nur-dev commited on
Commit
e7a0400
·
verified ·
1 Parent(s): 2d436d4

Add reviewer-requested controlled and sensor mechanism tests

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. README.md +38 -0
  2. SHA256SUMS.json +117 -2
  3. revision_r1/controlled/compute.json +33 -0
  4. revision_r1/controlled/family_sample_frontier.csv +25 -0
  5. revision_r1/controlled/final_audit.json +34 -0
  6. revision_r1/controlled/frontier_all_conditions.csv +0 -0
  7. revision_r1/controlled/indices.json +466 -0
  8. revision_r1/controlled/native_worlds.csv.gz +3 -0
  9. revision_r1/controlled/numeric_compute.json +9 -0
  10. revision_r1/controlled/numeric_diagnostic.json +154 -0
  11. revision_r1/controlled/numeric_push_sensitivity.csv.gz +3 -0
  12. revision_r1/controlled/paired_by_condition.csv +0 -0
  13. revision_r1/controlled/privileged_geometry.csv +0 -0
  14. revision_r1/controlled/reference_replay_discrepancies.csv.gz +3 -0
  15. revision_r1/controlled/replay_discrepancies.csv.gz +3 -0
  16. revision_r1/controlled/run.py +572 -0
  17. revision_r1/controlled/sample_frontier.csv +5 -0
  18. revision_r1/controlled/specification.json +108 -0
  19. revision_r1/controlled/specification_01.json +88 -0
  20. revision_r1/controlled/specification_02.json +96 -0
  21. revision_r1/controlled/summary.json +949 -0
  22. revision_r1/controlled/timing.json +0 -0
  23. revision_r1/external/article_table.csv +20 -0
  24. revision_r1/external/execution.json +24 -0
  25. revision_r1/external/implementation_correction.json +14 -0
  26. revision_r1/external/metrics.csv.gz +3 -0
  27. revision_r1/external/overall.json +0 -0
  28. revision_r1/external/paired_contrasts.json +1338 -0
  29. revision_r1/external/run.py +515 -0
  30. revision_r1/external/shift_by_task.json +743 -0
  31. revision_r1/external/shift_overall.json +102 -0
  32. revision_r1/external/specification.json +123 -0
  33. revision_r1/external/table_provenance.json +10 -0
  34. revision_r1/sensor/analysis_spec.json +79 -0
  35. revision_r1/sensor/export_safe.py +39 -0
  36. revision_r1/sensor/findings.json +123 -0
  37. revision_r1/sensor/full_replay.json +39 -0
  38. revision_r1/sensor/objective_extension.py +63 -0
  39. revision_r1/sensor/objective_extension_01/cell_compute.json +146 -0
  40. revision_r1/sensor/objective_extension_01/compute.json +17 -0
  41. revision_r1/sensor/objective_extension_01/hydraulic_n128_s20751.json +1 -0
  42. revision_r1/sensor/objective_extension_01/hydraulic_n128_s20751.safetensors +3 -0
  43. revision_r1/sensor/objective_extension_01/hydraulic_n128_s20752.json +1 -0
  44. revision_r1/sensor/objective_extension_01/hydraulic_n128_s20752.safetensors +3 -0
  45. revision_r1/sensor/objective_extension_01/hydraulic_n128_s20753.json +1 -0
  46. revision_r1/sensor/objective_extension_01/hydraulic_n128_s20753.safetensors +3 -0
  47. revision_r1/sensor/objective_extension_01/hydraulic_n64_s20751.json +1 -0
  48. revision_r1/sensor/objective_extension_01/hydraulic_n64_s20751.safetensors +3 -0
  49. revision_r1/sensor/objective_extension_01/hydraulic_n64_s20752.json +1 -0
  50. revision_r1/sensor/objective_extension_01/hydraulic_n64_s20752.safetensors +3 -0
README.md CHANGED
@@ -63,3 +63,41 @@ An action-value compression control, a complete 1,212-episode partition and norm
63
  ## Data and license
64
 
65
  Code is MIT; original models, generated data and research outputs are CC BY 4.0. Hydraulic features derive from Helwig, Pignanelli and Schuetze (UCI DOI 10.24432/C5CW21); temperature-modulated sensor features derive from Burgues (10.24432/C5S302). Gas-drift recordings (Vergara, 10.24432/C5RP6W) are retrieved separately and are not redistributed; preserve the source's legacy research-only notice. Full attribution and terms are in LICENSE.md.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
63
  ## Data and license
64
 
65
  Code is MIT; original models, generated data and research outputs are CC BY 4.0. Hydraulic features derive from Helwig, Pignanelli and Schuetze (UCI DOI 10.24432/C5CW21); temperature-modulated sensor features derive from Burgues (10.24432/C5S302). Gas-drift recordings (Vergara, 10.24432/C5RP6W) are retrieved separately and are not redistributed; preserve the source's legacy research-only notice. Full attribution and terms are in LICENSE.md.
66
+
67
+ ## Reviewer-revision experiments
68
+
69
+ `revision_r1/` adds exploratory mechanism tests on the already opened data:
70
+ full controlled action-value/loss comparisons, external sample-support and
71
+ fixed-policy transfer comparisons, and 318 sensor fits with all seeds and
72
+ budgets retained. Original results, protocols and promotion decisions are unchanged.
73
+ The 27 sensor state bundles use only Safetensors and JSON. Specifications,
74
+ all comparative scores, cost records and numerical tie-sensitivity diagnostics
75
+ are included; these are additional analyses, not new confirmation.
76
+
77
+ Replay every added sensor state without fitting:
78
+
79
+ ```bash
80
+ python revision_r1/sensor/verify_all.py --release . --base revision_r1/sensor/run_01 --extension revision_r1/sensor/objective_extension_01 --output revision_sensor_replay.json --device cuda:0
81
+ ```
82
+
83
+ Recompute the controlled interventions in a new directory:
84
+
85
+ ```bash
86
+ python revision_r1/controlled/run.py --prepare --release . --output new_controlled --device cuda:0
87
+ python revision_r1/controlled/run.py --release . --output new_controlled --device cuda:0
88
+ ```
89
+
90
+ The external analysis requires separately acquired gas-drift recordings. Retain
91
+ the provider's research-use notice; no transformed gas records are redistributed.
92
+
93
+ ```bash
94
+ python rebuild_external.py --output new_external_source --device cuda:0
95
+ python revision_r1/external/run.py specify --release . --input new_external_source --output new_external_analysis --device cuda:0
96
+ python revision_r1/external/run.py run --release . --input new_external_source --output new_external_analysis --device cuda:0
97
+ ```
98
+
99
+ Sensor refitting is optional through `revision_r1/sensor/review_r1_sensor.py`
100
+ and `objective_extension.py`, each accepting `--release .`, `--output` and
101
+ `--device`. Use a new output directory. Rank gains do not establish calibrated
102
+ acceptance, and sparse fitted-law ties can affect ancillary realized utilities;
103
+ the unchanged primary mixture frontiers and all disclosed discrepancies remain available.
