort": "max",
339
+ "n_runs": 1,
340
+ "test_overall": 0.5098476612243077,
341
+ "overall_gain_vs_harmony": -0.04115658408234979,
342
+ "paired_plate_band_2se_not_plotted": 0.002083827092896988,
343
+ "test_batch_mean": 0.4096857507996082,
344
+ "test_bio_mean": 0.5766222681741072,
345
+ "test_nonrep_mean_map": 0.5094418711383188,
346
+ "rounds": 0,
347
+ "tool_calls": 0,
348
+ "total_tokens": 116361,
349
+ "reasoning_tokens": 96500,
350
+ "canonical_reasoning_tokens": null,
351
+ "reasoning_token_source": "native usage.output_tokens_details.thinking_tokens",
352
+ "reply_errors": 3,
353
+ "truncated_replies": 3,
354
+ "outcome_status": "failed_no_tool_search",
355
+ "plot_included": false,
356
+ "wall_seconds": 815.3,
357
+ "from_call": null,
358
+ "decision": "ABSTAIN",
359
+ "stopped_because": "execution_error",
360
+ "non_default_parameters": null,
361
+ "study": "scenario2_sonnet55_effort_d9a07987_20261031"
362
+ }
363
+ ],
364
+ "provenance": {
365
+ "scenario": "scenario_2",
366
+ "n_runs_per_model_effort": 1,
367
+ "common_efforts": [
368
+ "low",
369
+ "medium",
370
+ "high",
371
+ "xhigh"
372
+ ],
373
+ "model_specific_efforts": {
374
+ "luna": [
375
+ "none"
376
+ ],
377
+ "sonnet": [
378
+ "max"
379
+ ]
380
+ },
381
+ "max_rounds": 8,
382
+ "max_tokens": 32000,
383
+ "timeout_seconds": 900,
384
+ "cores": 16,
385
+ "input_feature_sha256": "65366e8bbcef188def7801e5ef2a54fc7e740a2eedc41a010cfb2647b723604a",
386
+ "input_dimensions": {
387
+ "wells": 14194,
388
+ "plates": 37,
389
+ "features": 1027
390
+ },
391
+ "compound_counts": {
392
+ "calibration": 146,
393
+ "test": 147
394
+ },
395
+ "metadata_parquet_sha256": {
396
+ "luna": "e1c131da3e6c76b7974ab4246964662f2b842baf897b8484acfd2c25f921196f",
397
+ "sonnet": "fa7984a396f23dc2d0b5049e220ddd03d015eeec525d9c5f3d44053ae87e303b"
398
+ },
399
+ "sources": {
400
+ "luna": {
401
+ "path": "/home/xuekun/shape/code/SHAPE-Revision-batch-sonnet55-effort-20261031/benchmarks/batch_correction_JUMP/results/scenario2_effort_e3b9e47b_20261013T013018Z_47584915",
402
+ "report_sha256": "4e7ffc7b9fb5288af1fcc03d99d026d5e46aef4fbd476a798aa45507f86910f5",
403
+ "prepare_sha256": "c30e35b71ced53228ec7224f76807494eb82db40ab2109e5b162d49a9ec20f38"
404
+ },
405
+ "sonnet": {
406
+ "path": "/mnt/filesystem-e3/SHAPE-Revision/benchmarks/batch_correction_JUMP/results/scenario2_sonnet55_effort_d9a07987_20261031",
407
+ "report_sha256": "e3d63f6b82c50f507f7e23bf252663cdc81aee83684f1005ebf6d6ee3d9abdec",
408
+ "prepare_sha256": "1bb787ea23a2e929a6fe427821d86a84590ffb4f4535a7e059e20ebafaaa72e0"
409
+ }
410
+ },
411
+ "task_commits": {
412
+ "luna": "e3b9e47b",
413
+ "sonnet": "d9a07987ce103eb4ffa35bdcd50972cb3d1baa8e"
414
+ },
415
+ "historical_luna_repository_snapshot": "d9a07987ce103eb4ffa35bdcd50972cb3d1baa8e",
416
+ "sonnet_executed_source_sha256": {
417
+ "benchmarks/llm_reasoning/test_native_effort_engine.py": "f5b3fbb17e6a15e517433b0721bd1ce592349311eec1fa7c2c61b8b60d73737c",
418
+ "benchmarks/llm_reasoning/native_effort_engine.py": "d4cfd9c1b2404d4a7cdb6378303e470ed5012fef2eab919fc27f63cd20aa5e5a",
419
+ "benchmarks/llm_reasoning/run_batch_effort.py": "1729ad70da69d9a31fb0b53e5a57c35b4a14ff4b2a5fd50016d9c183d6f03be1"
420
+ },
421
+ "builder_sha256": "415335e03cabf22251760f22f5fbce02a702d178b0b9c7b5533a400e1918be4b",
422
+ "comparison_scope": "Descriptive comparison of existing measured single runs; no repeat-run error bars or significance tests.",
423
+ "version_notes": "Task runner, scorer, control and quantification code are unchanged between the two task commits. The batch verifier difference is an improved null-parameter error message. Package environments differ; metadata Parquet hashes are retained separately. This is not an isolated causal model-effect experiment.",