SHA256SUMS.json CHANGED
@@ -2,7 +2,7 @@
2
  ".gitattributes": "0aad2046ff3535fdb3750ad77100b4f6aed80009c097ffc8a9e264a5d1b0c4e7",
3
  "ENVIRONMENT.json": "786e73bb84b1300d1b5062961af74aacbab2bbfe1e12675ba6ceebd12d26b0eb",
4
  "LICENSE.md": "4a8b21f5c8ee1383fdf12d8622df51066334d446513d7318ab8a31763f435813",
5
- "README.md": "f6be7fca2e03477cf296dfc2c4633d934a32eac491e838ef1b7f46cef2a4ec75",
6
  "analyze.py": "3e2ee19da7cdec5b735db55a368132977d8aa6cfe4cfd8e49dbe86019a56a444",
7
  "data/controlled/manifest.json": "2b50155e7f4d198449e2499dad21ba82d51e94423a1711eca8de4a01ab9ca156",
8
  "data/controlled/worlds.json": "860e154c4ffd0690080ab8c263756d4340cc119fffa66e1cc41d1b42ac3848cb",
@@ -356,5 +356,120 @@
356
  "unionmind/predictive_transfer/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
357
  "unionmind/predictive_transfer/data.py": "259202796bf921db2beaa75166809e455e44aaddef73da96bdec970067abfbd6",
358
  "unionmind/supported_observations/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
359
- "unionmind/supported_observations/data.py": "d359f27c4cd23482ed4ef8326d3026c10b19caf9ba75b3da677249f5d9c9e43f"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
360
  }
 
2
  ".gitattributes": "0aad2046ff3535fdb3750ad77100b4f6aed80009c097ffc8a9e264a5d1b0c4e7",
3
  "ENVIRONMENT.json": "786e73bb84b1300d1b5062961af74aacbab2bbfe1e12675ba6ceebd12d26b0eb",
4
  "LICENSE.md": "4a8b21f5c8ee1383fdf12d8622df51066334d446513d7318ab8a31763f435813",
5
+ "README.md": "ba4c2be9b705ec3b49377003e9b9d6819efd8b8c654dd9a4eb125ac1fb7fb389",
6
  "analyze.py": "3e2ee19da7cdec5b735db55a368132977d8aa6cfe4cfd8e49dbe86019a56a444",
7
  "data/controlled/manifest.json": "2b50155e7f4d198449e2499dad21ba82d51e94423a1711eca8de4a01ab9ca156",
8
  "data/controlled/worlds.json": "860e154c4ffd0690080ab8c263756d4340cc119fffa66e1cc41d1b42ac3848cb",
 
356
  "unionmind/predictive_transfer/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
357
  "unionmind/predictive_transfer/data.py": "259202796bf921db2beaa75166809e455e44aaddef73da96bdec970067abfbd6",
358
  "unionmind/supported_observations/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
359
+ "unionmind/supported_observations/data.py": "d359f27c4cd23482ed4ef8326d3026c10b19caf9ba75b3da677249f5d9c9e43f",
360
+ "revision_r1/sensor/primary_table.csv": "4503de189e284a9c5a31bc144c3bdff69e75fdbf2bfd8a5d434bf74f7128e55b",
361
+ "revision_r1/sensor/safe_replay.json": "32c12d1bbd1e0161d272b9632c3a6e5d55953c92c9e53bdc3c9919325c8731c2",
362
+ "revision_r1/sensor/review_r1_sensor.py": "22e61a23b610a1399e478499e932bea5b79e3aa30603d8aa19e660562d8ec037",
363
+ "revision_r1/sensor/findings.json": "4aff5cf1f2f76094fd170e097e8ed6aa8b40d9199c35eb6527d2787cd69f2c58",
364
+ "revision_r1/sensor/full_replay.json": "eb948f80c7b116a7c174ebd63f64a1c4c16eb1f69a60cee61a8fd76f307e6286",
365
+ "revision_r1/sensor/analysis_spec.json": "4644bcb2ed1a9f152d6b5e74be0dafc339a3ae7ccbfae7aa0d79d9f4caa4ad02",
366
+ "revision_r1/sensor/verify_all.py": "5316a84ede79929187ffb31483e681c5860be267720e6920f15d029f5ccfef61",
367
+ "revision_r1/sensor/safe_tree.py": "4e89c1560a55fc91cd800fe88d3a8de8b6f1951effb4ce068b23ad808355853e",
368
+ "revision_r1/sensor/objective_extension_spec.json": "ea8ccb96ecf1b94e7c9cf68211428b23f6adf3de13653e38e9b61ce7c83cd18f",
369
+ "revision_r1/sensor/export_safe.py": "39a447b022689fd6d4ac8441d5c83dafe5d66136a632e80d2c7e244d2bd36a64",
370
+ "revision_r1/sensor/objective_extension.py": "bd24384d8512572d6cf880748a98714482f90c8da4dc0b14c75b4cfe53b527f3",
371
+ "revision_r1/controlled/numeric_push_sensitivity.csv.gz": "655e73bc4111c4c5dab7507e4fd029cd5fba7efdde6375757a6b281c94027358",
372
+ "revision_r1/controlled/summary.json": "903a81f9440772657a9590911e772ce2beba514a2d87f0b4789a8a0915742b0d",
373
+ "revision_r1/controlled/replay_discrepancies.csv.gz": "6ef693532219b9ba0abab9aaafcc07558a72a20f0ffe961b4bc42a138e873d25",
374
+ "revision_r1/controlled/reference_replay_discrepancies.csv.gz": "59fad6d56023fedd16c50f0faab3b4f85a8cb322c3ecbcb4f7d5f9c7e5076051",
375
+ "revision_r1/controlled/native_worlds.csv.gz": "1390af820583c90646ceb5edd7da2647ea0b9d8aeedfd516c921cdeea1b58581",
376
+ "revision_r1/controlled/specification.json": "d361b91b341dd596b246932f9113a0245adafcf5dcdce7b4e0f340d364b7b9da",
377
+ "revision_r1/controlled/privileged_geometry.csv": "8acd6348b334b91b9f9b0486a9e07b9ea4cd1eee71d4eb4124cbc8cfe7bca233",
378
+ "revision_r1/controlled/paired_by_condition.csv": "e432889ff30f13909602467258b211fa3f0248a11c7cd875b957d2942a3704af",
379
+ "revision_r1/controlled/frontier_all_conditions.csv": "5b5b9e761163a7f1800047503cc0eae9d9d3450e221161daad14f6a44ffa704d",
380
+ "revision_r1/controlled/run.py": "4a300dfc769473f5afb08e15e877f6c53dfff5f6d3a0abfe17e0043798eecbab",
381
+ "revision_r1/controlled/final_audit.json": "f3c2daa7eba8ddd2a364ee3cc363de6510c5fb0171d88db5b71480530c9d44a1",
382
+ "revision_r1/controlled/compute.json": "b07dc17453b5f7dd00512913def5dd5fb00d3244d8188ed159ff09f556364823",
383
+ "revision_r1/controlled/specification_02.json": "d0a57d8c485cb9ca380095c8ed018834d0df79382d9cffcc0709bc747129f9eb",
384
+ "revision_r1/controlled/timing.json": "852f0d26180eaebe616841f03894646a4252f57214a81856df984eccb8f99947",
385
+ "revision_r1/controlled/numeric_diagnostic.json": "cdd2528c9f74b817a1b006b75d16c80be2b8f8ed236cd22927f971e14afcfa89",
386
+ "revision_r1/controlled/family_sample_frontier.csv": "793bd7f0fe0da1a09dc6d89572d36f2afc8aeca149094942959ad4c13f441fb8",
387
+ "revision_r1/controlled/indices.json": "dae400dfe60ac990d206d0bf943e276e5ead6a454d317e34ae6ed57e0ad14e45",
388
+ "revision_r1/controlled/specification_01.json": "d820d635c001d28e9a884eaa1a5040dfacf9829ddb2387b301a13d07d33bbbc2",
389
+ "revision_r1/controlled/sample_frontier.csv": "f51b0fd71c1ffd9e14caaaa30a1677705190c08f02227ad72e5548ce33c4e12c",
390
+ "revision_r1/controlled/numeric_compute.json": "d6eda6bae1d8a6d023ccdb7d2e0677b3e409c0fc49d5e56f51101effcee39d08",
391
+ "revision_r1/external/shift_overall.json": "52aff394a2116b9b507482d7fdc96168036c0275db4f30f40f2d0671a7cb9965",
392
+ "revision_r1/external/execution.json": "0fae50882a58a2a589c673a9c6ae2e2ec83d464571bb48dfa2b0c8ada1924c6d",
393
+ "revision_r1/external/specification.json": "541dd0c1d4d7e242ee20c797bd9ef0928f7afdfffa4aacfeca5c3205cb212fab",
394
+ "revision_r1/external/shift_by_task.json": "92034faf5fe6f94508cb47cf4f4f6e207c9143a49ebe8f3ed312d2e29cb01c46",
395
+ "revision_r1/external/implementation_correction.json": "beb720147e0083ef857ebcb745c1403ccdf083c44ba70f6ed82d8a6d3e9e325e",
396
+ "revision_r1/external/metrics.csv.gz": "9813ff30348ad76cb8b672e022a5f3fe852b8f0ee13625e7a5799400cfc01120",
397
+ "revision_r1/external/run.py": "33e4d7ff42763ccd48db10013dd91831af13aa368a115c39987f65636a06c63f",
398
+ "revision_r1/external/paired_contrasts.json": "b1afae2e0ee103a5f9c65afb49781037e7b3d24e6427e890821931fe640cf3e5",
399
+ "revision_r1/external/overall.json": "af8fefc215e51f30684cdfefade17eac81f99a18302a864c5a9b48cf62c927a7",
400
+ "revision_r1/external/article_table.csv": "09b7110cc185574d01f5aeb4681aa20910048a50429b274a5422969c2654fad2",
401
+ "revision_r1/external/table_provenance.json": "d538e18c7367c0b27778d8a5cdf8a1cfdb1ecdc73a6906faa1980d20bb46b406",
402
+ "revision_r1/sensor/run_01/hydraulic_reversal_s20752.json": "3bf241d1167e33c19f17027ccdd39c8c96082c2977a67e9e8092928ba4dab2f0",
403
+ "revision_r1/sensor/run_01/hydraulic_n64_s20752.json": "ee4c987ec56a2d57909f89febb948f30ad1a13608c01dcebf9f304a4f7ba14e8",
404
+ "revision_r1/sensor/run_01/hydraulic_n64_s20753.json": "5105b5fcc456ddc4b1a6e046d9c63c60a4f080f1823da476870d2020fb6ee954",
405
+ "revision_r1/sensor/run_01/temperature_n128_s20853.json": "c67145613fcb17cf16ada5ec71cd3d127576e590759dacd6a4a5d36f35a05e93",
406
+ "revision_r1/sensor/run_01/hydraulic_n64_s20751.json": "461cded6ade8630bc2652e2e7bab15471f07d124eee50a44a1246611b01a8f46",
407
+ "revision_r1/sensor/run_01/cell_compute.json": "fd139f0c1ff8a284591455fdb2a61043ccc966c631df0d74c1360a5aa2e8c444",
408
+ "revision_r1/sensor/run_01/hydraulic_n128_s20752.safetensors": "7a2759068a289b615624b001dcd0115cd79cf6c9609b30fcc7529846379f18db",
409
+ "revision_r1/sensor/run_01/hydraulic_n128_s20753.json": "6783f3aa43c05882b5c46f42bbbf035945db49e4834a3f65e91ba63f858e4433",
410