424
+ "cancelled_repeat_study": {
425
+ "planned_runs": 20,
426
+ "executed_runs": 0,
427
+ "cancelled_by_user": true
428
+ },
429
+ "reply_errors_reviewed": true,
430
+ "native_usage_reconciliation": {
431
+ "low": {
432
+ "response_records": 7,
433
+ "usage_records": 7,
434
+ "thinking_tokens": 0,
435
+ "thinking_usage_coverage": 7,
436
+ "total_tokens": 96837,
437
+ "max_token_truncations": 0,
438
+ "source_sha256": "713b7b11bb121d54792c11cc3b9c8ad4524280f201963defea2537ef462a6961"
439
+ },
440
+ "medium": {
441
+ "response_records": 9,
442
+ "usage_records": 9,
443
+ "thinking_tokens": 0,
444
+ "thinking_usage_coverage": 9,
445
+ "total_tokens": 163416,
446
+ "max_token_truncations": 0,
447
+ "source_sha256": "f5f0f989ae3aee0e206104864b3d032a831cd3bba3a3e409505c92a64a3fecf0"
448
+ },
449
+ "high": {
450
+ "response_records": 8,
451
+ "usage_records": 8,
452
+ "thinking_tokens": 2041,
453
+ "thinking_usage_coverage": 8,
454
+ "total_tokens": 200781,
455
+ "max_token_truncations": 0,
456
+ "source_sha256": "04f04de9d6c82c484b1b02d34e06fc457cbe0a2097ba9bcb9caba4943e694186"
457
+ },
458
+ "xhigh": {
459
+ "response_records": 9,
460
+ "usage_records": 9,
461
+ "thinking_tokens": 10731,
462
+ "thinking_usage_coverage": 9,
463
+ "total_tokens": 359276,
464
+ "max_token_truncations": 0,
465
+ "source_sha256": "fec73d03704e648a0f8ccb5f6c7a48f356f2087e59b7c4b5f4ca9e936b5df106"
466
+ },
467
+ "max": {
468
+ "response_records": 4,
469
+ "usage_records": 4,
470
+ "thinking_tokens": 96500,
471
+ "thinking_usage_coverage": 4,
472
+ "total_tokens": 116361,
473
+ "max_token_truncations": 3,
474
+ "source_sha256": "28de4f069de010c7f4258a6a6f7deee015b0afc58e458ee45d523df1e6203572"
475
+ }
476
+ },
477
+ "usage_note": "Original canonical reports remain unchanged. The initial adapter omitted the provider nested thinking-token field; this derived comparison recovers exact reported values from raw native traces and verifies total tokens. No reasoning-token value is estimated.",
478
+ "failure_policy": "Keep recorded replies, truncations and harness fallback scores in the data. A condition with reply failures and no tool search is marked failed and omitted from normal performance bars; never fabricate a replacement."
479
+ }
480
+ }
logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/luna/PREPARE.json ADDED
@@ -0,0 +1,103 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "scenario": {
3
+ "scenario": "scenario_2",
4
+ "sources": [
5
+ "source_2",
6
+ "source_6",
7
+ "source_10"
8
+ ],
9
+ "plate_types": [
10
+ "TARGET2"
11
+ ],
12
+ "batch_key": "Metadata_Source",
13
+ "label_key": "Metadata_JCP2022",
14
+ "name": "scenario_2"
15
+ },
16
+ "preprocessing": {
17
+ "protocol": "Arevalo-2024-mad_int_featselect",
18
+ "vendor_commit": "cd9bcf99240880a5c9f9858debf70e94f5b4c0f7",
19
+ "fingerprint": {
20
+ "scenario": "scenario_2",
21
+ "raw_inputs": "/scratch/groups/joewu/xuekunwu/shape_benchmark/JUMP/inputs",
22
+ "vendor_sha256": "704ca2c6aea34448a4b805a012a169d7aa1a874b9c68992c1b419375e6575722",
23
+ "microscope_config_sha256": "f67c22d0bb891a3976722c801fd93b7dc07c6fddda5a4c64bd0aef20769235d4",
24
+ "packages": {
25
+ "numpy": "1.26.4",
26
+ "pandas": "2.3.3",
27
+ "pyarrow": "25.0.0",
28
+ "scipy": "1.15.3",
29
+ "scikit-learn": "1.7.2"
30
+ }
31
+ },
32
+ "reused_stages": [
33
+ "raw",
34
+ "neg_stats",
35
+ "variant_feats",
36
+ "mad",
37
+ "mad_int",
38
+ "mad_int_featselect"
39
+ ],
40
+ "artifacts": {
41
+ "raw": {
42
+ "path": "/scratch/groups/joewu/xuekunwu/shape_benchmark/results/batch_correction_JUMP/cache/arevalo/scenario_2/raw.parquet",
43
+ "sha256": "0364225e0550cf56be31ad55dfe8133cf1598509165e6cca2fe47526e35ad6a9"
44
+ },
45
+ "neg_stats": {
46
+ "path": "/scratch/groups/joewu/xuekunwu/shape_benchmark/results/batch_correction_JUMP/cache/arevalo/scenario_2/neg_stats.parquet",
47
+ "sha256": "91e8447b8a866889d25e6bd1973816fb7f45cabdc4115b2c87de251c474248d9"
48
+ },
49
+ "variant_feats": {
50
+ "path": "/scratch/groups/joewu/xuekunwu/shape_benchmark/results/batch_correction_JUMP/cache/arevalo/scenario_2/variant_feats.parquet",
51
+ "sha256": "c8d08ee058a9326fc2ccbe0ac21cee8da3bae867eafa653ebdd4841dbfc794ad"
52
+ },
53
+ "mad": {
54
+ "path": "/scratch/groups/joewu/xuekunwu/shape_benchmark/results/batch_correction_JUMP/cache/arevalo/scenario_2/mad.parquet",
55
+ "sha256": "8430417e1a278ab77d26c5e8c523f83c88ca4e4d18aca75316539ac78d91b0ad"
56
+ },
57
+ "mad_int": {
58
+ "path": "/scratch/groups/joewu/xuekunwu/shape_benchmark/results/batch_correction_JUMP/cache/arevalo/scenario_2/mad_int.parquet",
59
+ "sha256": "ec38a1de7163c580919d45a397854f5fa197e0883d02fbe5556933b22340e70b"
60
+ },
61
+ "mad_int_featselect": {
62
+ "path": "/scratch/groups/joewu/xuekunwu/shape_benchmark/results/batch_correction_JUMP/cache/arevalo/scenario_2/mad_int_featselect.parquet",
63
+ "sha256": "fa02f57e46a4ac6ebb67791faadd7732b9ebebcc56e0cd70aa097310a8c7a82c"
64
+ }
65
+ }
66
+ },
67
+ "raw_inputs": "/scratch/groups/joewu/xuekunwu/shape_benchmark/JUMP/inputs",
68
+ "n_wells": 14194,
69
+ "n_features": 1027,
70
+ "feature_names_head": [
71
+ "Cells_AreaShape_Compactness",
72
+ "Cells_AreaShape_MedianRadius",
73
+ "Cells_AreaShape_Orientation",
74
+ "Cells_AreaShape_Solidity",
75
+ "Cells_AreaShape_Zernike_0_0"
76
+ ],
77
+ "n_compounds": {
78
+ "val": 146,
79
+ "test": 147,
80
+ "poscon": 8
81
+ },
82
+ "n_wells_by_split": {
83
+ "val": 5719,
84
+ "test": 5669,
85
+ "negcon": 2399,
86
+ "poscon": 407
87
+ },
88
+ "n_wells_by_plate_type": {
89
+ "TARGET2": 14194
90
+ },
91
+ "n_wells_by_source": {
92
+ "source_10": 2302,
93
+ "source_2": 3828,
94
+ "source_6": 8064
95
+ },
96
+ "n_plates": 37,
97
+ "n_batches": 29,
98
+ "split_rule": "treatment compounds ordered by sha256('0:<Metadata_JCP2022>'); first half val (calibration), rest test (evaluation); controls (negcon, poscon) in neither",
99
+ "sha256": {
100
+ "features.npy": "65366e8bbcef188def7801e5ef2a54fc7e740a2eedc41a010cfb2647b723604a",
101
+ "meta.parquet": "e1c131da3e6c76b7974ab4246964662f2b842baf897b8484acfd2c25f921196f"
102
+ }
103
+ }
logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/luna/SWEEP_REPORT.json ADDED
@@ -0,0 +1,203 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "gpt-5.6-luna",
3
+ "max_tokens": 32000,
4
+ "timeout_seconds": 900,
5
+ "efforts": [
6
+ "none",
7
+ "low",
8
+ "medium",
9
+ "high",
10
+ "xhigh"
11
+ ],
12
+ "rows": [
13
+ {
14
+ "effort": "none",
15
+ "decision": "RUN",
16
+ "from_call": "agent_13",
17
+ "non_default_parameters": {
18
+ "theta": 4.0,
19
+ "lambda_value": 0.2,
20
+ "max_iter_harmony": 30,
21
+ "input_scale": "zscore",
22
+ "input_pca": 100
23
+ },
24
+ "test_overall": 0.5589373506038878,
25
+ "test_batch_mean": 0.5089071446119035,
26
+ "test_bio_mean": 0.5922908212652108,
27
+ "test_nonrep_mean_map": 0.5652392744438554,
28
+ "test_negcon_mean_map": 0.9685404480333976,
29
+ "agent_vs_harmony_overall": 0.007939896512330069,
30
+ "agent_vs_harmony_overall_2se": 0.0028078987308382108,
31
+ "agent_vs_raw_overall": 0.0490896814322942,
32
+ "agent_vs_raw_overall_2se": 0.003543119605840616,
33
+ "rounds": 3,
34
+ "stopped_because": "satisfied",
35
+ "tool_calls": 13,
36
+ "wall_seconds": 335.9,
37
+ "total_tokens": 67473,
38
+ "reasoning_tokens": 0,
39
+ "best_calibration_overall": 0.5679838610796037,