+ "revision_r1/sensor/run_01/reversal_diagnostics.json": "7ec82dc059717d2f382c924c8b29682c465c26508475a0955ea3576bfb3c8cd6",
411
+ "revision_r1/sensor/run_01/hydraulic_reversal_s20753.safetensors": "6fd698aa04579d0beb6ca3cdc08d5b8fd1f6edcb2991d062d65d74e86c0933e7",
412
+ "revision_r1/sensor/run_01/metrics.csv": "5aa8f4b6980113a976f79c9ee1d2acd2af9c545b06ce15416fad37f4337c5473",
413
+ "revision_r1/sensor/run_01/temperature_n128_s20853.safetensors": "8fce6292bf5d5740855525d26ac91e3363e0d2740f1271c9d56309ea0fed1d98",
414
+ "revision_r1/sensor/run_01/hydraulic_n128_s20752.json": "d0c21c696f9ee7e4eec30c19b1a797b401b185d9507a25922ff4d07e23014506",
415
+ "revision_r1/sensor/run_01/hydraulic_n128_s20751.safetensors": "2fda58bcf415701c8a2231e46fb0253e7e4aeb72b5bf296ad8b3ae4326c9eef4",
416
+ "revision_r1/sensor/run_01/hydraulic_reversal_s20753.json": "a2156b5048ee86e6022495b2e464c5c89d98db03757d9f70cd9d43a1777bf9ad",
417
+ "revision_r1/sensor/run_01/temperature_n64_s20852.safetensors": "77c7972c39b6f405d53686da88fafa680d6aec2303b175e1f9c858faf392f126",
418
+ "revision_r1/sensor/run_01/reversal_metrics.csv": "a8ebb206fc609e78b3ad8e9c327224dcc693d37b264bfdb0c1edd83e92750b7d",
419
+ "revision_r1/sensor/run_01/calibration_choices.csv": "7b822a9e32b6445168c22255b2e6e5789ecf2e8dfc857f2027dbce3e188ab6da",
420
+ "revision_r1/sensor/run_01/temperature_n128_s20851.safetensors": "00973079adcb876833d0af0f9399ca70a165639f99a225f44abb3292f0b914d2",
421
+ "revision_r1/sensor/run_01/hydraulic_n128_s20751.json": "8e91c7b827846c153095b14b8fed42028443f816f56856c06e26ff734be3c7fd",
422
+ "revision_r1/sensor/run_01/temperature_n64_s20851.safetensors": "6c64fb224e027bb5cab760fd37725567c4b470caff93a0384f6905aa02b4bdc1",
423
+ "revision_r1/sensor/run_01/temperature_n64_s20853.safetensors": "51819ee75d951c15106f03eb08132c305886b9f3fe44155e71efc5b7e0cd50a0",
424
+ "revision_r1/sensor/run_01/compute.json": "8f0adcf124145f1b91191b00ee452df2fb7679ba57f597f75cbd12511b74b11d",
425
+ "revision_r1/sensor/run_01/hydraulic_reversal_s20751.json": "6451cea208f09bd9e541ffaf8b3d1d112e5e6c153e5581ca354b3a98aa612f74",
426
+ "revision_r1/sensor/run_01/temperature_n64_s20851.json": "5c3cbc5181a479e94c9a3fa37f55d171619ae4f72fb3e3ce7d7b1660a3cc70cf",
427
+ "revision_r1/sensor/run_01/numerical_tests.json": "61759ebaf9ea643356170c39e30dfe666c82aa5cdf3146cb1b946e161f960edb",
428
+ "revision_r1/sensor/run_01/temperature_n128_s20852.json": "ee2fb82904fa151f1cc6b7b402700cbb5ca3e9257c377bd57d60e5658b59ef07",
429
+ "revision_r1/sensor/run_01/temperature_n64_s20852.json": "6251d047615ae184d78e974bb6d7ed67f148523fccd9970e7e9f9e3de1556530",
430
+ "revision_r1/sensor/run_01/summary.csv": "6a87bbb63231c3d2680e5f6329dea91bd4819016450cd6264633d3c9be9eda0b",
431
+ "revision_r1/sensor/run_01/reversal_summary.csv": "d9b2e591d90c92e202ef4278cf23c9dc01ed5829abde75174a826eddac18c91f",
432
+ "revision_r1/sensor/run_01/temperature_n128_s20852.safetensors": "dbf86d345e884c3ebe5729873f30cadedea12db9103c1106e08e8cb7098bcaec",
433
+ "revision_r1/sensor/run_01/hydraulic_n64_s20752.safetensors": "bffa83502f802cc8cb4475ebddef49479380fbf90114b744f734a8866b979813",
434
+ "revision_r1/sensor/run_01/replay_verification.json": "5cc75dafd0322febcee4ab1e4f2722423d24b5c8d4dfd8b9788639080e682c41",
435
+ "revision_r1/sensor/run_01/hydraulic_n64_s20753.safetensors": "a0bee153eeabb6ad7346aea4a0872c5b829b7e76b309b643d52c2412373ff3a9",
436
+ "revision_r1/sensor/run_01/hydraulic_reversal_s20751.safetensors": "b360e8dc3a93fe93ceba9e7bdbeb23fd5a1489934c1657bf9747a7152c0c6bc6",
437
+ "revision_r1/sensor/run_01/mean_diagnostics.csv": "66918f9e476716271f022e8a09ca25baaa40a33f63211ddaef2cc18584ca8d0f",
438
+ "revision_r1/sensor/run_01/run_manifest.json": "e1c73045c497d48a4765b6bdd6e99f29080d81dd585be03259877d20c35aa81a",
439
+ "revision_r1/sensor/run_01/temperature_n128_s20851.json": "95ce9f0bd3833eb99aec7ea044e033d6cbcd1846e13d76ed7c4bf545514dcaa2",
440
+ "revision_r1/sensor/run_01/hydraulic_n128_s20753.safetensors": "a0dedab19716b992c8a8fa5d6cc87de17ba472f88c27a3e8e2f24e98cb7b39fb",
441
+ "revision_r1/sensor/run_01/temperature_n64_s20853.json": "82e29ee85f769acfca702b26404d4b887e706e01a10d1c6b2642d71641481d42",
442
+ "revision_r1/sensor/run_01/hydraulic_n64_s20751.safetensors": "cbebd95806651b704991c11d0903046cb174d076b6eae52fb9fe72e946534d93",
443
+ "revision_r1/sensor/run_01/hydraulic_reversal_s20752.safetensors": "ee84572929b830064ff9e92e01ba503ddc2bbb359e6052c62b1f8ce1ec956303",
444
+ "revision_r1/sensor/objective_extension_01/hydraulic_n64_s20752.json": "e86e470016ba09d6f7da725a0f352118cb6012652caa4ea06e5b2921338b863e",
445
+ "revision_r1/sensor/objective_extension_01/hydraulic_n64_s20753.json": "8292642abe36e112830007a9a68d987bb09ab0fb738e50f99fb2e8d455593618",
446
+ "revision_r1/sensor/objective_extension_01/temperature_n128_s20853.json": "12a70593382be57116c35034ae0b9d22646a9acbfead91bf75015f85d6129ec4",
447
+ "revision_r1/sensor/objective_extension_01/hydraulic_n64_s20751.json": "026c7b03639d791649caaca18f9da894b114ae82905af552a10a7c89a49a2283",
448
+ "revision_r1/sensor/objective_extension_01/cell_compute.json": "041df92042872985572d039dc831ac4e31a89b963ef298a05ae559681f0e7dea",
449
+ "revision_r1/sensor/objective_extension_01/hydraulic_n128_s20752.safetensors": "84e37c6d7a7ec02217682fedad56211ef67be4757a0376682fc6428f70f92f61",
450
+ "revision_r1/sensor/objective_extension_01/summary.json": "39e95464252c3e3e948a6418ca33a77eedc236c672cbf28087ecec249650d6b8",
451
+ "revision_r1/sensor/objective_extension_01/hydraulic_n128_s20753.json": "51bbf95853904a4fcc7d51b9de5b1695a0c107df4a6ea6c18e818afa8c2630f2",
452
+ "revision_r1/sensor/objective_extension_01/metrics.csv": "714c629729eec808ced1d3b281989182d7a28b1299fb0d4f51f13b76dedce22d",
453
+ "revision_r1/sensor/objective_extension_01/temperature_n128_s20853.safetensors": "0e056e512d4ee1f5a3efe4344290d4668a8b3aa71f2e840ad1ab9276b76f1851",
454
+ "revision_r1/sensor/objective_extension_01/hydraulic_n128_s20752.json": "5891441524bac20f98801f92155f8f4a4fd55d01ee4135a3a452cd712a8761d0",
455
+ "revision_r1/sensor/objective_extension_01/hydraulic_n128_s20751.safetensors": "d7a21511d36785300dd873f79aadcec87bf8de59a73b621328e26cc45a137bda",
456
+ "revision_r1/sensor/objective_extension_01/temperature_n64_s20852.safetensors": "5ef676925efe8be2d36c8d26d21c874409db2547074ef99010628f0de46d0ae5",
457
+ "revision_r1/sensor/objective_extension_01/manifest.json": "74aa43959c81bf2cfc5b8f7e10653e05365250216451ee320e2964dc0a6b1733",
458
+ "revision_r1/sensor/objective_extension_01/temperature_n128_s20851.safetensors": "dfc1891afa7d41b2e26e19348598c522070cd20223d1a70a9e72030ed801e2a9",
459
+ "revision_r1/sensor/objective_extension_01/hydraulic_n128_s20751.json": "a4bfa9d2dfa4c95b8eca35b1edd0ec4d21988f8bb9d2b088b9553daafb72c3f4",
460
+ "revision_r1/sensor/objective_extension_01/temperature_n64_s20851.safetensors": "8e93164e9f58e98154ccd2bf69a3fd8323f9615b63ceae62a74dc5997b4dc3db",
461
+ "revision_r1/sensor/objective_extension_01/temperature_n64_s20853.safetensors": "3967dd4e5ba220d02fdc5808f3f6b08dcf643e43f39d4b70d9eaad9eecef3d88",
462
+ "revision_r1/sensor/objective_extension_01/compute.json": "c8daa706c9c4222737008ec8efd5ae98ecd921bc11642580d5821cf0da0960a5",
463
+ "revision_r1/sensor/objective_extension_01/temperature_n64_s20851.json": "95e4dedaa248d4d30912e2633f4e5d3f8a2846f9dc68682bec94a2206da61d4d",
464
+ "revision_r1/sensor/objective_extension_01/temperature_n128_s20852.json": "8d84d08589d246c2e413820e74572b96b16fdbac3dd810902319cd2b4e9611f6",
465
+ "revision_r1/sensor/objective_extension_01/temperature_n64_s20852.json": "b259c61869934fa7d19b37e154c669ccc39d19318ce7061dd53eea62a94e127f",
466
+ "revision_r1/sensor/objective_extension_01/summary.csv": "db94a3a67be24e9a3d7fae4fb8d475f329b78df7bfb528e9ce7600f32d220092",
467
+ "revision_r1/sensor/objective_extension_01/temperature_n128_s20852.safetensors": "d5fad93647f548d3385f5c946987d64d7aa78fd7fa87f03a884b0a5258cfb8cd",
468
+ "revision_r1/sensor/objective_extension_01/hydraulic_n64_s20752.safetensors": "e4ff49f92f19eeb70adf1a6c200eb238839153310e435c938951a6658adfbc7c",
469
+ "revision_r1/sensor/objective_extension_01/hydraulic_n64_s20753.safetensors": "e33317df751bf702ad4cc096ffe209c450ec2783c5f6c0e318b10f57d29f5879",
470
+ "revision_r1/sensor/objective_extension_01/summary_per_seed.csv": "8399b7bec3e4ea9c62ed2786c937d7a44495fed7982cdf0dfebc4f10112c0104",
471
+ "revision_r1/sensor/objective_extension_01/temperature_n128_s20851.json": "2f2e1c580bce42b35e93c8537b2fd219cc3b6de071df323560479270ae7bd108",
472
+ "revision_r1/sensor/objective_extension_01/hydraulic_n128_s20753.safetensors": "fb1eb8085bbd37b334f4a7033d7b0ca0863981f0263528ccced9ada34167dace",
473
+ "revision_r1/sensor/objective_extension_01/temperature_n64_s20853.json": "398edf09511ff818a3bb2e77829b121e5b1af9b1646eb29507556f6d756a5cba",
474
+ "revision_r1/sensor/objective_extension_01/hydraulic_n64_s20751.safetensors": "3e5a56962a870d83d2cd4037f8ec1c0aa39fe0b20111579bda7d435a98683f28"
475
  }
revision_r1/controlled/compute.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "created_utc": "2026-11-09T14:43:53.549254+00:00",
3
+ "status": "COMPLETED",
4
+ "source_release_files_hash_verified": 358,
5
+ "code_sha256": "aa82ccdc6a1bb3c7d1e92b9ab0c4c88478c7d26e0442fe91523036a63260b2aa",
6
+ "specification_sha256": "d0a57d8c485cb9ca380095c8ed018834d0df79382d9cffcc0709bc747129f9eb",
7
+ "command": [
8
+ "articles/make_1/reproduce/review_r1_controlled.py",
9
+ "--resume-aggregation",
10
+ "--device",
11
+ "cuda:4",
12
+ "--output",
13
+ "articles/make_1/metadata/review_r1/controlled"
14
+ ],
15
+ "device": "cuda:4",
16
+ "gpu": "NVIDIA H200",
17
+ "python": "3.13.11 | packaged by Anaconda, Inc. | (main, Dec 10 2025, 21:28:48) [GCC 14.3.0]",
18
+ "pytorch": "2.6.0+cu124",
19
+ "cuda": "12.4",
20
+ "elapsed_analysis_seconds": 10.689740672009066,
21
+ "elapsed_since_standard_library_import_seconds": 37.757182730943896,
22
+ "first_cuda_noop_seconds": 1.7166021279990673e-05,
23
+ "peak_cuda_allocated_bytes": 168696832,
24
+ "opened_worlds": true,
25
+ "new_backbone_executions": 0,
26
+ "new_trainable_parameters": 0,
27
+ "refitted_neural_models": 0,
28
+ "fitted_statistical_tables": "Recomputed original estimators from archived counts, unchanged formulas",
29
+ "original_files_modified": false,
30
+ "all_selected_units_retained": true,
31
+ "all_statistics_descriptive": true,
32
+ "promotion_decisions_changed": false
33
+ }
revision_r1/controlled/family_sample_frontier.csv ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ family,records,worlds,identity,query_joint,posterior_clusters,posterior_bins,future_marginal,raw_information,raw_entropy,hash_code,action_values,true_oracle,joint_minus_action,paired_95_low,paired_95_high,action_encoding_loss,joint_encoding_loss,action_estimation_loss,joint_estimation_loss,encoding_contribution,estimation_contribution,action_remaining_headroom,fraction_action_headroom_closed,true_code_joint_oracle,true_code_action_oracle,true_code_joint_fitted,true_code_action_fitted,joint_root_choice_mismatch,action_root_choice_mismatch,strongest_uncompressed_mean
2
+ context,64,32,0.09339283465187889,0.0691595408487013,0.05884192092177272,0.06287654697819856,0.07187382385734353,0.07728978551777194,0.09455913414529662,0.10673495166633204,0.06409177954684019,0.7880888584118835,0.005067761301861108,0.0015311813058195526,0.008840865316316502,0.3813222379944849,0.3669534072130477,0.34267484087055833,0.3519759103501344,0.014368830781437176,-0.00930106947957607,0.7239970788650432,0.00699969854823926,0.7880888584118835,0.7880888584118835,0.12334932494026522,0.12491782030181064,0.6849965145249336,0.6875732689753824,0.2821009589784481
3
+ context,256,32,0.34655353787902665,0.3367188419663062,0.16416722984358104,0.1463272094176618,0.1871258673613156,0.26397328263687364,0.25243683714316034,0.15529673741167777,0.31618507415861496,0.7880888584118835,0.0205337678076913,0.01188513381910629,0.03021118308609348,0.11119897518661376,0.07603887002395061,0.3607048090666547,0.37533114642162657,0.035160105162663144,-0.014626337354971848,0.47190378425326845,0.043512615267927005,0.7880888584118835,0.7880888584118835,0.3856346205277096,0.38320113863439054,0.629779568015901,0.6496391850613675,0.5904656904258859
4
+ context,1024,32,0.6650919319515614,0.6814637624980395,0.23779038974868094,0.17124022355036714,0.36268052770898085,0.5145775504161055,0.5794729711288229,0.2170459709045936,0.667719627246173,0.7880888584118835,0.013744135251866517,0.009521684888968644,0.01821770096658289,0.01430655071271671,0.006186382463298001,0.10606268045299369,0.10043871345054588,0.008120168249418708,0.005623967002447808,0.12036923116571041,0.11418312735540517,0.7880888584118835,0.7880888584118835,0.6988418074434672,0.6964377598607214,0.3473903728720782,0.3575560574608555,0.7340548744395632
5
+ context,4096,32,0.766972716874273,0.7766349438098745,0.26807855851736373,0.1733365486789426,0.5632767718308346,0.5928278598851834,0.6885210614964501,0.24707180576088097,0.7768669245376707,0.7880888584118835,-0.00023198072779632983,-0.001376293766611737,0.0008712872445568171,0.0006891208483935686,0.00018064349661515988,0.010532813025819078,0.011273271105393817,0.0005084773517784087,-0.0007404580795747386,0.011221933874212647,-0.02067208115790168,0.7880888584118835,0.7880888584118835,0.7795383990383913,0.7790645144622885,0.1620783588267283,0.1554586985921155,0.7708361142534652
6
+ independent,64,32,0.909510548381619,0.9071402048294446,0.9072370207266524,0.9093082065269007,0.9074681824272653,0.8896605460004141,0.688211917829685,0.44151472051268514,0.9068080779301189,0.928933733504975,0.00033212689932572356,-0.0004962724151037125,0.0012075628763129809,0.009646569741400077,0.00982195667832023,0.012479085833456041,0.011971571997210165,-0.00017538693692015217,0.0005075138362458757,0.02212565557485612,0.015010940498556656,0.927818092070147,0.9281063650817161,0.9185297572467013,0.9192315821057243,0.23523002477266935,0.23312046998815447,0.9136191430312471
7
+ independent,256,32,0.9116024194164163,0.9098670177155274,0.9100779075741663,0.9126143882919777,0.905503332137763,0.896961421482459,0.6703717514265709,0.4619766730060043,0.9090004625042527,0.928933733504975,0.0008665552112746971,-0.0004301648384077622,0.0021276027149003647,0.006369755441728155,0.0065092495744983395,0.01356351555899413,0.01255746621494925,-0.00013949413277018416,0.0010060493440448813,0.019933271000722286,0.04347280540375422,0.927818092070147,0.9281063650817161,0.9170722759812633,0.9171807569702725,0.23936604222246965,0.23868371954799117,0.9251088162712999
8
+ independent,1024,32,0.9100079337524132,0.913013489675506,0.915318662687657,0.9168213117935321,0.8954121957914773,0.9031660787545921,0.6489306322062101,0.470594937948292,0.9133526580360904,0.928933733504975,-0.00033916836058426117,-0.0011080639841925125,0.00048298344501220073,0.006376698269395811,0.00708733789545923,0.009204377199488828,0.00883290593400967,-0.0007106396260634189,0.0003714712654791577,0.01558107546888464,-0.021767968537318196,0.927818092070147,0.9281063650817161,0.9209817567018171,0.9206480724749362,0.22556813929192912,0.21870599191031026,0.9280646584591856
9
+ independent,4096,32,0.9190137606668483,0.9212120370356884,0.9233281672038094,0.9234185604697762,0.9120224134643348,0.9084610246510227,0.6461753611274292,0.47480680272329256,0.9224258752130949,0.928933733504975,-0.0012138381774065114,-0.0017147112652704927,-0.000726757585522743,0.0038865080122648353,0.004790920676205797,0.002621350279615183,0.002930775793080733,-0.0009044126639409615,-0.0003094255134655499,0.0065078582918800185,-0.18651883968048918,0.927818092070147,0.9281063650817161,0.9258990135155555,0.9259546622363166,0.21008579345280287,0.19455413681065953,0.9287125775280298
10
+ orientation_alias,64,32,0.1635502787070287,0.12885644874690116,0.11989069408297903,0.12129218066799444,0.11565627135211173,0.1194667363872123,0.08451253981799893,0.062005774703194766,0.12837855808157933,0.828928467957698,0.00047789066532179854,-0.005027820336695715,0.005779686114251358,0.3951517175523762,0.4017012049487419,0.3053981923237425,0.29837081426205503,-0.0065494873963657056,0.007027378061687502,0.7005499098761186,0.0006821650514612244,0.828928467957698,0.828928467957698,0.23582020614293137,0.23582020614293134,0.6245539383534132,0.6283048497052103,0.3744516848229764
11
+ orientation_alias,256,32,0.48105690410695323,0.45413249913205544,0.3267375095162406,0.3044234970537981,0.17946386454389426,0.3385633917456905,0.26942870751962467,0.14858964279293768,0.4108957134778612,0.828928467957698,0.04323678565419422,0.03091808741028508,0.056253384606153346,0.22593360992407183,0.15025128394784587,0.19209914455576496,0.22454468487779672,0.07568232597622598,-0.03244554032203176,0.4180327544798368,0.1034291815434278,0.828928467957698,0.828928467957698,0.5438758040593306,0.5438758040593306,0.5760678352632569,0.6058617660589638,0.6522145924562783