40
+ "replies": 7,
41
+ "errors": 0,
42
+ "token_limit_errors": 0,
43
+ "reasoning_tokens_mean": 0.0,
44
+ "reasoning_tokens_max": 0,
45
+ "completion_tokens_max": 1424,
46
+ "reply_seconds_mean": 20.93
47
+ },
48
+ {
49
+ "effort": "low",
50
+ "decision": "RUN",
51
+ "from_call": "agent_22",
52
+ "non_default_parameters": {
53
+ "theta": 16.0,
54
+ "lambda_value": 1.0,
55
+ "nclust": 300,
56
+ "max_iter_harmony": 35,
57
+ "block_size": 0.1,
58
+ "input_scale": "zscore",
59
+ "input_pca": 400
60
+ },
61
+ "test_overall": 0.5631142086462773,
62
+ "test_batch_mean": 0.5212980495643695,
63
+ "test_bio_mean": 0.5909916480342158,
64
+ "test_nonrep_mean_map": 0.5543929767588093,
65
+ "test_negcon_mean_map": 0.9690939937679726,
66
+ "agent_vs_harmony_overall": 0.0121167545547195,
67
+ "agent_vs_harmony_overall_2se": 0.0025752097311447313,
68
+ "agent_vs_raw_overall": 0.05326653947468363,
69
+ "agent_vs_raw_overall_2se": 0.0035433480710703484,
70
+ "rounds": 4,
71
+ "stopped_because": "plateau",
72
+ "tool_calls": 27,
73
+ "wall_seconds": 714.6,
74
+ "total_tokens": 89710,
75
+ "reasoning_tokens": 848,
76
+ "best_calibration_overall": 0.5755456241380497,
77
+ "replies": 7,
78
+ "errors": 0,
79
+ "token_limit_errors": 0,
80
+ "reasoning_tokens_mean": 121.1,
81
+ "reasoning_tokens_max": 262,
82
+ "completion_tokens_max": 1725,
83
+ "reply_seconds_mean": 11.45
84
+ },
85
+ {
86
+ "effort": "medium",
87
+ "decision": "RUN",
88
+ "from_call": "agent_22",
89
+ "non_default_parameters": {
90
+ "theta": 6.0,
91
+ "lambda_value": 0.5,
92
+ "nclust": 200,
93
+ "max_iter_harmony": 20,
94
+ "epsilon_harmony": 0.005,
95
+ "input_scale": "zscore",
96
+ "input_pca": 128
97
+ },
98
+ "test_overall": 0.5678742039333458,
99
+ "test_batch_mean": 0.5315159095120535,
100
+ "test_bio_mean": 0.592113066880874,
101
+ "test_nonrep_mean_map": 0.5650862454436287,
102
+ "test_negcon_mean_map": 0.9689527124932255,
103
+ "agent_vs_harmony_overall": 0.016876749841787997,
104
+ "agent_vs_harmony_overall_2se": 0.002743804854260152,
105
+ "agent_vs_raw_overall": 0.058026534761752124,
106
+ "agent_vs_raw_overall_2se": 0.003822075186018169,
107
+ "rounds": 3,
108
+ "stopped_because": "satisfied",
109
+ "tool_calls": 35,
110
+ "wall_seconds": 678.7,
111
+ "total_tokens": 109831,
112
+ "reasoning_tokens": 1621,
113
+ "best_calibration_overall": 0.5783344514027967,
114
+ "replies": 7,
115
+ "errors": 0,
116
+ "token_limit_errors": 0,
117
+ "reasoning_tokens_mean": 231.6,
118
+ "reasoning_tokens_max": 512,
119
+ "completion_tokens_max": 2683,
120
+ "reply_seconds_mean": 14.87
121
+ },
122
+ {
123
+ "effort": "high",
124
+ "decision": "RUN",
125
+ "from_call": "agent_31",
126
+ "non_default_parameters": {
127
+ "theta": 20.0,
128
+ "lambda_value": 0.5,
129
+ "sigma": 0.085,
130
+ "nclust": 500,
131
+ "max_iter_harmony": 40,
132
+ "epsilon_harmony": 0.0001,
133
+ "input_scale": "zscore",
134
+ "input_pca": 800
135
+ },
136
+ "test_overall": 0.5717093741326299,
137
+ "test_batch_mean": 0.549205162791825,
138
+ "test_bio_mean": 0.5867121816931665,
139
+ "test_nonrep_mean_map": 0.5503436432783145,
140
+ "test_negcon_mean_map": 0.9698097743221602,
141
+ "agent_vs_harmony_overall": 0.02071192004107214,
142
+ "agent_vs_harmony_overall_2se": 0.004071871255351384,
143
+ "agent_vs_raw_overall": 0.06186170496103627,
144
+ "agent_vs_raw_overall_2se": 0.004514571244202396,
145
+ "rounds": 3,
146
+ "stopped_because": "satisfied",
147
+ "tool_calls": 36,
148
+ "wall_seconds": 6704.7,
149
+ "total_tokens": 128376,
150
+ "reasoning_tokens": 10082,
151
+ "best_calibration_overall": 0.5817579090862024,
152
+ "replies": 7,
153
+ "errors": 0,
154
+ "token_limit_errors": 0,
155