12
+ orientation_alias,1024,32,0.7375432253689826,0.749695480009097,0.5557157182324963,0.5477364417189148,0.18531714088359752,0.6134052546318305,0.6043586411911206,0.2567112571360594,0.7156206511171621,0.828928467957698,0.03407482889193493,0.026490100551048997,0.0415008522799025,0.040254364563248254,0.01143065598560973,0.07305345227728763,0.06780233196299122,0.028823708577638524,0.00525112031429641,0.1133078168405359,0.30072796248373795,0.828928467957698,0.828928467957698,0.766961318774352,0.7669613187743521,0.3803245810966434,0.4051527363935008,0.767114986178982
13
+ orientation_alias,4096,32,0.804304431565352,0.8195916562097663,0.6530579291374521,0.648285517967734,0.18829678111200815,0.6964865610322947,0.706006590523646,0.3003141222527754,0.819480859898903,0.828928467957698,0.00011079631086339387,-2.228212253385705e-05,0.0002634376906898875,0.0002445290260795034,0.00019810226223412192,0.009203079032715649,0.009138709485697637,4.64267638453815e-05,6.436954701801237e-05,0.009447608058795152,0.011727445737998115,0.828928467957698,0.828928467957698,0.8198637628550727,0.8198637628550727,0.21189316146075493,0.21099546987915718,0.8096451880020169
14
+ parity,64,32,0.06273197484460194,0.02431126538233747,0.011591763903878482,0.011981140143721973,0.015523777582997956,0.011150427588187746,0.020184840684643574,0.024015757122275827,0.02252062020077706,0.8347957761429843,0.0017906451815604108,-0.0038315227161669556,0.006652954501555841,0.46294985993286397,0.4524966576959601,0.34932529600934337,0.35798785306468683,0.010453202236903802,-0.008662557055343396,0.8122751559422072,0.0022044810412591444,0.8347957761429844,0.8347957761429844,0.12286533950243893,0.12288347183177817,0.5209659718033554,0.5285378271397623,0.385086063569183
15
+ parity,256,32,0.4475888678222731,0.41757388132172324,0.2750026478906139,0.23783397003296353,0.1298958512570368,0.20886862239855064,0.18373006939535583,0.11691960704302928,0.38718031379931517,0.8347957761429843,0.030393567522408023,0.019497351412313457,0.041409246567013835,0.22704928735222724,0.15105877295815492,0.220566174991442,0.2661631218631063,0.07599051439407231,-0.04559694687166428,0.4476154623436692,0.06790106705266699,0.8347957761429844,0.8347957761429844,0.4773303993370155,0.4773303993370155,0.5556145869981621,0.5840915047008,0.6622060304284867
16
+ parity,1024,32,0.7505863638432653,0.7643539645704318,0.5785860682616655,0.5650067039347257,0.17619358394784873,0.491345522934861,0.5428645204900009,0.20833025114024675,0.749467269935558,0.8347957761429843,0.01488669463487384,0.010441901192385099,0.019552134986879535,0.021393038552985848,0.0075410150709803,0.06393546765444057,0.06290079650157226,0.013852023482005548,0.0010346711528682918,0.08532850620742641,0.17446332177297866,0.8347957761429844,0.8347957761429844,0.7761508394516206,0.7761508394516206,0.3591226824418755,0.35639638530427453,0.7761873106567354
17
+ parity,4096,32,0.811457808557315,0.824131626938427,0.694169383413467,0.6900830814062695,0.18472462130093503,0.6242484323345312,0.6210217172753493,0.2392919195668825,0.824102588730724,0.8347957761429843,2.9038207703040858e-05,-1.3477962034352808e-06,7.886248154405023e-05,9.112519989473711e-05,6.378990475829419e-05,0.010602062212365704,0.010600359299799106,2.7335295136442922e-05,1.7029125665979361e-06,0.010693187412260441,0.0027155801711421096,0.8347957761429844,0.8347957761429844,0.8242244796942859,0.8242244796942859,0.199476969506594,0.1989397435570667,0.8150734693187245
18
+ redundant,64,32,0.5363358739730864,0.5113855171174998,0.5036569357456289,0.5085041397628068,0.5000048354948977,0.5772794551090056,0.4124257595149454,-0.04050167310864608,0.5053617140765039,0.8364283751534423,0.006023803040995961,0.0009731619737643064,0.011502247775086136,0.14576415220704214,0.13364129983182016,0.1853025088698963,0.1914015582041223,0.012122852375221975,-0.006099049334226014,0.3310666610769384,0.018195136355321676,0.8364283751534423,0.8364283751534423,0.5890955182088521,0.5890955182088521,0.8539832767484425,0.8572854915108054,0.5662414160281566
19
+ redundant,256,32,0.6453415693659241,0.6198034615018286,0.5877161849929605,0.591621214981205,0.42260265019739973,0.6312017727896639,0.5514639211048289,0.019364542901451464,0.6050694499969314,0.8364283751534423,0.014734011504897247,0.009603019224582635,0.01995868486298478,0.0932629369589013,0.07785092527231657,0.13809598819760954,0.13877398837929703,0.015412011686584716,-0.0006780001816874687,0.23135892515651085,0.06368464711240299,0.8364283751534423,0.8364283751534423,0.653629840527671,0.653629840527671,0.763869962732947,0.7826760570524698,0.7105456685441063
20
+ redundant,1024,32,0.7653668596783447,0.7545480306631399,0.6989231534206255,0.6916977366851904,0.29795754465378854,0.7244842637951285,0.6854194139838374,0.04484946990041256,0.7412029739417946,0.8364283751534423,0.013345056721345335,0.010143917478369981,0.016472364541550773,0.039402157560219295,0.02814080307390973,0.05582324365142841,0.05373954141639264,0.01126135448630956,0.0020837022350357753,0.0952254012116477,0.14014177468976635,0.8364283751534423,0.8364283751534423,0.7721932591100672,0.7721932591100673,0.4023677500023982,0.43296693574774525,0.7863078402596235
21
+ redundant,4096,32,0.8155857149751549,0.8194726488422938,0.7193415226940727,0.6960627204154617,0.18000890013383308,0.7543788456935281,0.752804281570126,0.05292377176364181,0.8165418628657128,0.8364283751534423,0.002930785976580884,0.0019304463049228831,0.004100380433294095,0.00730164446795117,0.0038787169871150502,0.012584867819778103,0.013077009324033339,0.00342292748083612,-0.0004921415042552363,0.019886512287729273,0.14737556461266912,0.8364283751534423,0.8364283751534423,0.8226174136373886,0.8226174136373885,0.1746054019842278,0.18535970463320198,0.8207655371219653
22
+ uninformative,64,32,-0.04890966142219218,-0.04575341745644379,-0.047049768625332646,-0.04950046362247272,-0.026933077678025653,-0.041441113900889705,-0.010995531167365718,0.02058205608137745,-0.043723918882921195,0.0667348907999358,-0.0020294985735225856,-0.0039889178805482635,-0.000114955541232761,-2.0599841277224584e-18,-9.75781955236954e-19,0.11045880968285701,0.11248830825637958,-1.0842021724855044e-18,-0.002029498573522584,0.11045880968285701,-0.018373351834494372,0.0667348907999358,0.0667348907999358,-0.008709004374880524,-0.01331114610045102,0.5172644451120586,0.5137298310222932,-0.04890966142219218
23
+ uninformative,256,32,0.015731854753485836,0.03111698106830645,0.02924553921598229,0.03234940005689452,0.046069027949175986,0.03744741765448152,0.04500280677191939,0.04957765804565338,0.024803729988287868,0.0667348907999358,0.006313251080018578,0.0035633612177853723,0.009296090920171755,-1.1926223897340549e-18,5.421010862427522e-19,0.04193116081164793,0.03561790973162935,-1.734723475976807e-18,0.006313251080018579,0.04193116081164793,0.15056227773844122,0.0667348907999358,0.0667348907999358,0.046611496345385736,0.04479898126383822,0.318326143510767,0.3350521771578457,0.026880985378230898
24
+ uninformative,1024,32,0.02450926400985953,0.046334453063967655,0.041247390040046704,0.048595464204311184,0.055973307164560476,0.05435897532129334,0.05452489439127397,0.05498349310416952,0.038204503983425536,0.0667348907999358,0.008129949080542126,0.0048984844703365985,0.011408492793951204,4.553649124439119e-18,-3.3610267347050637e-18,0.028530386816510264,0.020400437735968148,7.914675859144182e-18,0.008129949080542116,0.028530386816510267,0.2849575483441185,0.0667348907999358,0.0667348907999358,0.05766653710546027,0.0569982079524269,0.27031338447698905,0.29409977769850704,0.05871121561510939
25
+ uninformative,4096,32,0.044186530387099784,0.05829518844278467,0.05390322882895525,0.05970059877167674,0.06250539980375937,0.06113657669308175,0.061513705438783525,0.0609891079973558,0.05473918863077931,0.0667348907999358,0.0035559998120053464,0.002404317887698644,0.004959589260649769,4.7704895589362195e-18,5.421010862427522e-19,0.011995702169156477,0.008439702357151135,4.228388472693467e-18,0.003555999812005342,0.011995702169156483,0.2964394882317588,0.0667348907999358,0.0667348907999358,0.06340955513774679,0.0624701246085254,0.21447414713172985,0.22634002044987156,0.06369397045613248
revision_r1/controlled/final_audit.json ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "created_utc": "2026-11-09T14:50:38.476227+00:00",
3
+ "status": "COMPLETED_EXPLORATORY_WITH_DISCLOSED_ARCHIVAL_TIE_SENSITIVITY",
4
+ "current_code_sha256": "4a300dfc769473f5afb08e15e877f6c53dfff5f6d3a0abfe17e0043798eecbab",
5
+ "specification_sha256": "d361b91b341dd596b246932f9113a0245adafcf5dcdce7b4e0f340d364b7b9da",
6
+ "primary_frontier_exact_replay_maximum_error": 3.3306690738754696e-16,
7
+ "original_reference_replay_changed_rows": 968,
8
+ "original_reference_replay_maximum_error": 0.1456998408107582,
9
+ "decomposition_identity_maximum_error": 1.6653345369377348e-16,
10
+ "all_192_worlds_6_families_3_streams_4_sizes_3_estimators_2_availability_9_methods_9_budget_cost_cells_retained": true,
11
+ "new_decomposition_grid_rows": 124416,
12
+ "new_decomposition_cell_rows": 1119744,
13
+ "no_source_results_protocols_models_or_manuscripts_modified": true,
14
+ "compute_accounting": "compute.json records only resumed aggregation/benchmark process. numeric_compute.json records the additional accumulation-order experiment. Initial full-grid collection completed but validation failed before its process ledger; its exact elapsed time and CUDA peak were not captured. Do not treat resumed process totals as total revision compute. No backbones or neural fits executed.",
15
+ "alias_caution": "Root-action mismatch can reflect equal-value zero-cost query ties; it is not itself decision regret. The utility loss decomposition and true-law encoders support the mechanism explanation.",
16
+ "paired_intervals": "All descriptive on opened worlds, uncorrected; world is paired unit, streams/costs/budgets averaged before primary resampling. No new confirmation or promotion.",
17
+ "statistical_fallacy_scan": {
18
+ "checked": 11,
19
+ "total": 11,
20
+ "items": {
21
+ "Simpson_paradox": "All six family contrasts and each cost/budget retained; heterogeneous signs disclosed.",
22
+ "Ecological_fallacy": "World-level intervals describe designed world population only; no sensor/person extrapolation.",
23
+ "Berkson_paradox": "No selection on results; every archived unit retained; constructed family selection limits generalization.",
24
+ "Collider_bias": "No covariate conditioning or exclusions based on outcomes.",
25
+ "Base_rate_neglect": "World-specific generating priors and true-law utility retain base rates; no sensitivity-only inference.",
26
+ "Regression_to_mean": "No extreme-world selection or pre/post effectiveness claim.",
27
+ "Survivorship_bias": "All 192 worlds retained; numerical discrepancies preserved.",
28
+ "Look_elsewhere": "Full condition grid and negative family results retained; intervals descriptive and not multiplicity adjusted.",
29
+ "Forking_paths": "Original specification and both post-failure amendments timestamped; no tuning or confirmatory relabeling.",
30
+ "Correlation_causation": "Fixed-code true-law replanning and true-code interventions isolate finite-grid losses; no unique general theoretical cause claimed.",
31
+ "Reverse_causality": "Encoders/policies constructed before separate true-law scoring; no target outcomes enter native fitted encoders."
32
+ }
33
+ }
34
+ }
revision_r1/controlled/frontier_all_conditions.csv ADDED
The diff for this file is too large to render. See raw diff
 
revision_r1/controlled/indices.json ADDED
@@ -0,0 +1,466 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "axes": [
3
+ "availability",
4
+ "labels",
5
+ "estimator",
6
+ "stream",
7
+ "method",
8
+ "world",
9
+ "budget",
10
+ "cost",
11
+ "metric"
12
+ ],
13
+ "availability": [
14
+ 6,
15
+ 3
16
+ ],
17
+ "labels": [
18
+ 64,
19
+ 256,
20
+ 1024,
21
+ 4096
22
+ ],
23
+ "estimator": [
24
+ "joint_jeffreys",
25
+ "naive_bayes",
26
+ "support_mixture"
27
+ ],
28
+ "stream": [
29
+ 0,
30
+ 1,
31
+ 2
32
+ ],
33
+ "method": [
34
+ "identity",
35
+ "query_joint",
36
+ "posterior_clusters",
37
+ "posterior_bins",
38
+ "future_marginal",
39
+ "raw_information",
40
+ "raw_entropy",
41
+ "hash_code",
42
+ "action_values"
43
+ ],
44
+ "world": [
45
+ "confirmation:independent:000",
46
+ "confirmation:independent:001",
47
+ "confirmation:independent:002",
48
+ "confirmation:independent:003",
49
+ "confirmation:independent:004",
50
+ "confirmation:independent:005",
51
+ "confirmation:independent:006",
52
+ "confirmation:independent:007",
53
+ "confirmation:independent:008",
54
+ "confirmation:independent:009",
55
+ "confirmation:independent:010",
56
+ "confirmation:independent:011",
57
+ "confirmation:independent:012",
58
+ "confirmation:independent:013",
59
+ "confirmation:independent:014",
60
+ "confirmation:independent:015",
61
+ "confirmation:independent:016",
62
+ "confirmation:independent:017",
63
+ "confirmation:independent:018",
64
+ "confirmation:independent:019",
65
+ "confirmation:independent:020",
66
+ "confirmation:independent:021",
67
+ "confirmation:independent:022",
68
+ "confirmation:independent:023",
69
+ "confirmation:independent:024",
70
+ "confirmation:independent:025",
71
+ "confirmation:independent:026",
72
+ "confirmation:independent:027",
73
+ "confirmation:independent:028",
74
+ "confirmation:independent:029",
75
+ "confirmation:independent:030",
76
+ "confirmation:independent:031",
77
+ "confirmation:context:000",
78
+ "confirmation:context:001",
79
+ "confirmation:context:002",
80
+ "confirmation:context:003",
81
+ "confirmation:context:004",
82
+ "confirmation:context:005",
83
+ "confirmation:context:006",
84
+ "confirmation:context:007",
85
+ "confirmation:context:008",
86
+ "confirmation:context:009",
87
+ "confirmation:context:010",
88
+ "confirmation:context:011",
89
+ "confirmation:context:012",
90
+ "confirmation:context:013",
91
+ "confirmation:context:014",
92
+ "confirmation:context:015",
93
+ "confirmation:context:016",
94
+ "confirmation:context:017",
95
+ "confirmation:context:018",
96
+ "confirmation:context:019",
97
+ "confirmation:context:020",
98
+ "confirmation:context:021",
99
+ "confirmation:context:022",
100
+ "confirmation:context:023",
101
+ "confirmation:context:024",
102
+ "confirmation:context:025",
103
+ "confirmation:context:026",
104
+ "confirmation:context:027",
105
+ "confirmation:context:028",
106
+ "confirmation:context:029",
107
+ "confirmation:context:030",
108
+ "confirmation:context:031",
109
+ "confirmation:parity:000",
110
+ "confirmation:parity:001",
111
+ "confirmation:parity:002",
112
+ "confirmation:parity:003",
113
+ "confirmation:parity:004",
114
+ "confirmation:parity:005",
115
+ "confirmation:parity:006",
116
+ "confirmation:parity:007",
117
+ "confirmation:parity:008",
118
+ "confirmation:parity:009",
119
+ "confirmation:parity:010",
120
+ "confirmation:parity:011",
121
+ "confirmation:parity:012",
122
+ "confirmation:parity:013",
123
+ "confirmation:parity:014",
124
+ "confirmation:parity:015",
125
+ "confirmation:parity:016",
126
+ "confirmation:parity:017",
127
+ "confirmation:parity:018",
128
+ "confirmation:parity:019",
129
+ "confirmation:parity:020",
130
+ "confirmation:parity:021",
131
+ "confirmation:parity:022",
132
+ "confirmation:parity:023",
133
+ "confirmation:parity:024",
134
+ "confirmation:parity:025",
135
+ "confirmation:parity:026",
136
+ "confirmation:parity:027",
137
+ "confirmation:parity:028",
138
+ "confirmation:parity:029",
139
+ "confirmation:parity:030",
140
+ "confirmation:parity:031",
141
+ "confirmation:uninformative:000",
142
+ "confirmation:uninformative:001",
143
+ "confirmation:uninformative:002",
144
+ "confirmation:uninformative:003",
145
+ "confirmation:uninformative:004",
146
+ "confirmation:uninformative:005",
147
+ "confirmation:uninformative:006",
148
+ "confirmation:uninformative:007",
149
+ "confirmation:uninformative:008",
150
+ "confirmation:uninformative:009",
151
+ "confirmation:uninformative:010",
152
+ "confirmation:uninformative:011",
153
+ "confirmation:uninformative:012",
154
+ "confirmation:uninformative:013",
155
+ "confirmation:uninformative:014",
156
+ "confirmation:uninformative:015",
157
+ "confirmation:uninformative:016",
158
+ "confirmation:uninformative:017",
159
+ "confirmation:uninformative:018",
160
+ "confirmation:uninformative:019",
161