+ "reasoning_tokens_mean": 1440.3,
156
+ "reasoning_tokens_max": 5012,
157
+ "completion_tokens_max": 7535,
158
+ "reply_seconds_mean": 37.09
159
+ },
160
+ {
161
+ "effort": "xhigh",
162
+ "decision": "RUN",
163
+ "from_call": "agent_27",
164
+ "non_default_parameters": {
165
+ "theta": 6.0,
166
+ "lambda_value": 5.0,
167
+ "sigma": 0.05,
168
+ "nclust": 200,
169
+ "max_iter_harmony": 30,
170
+ "max_iter_kmeans": 10,
171
+ "epsilon_cluster": 0.0001,
172
+ "epsilon_harmony": 0.001,
173
+ "block_size": 0.1,
174
+ "tau": 50.0,
175
+ "input_pca": 100
176
+ },
177
+ "test_overall": 0.5684772870340074,
178
+ "test_batch_mean": 0.5317450306175225,
179
+ "test_bio_mean": 0.5929654579783307,
180
+ "test_nonrep_mean_map": 0.5612766108681524,
181
+ "test_negcon_mean_map": 0.967192566040019,
182
+ "agent_vs_harmony_overall": 0.01747983294244959,
183
+ "agent_vs_harmony_overall_2se": 0.0028938841405390884,
184
+ "agent_vs_raw_overall": 0.05862961786241372,
185
+ "agent_vs_raw_overall_2se": 0.003291797334499316,
186
+ "rounds": 5,
187
+ "stopped_because": "plateau",
188
+ "tool_calls": 83,
189
+ "wall_seconds": 2477.3,
190
+ "total_tokens": 270770,
191
+ "reasoning_tokens": 27823,
192
+ "best_calibration_overall": 0.5771138485467787,
193
+ "replies": 8,
194
+ "errors": 0,
195
+ "token_limit_errors": 0,
196
+ "reasoning_tokens_mean": 3477.9,
197
+ "reasoning_tokens_max": 5752,
198
+ "completion_tokens_max": 8897,
199
+ "reply_seconds_mean": 95.53
200
+ }
201
+ ],
202
+ "rule": "one SHAPE-arm run per level (n = 1); test view; deltas vs harmony / raw on the paired plate halves (+- 2SE); every level the same control, token cap and timeout"
203
+ }
logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/sonnet/DATA_PROVENANCE.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "source_prep": "/mnt/filesystem-e3/SHAPE-Revision/benchmarks/batch_correction_JUMP/results/scenes_39d10dd1/scenario_2/prep",
3
+ "prep": "/mnt/filesystem-e3/SHAPE-Revision/benchmarks/batch_correction_JUMP/results/scenario2_sonnet55_effort_d9a07987_20261031/prep",
4
+ "copied_without_reprocessing": true,
5
+ "n_wells": 14194,
6
+ "n_plates": 37,
7
+ "n_features": 1027,
8
+ "input_hashes": {
9
+ "features.npy": "65366e8bbcef188def7801e5ef2a54fc7e740a2eedc41a010cfb2647b723604a",
10
+ "meta.parquet": "fa7984a396f23dc2d0b5049e220ddd03d015eeec525d9c5f3d44053ae87e303b"
11
+ },
12
+ "source_prepare_sha256": "1bb787ea23a2e929a6fe427821d86a84590ffb4f4535a7e059e20ebafaaa72e0",
13
+ "control_provenance": {
14
+ "tool_version": "1.4.0",
15
+ "tool_sha256": "1c32f0e7272cdf5993e3b9fd1f451117a85524258a36e2e8d7da0ff686ef305d",
16
+ "quant_sha256": "866ca4877fd5d0e46dae20677b457e43593b9fc0ae9c1b9311aed43418ddc83c",
17
+ "control_sha256": "1d3c60ee1bc2685b80976724265b215bf85bd2d9a1461a53042469d1ba7dada2",
18
+ "harmonypy": "2.0.0",
19
+ "scib_metrics": "0.5.10",
20
+ "copairs": "0.5.4",
21
+ "scanorama": "1.7.4"
22
+ },
23
+ "matched_original_control": true
24
+ }
logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/sonnet/PREFLIGHT.json ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "effort": "low",
4
+ "passed": true,
5
+ "transport": "anthropic-native",
6
+ "thinking": "adaptive"
7
+ },
8
+ {
9
+ "effort": "medium",
10
+ "passed": true,
11
+ "transport": "anthropic-native",
12
+ "thinking": "adaptive"
13
+ },
14
+ {
15
+ "effort": "high",
16
+ "passed": true,
17
+ "transport": "anthropic-native",
18
+ "thinking": "adaptive"
19
+ },
20
+ {
21
+ "effort": "xhigh",
22
+ "passed": true,
23
+ "transport": "anthropic-native",
24
+ "thinking": "adaptive"
25
+ },
26
+ {
27
+ "effort": "max",
28
+ "passed": true,
29
+ "transport": "anthropic-native",
30
+ "thinking": "adaptive"
31
+ }
32