+ "confirmation:uninformative:020",
162
+ "confirmation:uninformative:021",
163
+ "confirmation:uninformative:022",
164
+ "confirmation:uninformative:023",
165
+ "confirmation:uninformative:024",
166
+ "confirmation:uninformative:025",
167
+ "confirmation:uninformative:026",
168
+ "confirmation:uninformative:027",
169
+ "confirmation:uninformative:028",
170
+ "confirmation:uninformative:029",
171
+ "confirmation:uninformative:030",
172
+ "confirmation:uninformative:031",
173
+ "confirmation:redundant:000",
174
+ "confirmation:redundant:001",
175
+ "confirmation:redundant:002",
176
+ "confirmation:redundant:003",
177
+ "confirmation:redundant:004",
178
+ "confirmation:redundant:005",
179
+ "confirmation:redundant:006",
180
+ "confirmation:redundant:007",
181
+ "confirmation:redundant:008",
182
+ "confirmation:redundant:009",
183
+ "confirmation:redundant:010",
184
+ "confirmation:redundant:011",
185
+ "confirmation:redundant:012",
186
+ "confirmation:redundant:013",
187
+ "confirmation:redundant:014",
188
+ "confirmation:redundant:015",
189
+ "confirmation:redundant:016",
190
+ "confirmation:redundant:017",
191
+ "confirmation:redundant:018",
192
+ "confirmation:redundant:019",
193
+ "confirmation:redundant:020",
194
+ "confirmation:redundant:021",
195
+ "confirmation:redundant:022",
196
+ "confirmation:redundant:023",
197
+ "confirmation:redundant:024",
198
+ "confirmation:redundant:025",
199
+ "confirmation:redundant:026",
200
+ "confirmation:redundant:027",
201
+ "confirmation:redundant:028",
202
+ "confirmation:redundant:029",
203
+ "confirmation:redundant:030",
204
+ "confirmation:redundant:031",
205
+ "confirmation:orientation_alias:000",
206
+ "confirmation:orientation_alias:001",
207
+ "confirmation:orientation_alias:002",
208
+ "confirmation:orientation_alias:003",
209
+ "confirmation:orientation_alias:004",
210
+ "confirmation:orientation_alias:005",
211
+ "confirmation:orientation_alias:006",
212
+ "confirmation:orientation_alias:007",
213
+ "confirmation:orientation_alias:008",
214
+ "confirmation:orientation_alias:009",
215
+ "confirmation:orientation_alias:010",
216
+ "confirmation:orientation_alias:011",
217
+ "confirmation:orientation_alias:012",
218
+ "confirmation:orientation_alias:013",
219
+ "confirmation:orientation_alias:014",
220
+ "confirmation:orientation_alias:015",
221
+ "confirmation:orientation_alias:016",
222
+ "confirmation:orientation_alias:017",
223
+ "confirmation:orientation_alias:018",
224
+ "confirmation:orientation_alias:019",
225
+ "confirmation:orientation_alias:020",
226
+ "confirmation:orientation_alias:021",
227
+ "confirmation:orientation_alias:022",
228
+ "confirmation:orientation_alias:023",
229
+ "confirmation:orientation_alias:024",
230
+ "confirmation:orientation_alias:025",
231
+ "confirmation:orientation_alias:026",
232
+ "confirmation:orientation_alias:027",
233
+ "confirmation:orientation_alias:028",
234
+ "confirmation:orientation_alias:029",
235
+ "confirmation:orientation_alias:030",
236
+ "confirmation:orientation_alias:031"
237
+ ],
238
+ "family": [
239
+ "independent",
240
+ "independent",
241
+ "independent",
242
+ "independent",
243
+ "independent",
244
+ "independent",
245
+ "independent",
246
+ "independent",
247
+ "independent",
248
+ "independent",
249
+ "independent",
250
+ "independent",
251
+ "independent",
252
+ "independent",
253
+ "independent",
254
+ "independent",
255
+ "independent",
256
+ "independent",
257
+ "independent",
258
+ "independent",
259
+ "independent",
260
+ "independent",
261
+ "independent",
262
+ "independent",
263
+ "independent",
264
+ "independent",
265
+ "independent",
266
+ "independent",
267
+ "independent",
268
+ "independent",
269
+ "independent",
270
+ "independent",
271
+ "context",
272
+ "context",
273
+ "context",
274
+ "context",
275
+ "context",
276
+ "context",
277
+ "context",
278
+ "context",
279
+ "context",
280
+ "context",
281
+ "context",
282
+ "context",
283
+ "context",
284
+ "context",
285
+ "context",
286
+ "context",
287
+ "context",
288
+ "context",
289
+ "context",
290
+ "context",
291
+ "context",
292
+ "context",
293
+ "context",
294
+ "context",
295
+ "context",
296
+ "context",
297
+ "context",
298
+ "context",
299
+ "context",
300
+ "context",
301
+ "context",
302
+ "context",
303
+ "parity",
304
+ "parity",
305
+ "parity",
306
+ "parity",
307
+ "parity",
308
+ "parity",
309
+ "parity",
310
+ "parity",
311
+ "parity",
312
+ "parity",
313
+ "parity",
314
+ "parity",
315
+ "parity",
316
+ "parity",
317
+ "parity",
318
+ "parity",
319
+ "parity",
320
+ "parity",
321
+ "parity",
322
+ "parity",
323
+ "parity",
324
+ "parity",
325
+ "parity",
326
+ "parity",
327
+ "parity",
328
+ "parity",
329
+ "parity",
330
+ "parity",
331
+ "parity",
332
+ "parity",
333
+ "parity",
334
+ "parity",
335
+ "uninformative",
336
+ "uninformative",
337
+ "uninformative",
338
+ "uninformative",
339
+ "uninformative",
340
+ "uninformative",
341
+ "uninformative",
342
+ "uninformative",
343
+ "uninformative",
344
+ "uninformative",
345
+ "uninformative",
346
+ "uninformative",
347
+ "uninformative",
348
+ "uninformative",
349
+ "uninformative",
350
+ "uninformative",
351
+ "uninformative",
352
+ "uninformative",
353
+ "uninformative",
354
+ "uninformative",
355
+ "uninformative",
356
+ "uninformative",
357
+ "uninformative",
358
+ "uninformative",
359
+ "uninformative",
360
+ "uninformative",
361
+ "uninformative",
362
+ "uninformative",
363
+ "uninformative",
364
+ "uninformative",
365
+ "uninformative",
366
+ "uninformative",
367
+ "redundant",
368
+ "redundant",
369
+ "redundant",
370
+ "redundant",
371
+ "redundant",
372
+ "redundant",
373
+ "redundant",
374
+ "redundant",
375
+ "redundant",
376
+ "redundant",
377
+ "redundant",
378
+ "redundant",
379
+ "redundant",
380
+ "redundant",
381
+ "redundant",
382
+ "redundant",
383
+ "redundant",
384
+ "redundant",
385
+ "redundant",
386
+ "redundant",
387
+ "redundant",
388
+ "redundant",
389
+ "redundant",
390
+ "redundant",
391
+ "redundant",
392
+ "redundant",
393
+ "redundant",
394
+ "redundant",
395
+ "redundant",
396
+ "redundant",
397
+ "redundant",
398
+ "redundant",
399
+ "orientation_alias",
400
+ "orientation_alias",
401
+ "orientation_alias",
402
+ "orientation_alias",
403
+ "orientation_alias",
404
+ "orientation_alias",
405
+ "orientation_alias",
406
+ "orientation_alias",
407
+ "orientation_alias",
408
+ "orientation_alias",
409
+ "orientation_alias",
410
+ "orientation_alias",
411
+ "orientation_alias",
412
+ "orientation_alias",
413
+ "orientation_alias",
414
+ "orientation_alias",
415
+ "orientation_alias",
416
+ "orientation_alias",
417
+ "orientation_alias",
418
+ "orientation_alias",
419
+ "orientation_alias",
420
+ "orientation_alias",
421
+ "orientation_alias",
422
+ "orientation_alias",
423
+ "orientation_alias",
424
+ "orientation_alias",
425
+ "orientation_alias",
426
+ "orientation_alias",
427
+ "orientation_alias",
428
+ "orientation_alias",
429
+ "orientation_alias",
430
+ "orientation_alias"
431
+ ],
432
+ "budget": [
433
+ 0,
434
+ 1,
435
+ 2
436
+ ],
437
+ "cost": [
438
+ 0.0,
439
+ 0.01,
440
+ 0.05
441
+ ],
442
+ "metric": [
443
+ "utility",
444
+ "predicted_utility",
445
+ "queries",
446
+ "abstention",
447
+ "available_oracle",
448
+ "encoded_oracle",
449
+ "true_encoder_oracle",
450
+ "true_encoder_fitted_utility",
451
+ "true_conditional_dispersion",
452
+ "root_choice_mismatch"
453
+ ],
454
+ "score_shape": [
455
+ 2,
456
+ 4,
457
+ 3,
458
+ 3,
459
+ 9,
460
+ 192,
461
+ 3,
462
+ 3,
463
+ 10
464
+ ],
465
+ "all_units_retained": true
466
+ }
revision_r1/controlled/native_worlds.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1390af820583c90646ceb5edd7da2647ea0b9d8aeedfd516c921cdeea1b58581
3
+ size 3779770
revision_r1/controlled/numeric_compute.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "code_sha256": "4a300dfc769473f5afb08e15e877f6c53dfff5f6d3a0abfe17e0043798eecbab",
3
+ "elapsed_analysis_seconds": 19.021096944925375,
4
+ "elapsed_process_seconds": 37.11383769393433,
5
+ "peak_cuda_allocated_bytes": 167947264,
6
+ "device": "cuda:4",
7
+ "original_files_modified": false,
8
+ "new_trainable_parameters": 0
9
+ }
revision_r1/controlled/numeric_diagnostic.json ADDED
@@ -0,0 +1,154 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "POST_REVIEW_NUMERICAL_DIAGNOSTIC_NOT_A_COMPETITOR",