+ ]
logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/sonnet/PREPARE.json ADDED
@@ -0,0 +1,96 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "scenario": {
3
+ "scenario": "scenario_2",
4
+ "sources": [
5
+ "source_2",
6
+ "source_6",
7
+ "source_10"
8
+ ],
9
+ "plate_types": [
10
+ "TARGET2"
11
+ ],
12
+ "batch_key": "Metadata_Source",
13
+ "label_key": "Metadata_JCP2022",
14
+ "name": "scenario_2"
15
+ },
16
+ "preprocessing": {
17
+ "protocol": "Arevalo-2024-mad_int_featselect",
18
+ "vendor_commit": "cd9bcf99240880a5c9f9858debf70e94f5b4c0f7",
19
+ "fingerprint": {
20
+ "scenario": "scenario_2",
21
+ "raw_inputs": "/mnt/filesystem-r8/SHAPE-Revision/data/JUMP/inputs",
22
+ "vendor_sha256": "704ca2c6aea34448a4b805a012a169d7aa1a874b9c68992c1b419375e6575722",
23
+ "microscope_config_sha256": "f67c22d0bb891a3976722c801fd93b7dc07c6fddda5a4c64bd0aef20769235d4",
24
+ "packages": {
25
+ "numpy": "1.26.4",
26
+ "pandas": "2.3.3",
27
+ "pyarrow": "25.0.1",
28
+ "scipy": "1.17.1",
29
+ "scikit-learn": "1.9.1"
30
+ }
31
+ },
32
+ "reused_stages": [],
33
+ "artifacts": {
34
+ "raw": {
35
+ "path": "/mnt/filesystem-r8/SHAPE-Revision/benchmarks/batch_correction_JUMP/results/scenes_39d10dd1/cache/arevalo/scenario_2/raw.parquet",
36
+ "sha256": "1519164b2a36e68b9cb056578b3d91280c7330142c57fb1673d4db3306ff9aea"
37
+ },
38
+ "neg_stats": {
39
+ "path": "/mnt/filesystem-r8/SHAPE-Revision/benchmarks/batch_correction_JUMP/results/scenes_39d10dd1/cache/arevalo/scenario_2/neg_stats.parquet",
40
+ "sha256": "e5ea4b1a926493c4d0023283f55fdab87aa7b0369622f093b66bd522370d162a"
41
+ },
42
+ "variant_feats": {
43
+ "path": "/mnt/filesystem-r8/SHAPE-Revision/benchmarks/batch_correction_JUMP/results/scenes_39d10dd1/cache/arevalo/scenario_2/variant_feats.parquet",
44
+ "sha256": "b9504caa59f821b56ec804ad6ba68f3e8b961d2dc95dfbd4e7edfbcc73d6c32f"
45
+ },
46
+ "mad": {
47
+ "path": "/mnt/filesystem-r8/SHAPE-Revision/benchmarks/batch_correction_JUMP/results/scenes_39d10dd1/cache/arevalo/scenario_2/mad.parquet",
48
+ "sha256": "b6ec71c783633cfee50a9e1c37c4a635fd8521f7922d85bcd27c387ab4b4a5df"
49
+ },
50
+ "mad_int": {
51
+ "path": "/mnt/filesystem-r8/SHAPE-Revision/benchmarks/batch_correction_JUMP/results/scenes_39d10dd1/cache/arevalo/scenario_2/mad_int.parquet",
52
+ "sha256": "72158aa5642972ce82bb3c2bd7a5e95d89965aac4538ce34a580e534898b66c4"
53
+ },
54
+ "mad_int_featselect": {
55
+ "path": "/mnt/filesystem-r8/SHAPE-Revision/benchmarks/batch_correction_JUMP/results/scenes_39d10dd1/cache/arevalo/scenario_2/mad_int_featselect.parquet",
56
+ "sha256": "5a33a64f39a084b1bace43d44545939587ae602c247305a585e6cd0d3034ae38"
57
+ }
58
+ }
59
+ },
60
+ "raw_inputs": "/mnt/filesystem-r8/SHAPE-Revision/data/JUMP/inputs",
61
+ "n_wells": 14194,
62
+ "n_features": 1027,
63
+ "feature_names_head": [
64
+ "Cells_AreaShape_Compactness",
65
+ "Cells_AreaShape_MedianRadius",
66
+ "Cells_AreaShape_Orientation",
67
+ "Cells_AreaShape_Solidity",
68
+ "Cells_AreaShape_Zernike_0_0"
69
+ ],
70
+ "n_compounds": {
71
+ "val": 146,
72
+ "test": 147,
73
+ "poscon": 8
74
+ },
75
+ "n_wells_by_split": {
76
+ "val": 5719,
77
+ "test": 5669,
78
+ "negcon": 2399,
79
+ "poscon": 407
80
+ },
81
+ "n_wells_by_plate_type": {
82
+ "TARGET2": 14194
83
+ },
84
+ "n_wells_by_source": {
85
+ "source_10": 2302,
86
+ "source_2": 3828,
87
+ "source_6": 8064
88
+ },
89
+ "n_plates": 37,
90
+ "n_batches": 29,
91
+ "split_rule": "treatment compounds ordered by sha256('0:<Metadata_JCP2022>'); first half val (calibration), rest test (evaluation); controls (negcon, poscon) in neither",
92
+ "sha256": {
93
+ "features.npy": "65366e8bbcef188def7801e5ef2a54fc7e740a2eedc41a010cfb2647b723604a",
94
+ "meta.parquet": "fa7984a396f23dc2d0b5049e220ddd03d015eeec525d9c5f3d44053ae87e303b"
95
+ }
96
+ }
logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/sonnet/STUDY.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "claude-sonnet-5-5",
3
+ "efforts": [
4
+ "low",
5
+ "medium",
6
+ "high",
7
+ "xhigh",
8
+ "max"
9
+ ],
10
+ "max_rounds": 8,
11
+ "max_tokens": 32000,
12
+ "timeout": 900,
13
+ "cores": 16,
14
+ "noise_band": 0.002,
15
+ "source_commit": "d9a07987ce103eb4ffa35bdcd50972cb3d1baa8e",
16
+ "study_source_sha256": {
17
+ "benchmarks/llm_reasoning/test_native_effort_engine.py": "f5b3fbb17e6a15e517433b0721bd1ce592349311eec1fa7c2c61b8b60d73737c",
18
+ "benchmarks/llm_reasoning/native_effort_engine.py": "d4cfd9c1b2404d4a7cdb6378303e470ed5012fef2eab919fc27f63cd20aa5e5a",
19
+ "benchmarks/llm_reasoning/run_batch_effort.py": "1729ad70da69d9a31fb0b53e5a57c35b4a14ff4b2a5fd50016d9c183d6f03be1"
20
+ },
21
+ "prep": "/mnt/filesystem-e3/SHAPE-Revision/benchmarks/batch_correction_JUMP/results/scenario2_sonnet55_effort_d9a07987_20261031/prep",
22
+ "native_effort_field": "output_config.effort",
23
+ "thinking": "adaptive",
24
+ "n_runs_per_effort": 1,
25
+ "heldout_labels_never_sent_to_reasoner": true
26
+ }
logs/reasoning_effort/comparison_scenario2_luna_sonnet_20261031/sonnet/SWEEP_REPORT.json ADDED
@@ -0,0 +1,187 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "claude-sonnet-5-5",
3
+ "max_tokens": 32000,
4
+ "timeout_seconds": 900,
5
+ "efforts": [
6
+ "low",
7
+ "medium",
8
+ "high",
9
+ "xhigh",
10
+ "max"
11
+ ],
12
+ "rows": [
13
+ {
14
+ "effort": "low",
15
+ "decision": "RUN",
16
+ "from_call": "agent_16",
17
+ "non_default_parameters": {
18
+ "theta": 8.0,
19
+ "lambda_value": 0.3,
20
+ "max_iter_harmony": 30,
21
+ "epsilon_harmony": 0.001,
22
+ "input_scale": "zscore",
23
+ "input_pca": 50
24
+ },
25
+ "test_overall": 0.5641117776830196,
26
+ "test_batch_mean": 0.5166596999139929,
27
+ "test_bio_mean": 0.5957464961957041,
28
+ "test_nonrep_mean_map": 0.5719303606516389,
29
+ "test_negcon_mean_map": 0.9690683958195484,
30
+ "agent_vs_harmony_overall": 0.013107532376362085,
31
+ "agent_vs_harmony_overall_2se": 0.0033610760681294432,
32
+ "agent_vs_raw_overall": 0.054264124189495666,
33
+ "agent_vs_raw_overall_2se": 0.004316336346112574,
34
+ "rounds": 4,
35
+ "stopped_because": "satisfied",
36
+ "tool_calls": 19,
37
+ "wall_seconds": 308.9,
38
+ "total_tokens": 96837,
39
+ "reasoning_tokens": null,
40
+ "best_calibration_overall": 0.57239673790964,
41
+ "replies": 7,
42
+ "errors": 0,
43
+ "token_limit_errors": 0,
44
+ "reasoning_tokens_mean": null,
45
+ "reasoning_tokens_max": null,
46
+ "completion_tokens_max": 713,
47
+ "reply_seconds_mean": 4.75
48
+ },
49
+ {
50
+ "effort": "medium",
51
+ "decision": "RUN",
52
+ "from_call": "agent_23",
53
+ "non_default_parameters": {
54
+ "theta": 3.0,
55
+ "nclust": 2000,
56
+ "epsilon_harmony": 0.001,
57
+ "input_pca": 100
58
+ },
59
+ "test_overall": 0.5789328652909089,
60
+ "test_batch_mean": 0.548963537300199,
61
+ "test_bio_mean": 0.5989124172847157,
62
+ "test_nonrep_mean_map": 0.569990239246832,
63
+ "test_negcon_mean_map": 0.9705183107707358,
64
+ "agent_vs_harmony_overall": 0.027928619984251446,
65
+ "agent_vs_harmony_overall_2se": 0.0035681044753392097,
66
+ "agent_vs_raw_overall": 0.06908521179738503,
67
+ "agent_vs_raw_overall_2se": 0.004132520762405466,
68
+ "rounds": 6,
69
+ "stopped_because": "satisfied",
70
+ "tool_calls": 26,
71
+ "wall_seconds": 1354.2,
72
+ "total_tokens": 163416,
73
+ "reasoning_tokens": null,
74
+ "best_calibration_overall": 0.5885191648471167,
75
+ "replies": 9,
76
+ "errors": 0,
77
+ "token_limit_errors": 0,
78
+ "reasoning_tokens_mean": null,
79
+ "reasoning_tokens_max": null,
80
+ "completion_tokens_max": 927,
81
+ "reply_seconds_mean": 4.62
82
+ },
83
+ {
84
+ "effort": "high",
85
+ "decision": "RUN",
86
+ "from_call": "agent_35",
87
+ "non_default_parameters": {
88
+ "theta": 12.0,
89
+ "sigma": 0.07,
90
+ "nclust": 1500,
91
+ "max_iter_kmeans": 20,
92
+ "input_scale": "zscore",
93
+ "input_pca": 50
94
+ },
95
+ "test_overall": 0.5902377036116497,
96
+ "test_batch_mean": 0.5863485084238774,
97
+ "test_bio_mean": 0.592830500403498,
98
+ "test_nonrep_mean_map": 0.5755813568696371,
99
+ "test_negcon_mean_map": 0.9705413342957637,
100
+ "agent_vs_harmony_overall": 0.03923345830499225,
101
+ "agent_vs_harmony_overall_2se": 0.003724520645030581,
102
+ "agent_vs_raw_overall": 0.08039005011812583,
103
+ "agent_vs_raw_overall_2se": 0.004259007918584982,
104
+ "rounds": 5,
105
+ "stopped_because": "satisfied",
106
+ "tool_calls": 40,
107
+ "wall_seconds": 2625.8,
108
+ "total_tokens": 200781,
109
+ "reasoning_tokens": null,
110
+ "best_calibration_overall": 0.5983310172341478,
111
+ "replies": 8,
112
+ "errors": 0,
113
+ "token_limit_errors": 0,
114
+ "reasoning_tokens_mean": null,
115
+ "reasoning_tokens_max": null,
116
+ "completion_tokens_max": 2061,
117
+ "reply_seconds_mean": 9.9
118
+ },
119
+ {
120
+ "effort": "xhigh",
121
+ "decision": "RUN",
122
+ "from_call": "agent_49",
123
+ "non_default_parameters": {
124
+ "theta": 5.0,
125
+ "sigma": 0.06,
126
+ "nclust": 800,
127
+ "max_iter_harmony": 40,
128
+ "epsilon_harmony": 0.0001,
129
+ "input_scale": "zscore",
130
+ "input_pca": 50
131
+ },
132
+ "test_overall": 0.5897806350012761,
133
+ "test_batch_mean": 0.5851601200331521,
134
+ "test_bio_mean": 0.5928609783133587,
135
+ "test_nonrep_mean_map": 0.5755300812664508,
136
+ "test_negcon_mean_map": 0.9674684241319893,
137
+ "agent_vs_harmony_overall": 0.03877638969461861,
138
+ "agent_vs_harmony_overall_2se": 0.003844277805579813,
139
+ "agent_vs_raw_overall": 0.07993298150775219,
140
+ "agent_vs_raw_overall_2se": 0.0044213271571133935,
141
+ "rounds": 6,
142
+ "stopped_because": "satisfied",
143
+ "tool_calls": 63,
144
+ "wall_seconds": 6486.9,
145
+ "total_tokens": 359276,
146
+ "reasoning_tokens": null,
147
+ "best_calibration_overall": 0.5982290475332367,
148
+ "replies": 9,
149
+ "errors": 0,
150
+ "token_limit_errors": 0,
151
+ "reasoning_tokens_mean": null,
152
+ "reasoning_tokens_max": null,
153
+ "completion_tokens_max": 5417,
154
+ "reply_seconds_mean": 22.96
155
+ },
156
+ {
157
+ "effort": "max",
158
+ "decision": "ABSTAIN",
159
+ "from_call": null,
160
+ "non_default_parameters": null,
161
+ "test_overall": 0.5098476612243077,
162
+ "test_batch_mean": 0.4096857507996082,
163
+ "test_bio_mean": 0.5766222681741072,
164
+ "test_nonrep_mean_map": 0.5094418711383188,
165
+ "test_negcon_mean_map": 0.9461091695250572,
166
+ "agent_vs_harmony_overall": -0.04115658408234979,
167
+ "agent_vs_harmony_overall_2se": 0.002083827092896988,
168
+ "agent_vs_raw_overall": 7.730783790904638e-09,
169
+ "agent_vs_raw_overall_2se": 5.979338128313691e-08,
170
+ "rounds": 0,
171
+ "stopped_because": "execution_error",
172
+ "tool_calls": 0,
173
+ "wall_seconds": 815.3,
174
+ "total_tokens": 116361,
175
+ "reasoning_tokens": null,
176
+ "best_calibration_overall": null,
177
+ "replies": 4,
178
+ "errors": 3,
179
+ "token_limit_errors": 0,
180
+ "reasoning_tokens_mean": null,
181
+ "reasoning_tokens_max": null,
182
+ "completion_tokens_max": 862,
183
+ "reply_seconds_mean": 171.44
184
+ }
185
+ ],
186
+ "rule": "one SHAPE-arm run per level (n = 1); test view; deltas vs harmony / raw on the paired plate halves (+- 2SE); every level the same control, token cap and timeout"
187
+ }