3
+ "code_sha256": "4a300dfc769473f5afb08e15e877f6c53dfff5f6d3a0abfe17e0043798eecbab",
4
+ "operation": "Hold each encoder fixed. Compare ordinary pushforward to reversal of raw-state accumulation order, which is the same probability law in exact arithmetic.",
5
+ "metric_axis": [
6
+ "ordinary_realized_utility",
7
+ "reversed_sum_realized_utility",
8
+ "absolute_fitted_optimal_value_difference",
9
+ "maximum_law_cell_difference"
10
+ ],
11
+ "all_world_stream_method_rows": 124416,
12
+ "changed_world_stream_method_rows": 2637,
13
+ "maximum_law_cell_difference": 2.220446049250313e-16,
14
+ "maximum_fitted_optimal_value_difference": 6.106226635438361e-16,
15
+ "maximum_realized_utility_difference": 0.27986264044260756,
16
+ "changed_by_estimator": {
17
+ "joint_jeffreys": 2609,
18
+ "naive_bayes": 26,
19
+ "support_mixture": 2
20
+ },
21
+ "interpretation": "Tiny law perturbations can select different fitted-optimal tied or nearly tied actions whose true-law utilities differ. This demonstrates numerical tie sensitivity; it does not uniquely reconstruct every archived discrepancy or justify replacing frozen scores.",
22
+ "strongest_witnesses": [
23
+ {
24
+ "available": 6,
25
+ "labels": 64,
26
+ "estimator": "joint_jeffreys",
27
+ "stream": 1,
28
+ "method": "raw_entropy",
29
+ "world": "confirmation:redundant:009",
30
+ "family": "redundant",
31
+ "maximum_law_cell_difference": 2.7755575615628914e-17,
32
+ "maximum_fitted_optimal_value_difference": 2.7755575615628914e-17,
33
+ "maximum_realized_utility_difference": 0.27986264044260756,
34
+ "positive_budget_average_utility_difference": 0.0
35
+ },
36
+ {
37
+ "available": 6,
38
+ "labels": 64,
39
+ "estimator": "joint_jeffreys",
40
+ "stream": 0,
41
+ "method": "raw_information",
42
+ "world": "confirmation:redundant:030",
43
+ "family": "redundant",
44
+ "maximum_law_cell_difference": 2.7755575615628914e-17,
45
+ "maximum_fitted_optimal_value_difference": 4.163336342344337e-17,
46
+ "maximum_realized_utility_difference": 0.2505645108394803,
47
+ "positive_budget_average_utility_difference": 0.0
48
+ },
49
+ {
50
+ "available": 3,
51
+ "labels": 64,
52
+ "estimator": "joint_jeffreys",
53
+ "stream": 2,
54
+ "method": "posterior_bins",
55
+ "world": "confirmation:orientation_alias:007",
56
+ "family": "orientation_alias",
57
+ "maximum_law_cell_difference": 1.3877787807814457e-17,
58
+ "maximum_fitted_optimal_value_difference": 1.3877787807814457e-17,
59
+ "maximum_realized_utility_difference": 0.24603762586320843,
60
+ "positive_budget_average_utility_difference": 0.0410062709772014
61
+ },
62
+ {
63
+ "available": 6,
64
+ "labels": 64,
65
+ "estimator": "joint_jeffreys",
66
+ "stream": 0,
67
+ "method": "raw_information",
68
+ "world": "confirmation:redundant:027",
69
+ "family": "redundant",
70
+ "maximum_law_cell_difference": 1.3877787807814457e-17,
71
+ "maximum_fitted_optimal_value_difference": 2.7755575615628914e-17,
72
+ "maximum_realized_utility_difference": 0.22993269695902768,
73
+ "positive_budget_average_utility_difference": 0.0
74
+ },
75
+ {
76
+ "available": 3,
77
+ "labels": 64,
78
+ "estimator": "joint_jeffreys",
79
+ "stream": 0,
80
+ "method": "posterior_bins",
81
+ "world": "confirmation:orientation_alias:005",
82
+ "family": "orientation_alias",
83
+ "maximum_law_cell_difference": 2.7755575615628914e-17,
84
+ "maximum_fitted_optimal_value_difference": 5.551115123125783e-17,
85
+ "maximum_realized_utility_difference": 0.21863751129021827,
86
+ "positive_budget_average_utility_difference": 0.10931875564510912
87
+ },
88
+ {
89
+ "available": 6,
90
+ "labels": 64,
91
+ "estimator": "naive_bayes",
92
+ "stream": 1,
93
+ "method": "raw_entropy",
94
+ "world": "confirmation:parity:009",
95
+ "family": "parity",
96
+ "maximum_law_cell_difference": 6.938893903907228e-18,
97
+ "maximum_fitted_optimal_value_difference": 2.7755575615628914e-17,
98
+ "maximum_realized_utility_difference": 0.20848183477792573,
99
+ "positive_budget_average_utility_difference": 0.1036287884194994
100
+ },
101
+ {
102
+ "available": 6,
103
+ "labels": 64,
104
+ "estimator": "joint_jeffreys",
105
+ "stream": 1,
106
+ "method": "raw_information",
107
+ "world": "confirmation:orientation_alias:012",
108
+ "family": "orientation_alias",
109
+ "maximum_law_cell_difference": 1.3877787807814457e-17,
110
+ "maximum_fitted_optimal_value_difference": 1.3877787807814457e-17,
111
+ "maximum_realized_utility_difference": 0.19696465415136627,
112
+ "positive_budget_average_utility_difference": -0.03282744235856104
113
+ },
114
+ {
115
+ "available": 6,
116
+ "labels": 64,
117
+ "estimator": "joint_jeffreys",
118
+ "stream": 0,
119
+ "method": "raw_information",
120
+ "world": "confirmation:orientation_alias:003",
121
+ "family": "orientation_alias",
122
+ "maximum_law_cell_difference": 1.3877787807814457e-17,
123
+ "maximum_fitted_optimal_value_difference": 1.3877787807814457e-17,
124
+ "maximum_realized_utility_difference": 0.15956330285126807,
125
+ "positive_budget_average_utility_difference": 0.026593883808544678
126
+ },
127
+ {
128
+ "available": 6,
129
+ "labels": 64,
130
+ "estimator": "joint_jeffreys",
131
+ "stream": 1,
132
+ "method": "posterior_clusters",
133
+ "world": "confirmation:redundant:014",
134
+ "family": "redundant",
135
+ "maximum_law_cell_difference": 3.469446951953614e-18,
136
+ "maximum_fitted_optimal_value_difference": 1.3877787807814457e-17,
137
+ "maximum_realized_utility_difference": 0.15947780468816652,
138
+ "positive_budget_average_utility_difference": -0.02657963411469442
139
+ },
140
+ {
141
+ "available": 6,
142
+ "labels": 64,
143
+ "estimator": "joint_jeffreys",
144
+ "stream": 1,
145
+ "method": "posterior_bins",
146
+ "world": "confirmation:redundant:014",
147
+ "family": "redundant",
148
+ "maximum_law_cell_difference": 1.3877787807814457e-17,
149
+ "maximum_fitted_optimal_value_difference": 1.3877787807814457e-17,
150
+ "maximum_realized_utility_difference": 0.15947780468816652,
151
+ "positive_budget_average_utility_difference": -0.02657963411469442
152
+ }
153
+ ]
154
+ }
revision_r1/controlled/numeric_push_sensitivity.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:655e73bc4111c4c5dab7507e4fd029cd5fba7efdde6375757a6b281c94027358
3
+ size 597542
revision_r1/controlled/paired_by_condition.csv ADDED
The diff for this file is too large to render. See raw diff
 
revision_r1/controlled/privileged_geometry.csv ADDED
The diff for this file is too large to render. See raw diff
 
revision_r1/controlled/reference_replay_discrepancies.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:59fad6d56023fedd16c50f0faab3b4f85a8cb322c3ecbcb4f7d5f9c7e5076051
3
+ size 34992
revision_r1/controlled/replay_discrepancies.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6ef693532219b9ba0abab9aaafcc07558a72a20f0ffe961b4bc42a138e873d25
3
+ size 34118
revision_r1/controlled/run.py ADDED
@@ -0,0 +1,572 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Post-review opened-world diagnostics; frozen confirmation results are immutable.
2
+
3
+ First run --prepare, then run without --prepare using the same output directory.
4
+ All statistical tensors and decision calculations use CUDA float64. CPU work is
5
+ limited to provenance, output, and descriptive paired-bootstrap aggregation.
6
+ """
7
+ from __future__ import annotations
8
+
9
+ import argparse
10
+ import csv
11
+ import datetime as dt
12
+ import gzip
13
+ import hashlib
14
+ import json
15
+ import os
16
+ from pathlib import Path
17
+ import sys
18
+ import time
19
+
20
+ PROCESS_START = time.perf_counter()
21
+ HERE = Path(__file__).resolve()
22
+ ARTICLE = HERE.parents[1]
23
+ DEFAULT_RELEASE = next((p for p in [HERE.parent, HERE.parent.parent, ARTICLE / "release/huggingface-revised"]
24
+ if (p / "data/controlled/manifest.json").is_file()), ARTICLE / "release/huggingface-revised")
25
+ METHODS = ["identity", "query_joint", "posterior_clusters", "posterior_bins",
26
+ "future_marginal", "raw_information", "raw_entropy", "hash_code", "action_values"]
27
+ AVAILABILITY = [6, 3]
28
+ SIZES = [64, 256, 1024, 4096]
29
+ ESTIMATORS = ["joint_jeffreys", "naive_bayes", "support_mixture"]
30
+ COSTS = [0.0, 0.01, 0.05]
31
+ METRICS = ["utility", "predicted_utility", "queries", "abstention", "available_oracle",
32
+ "encoded_oracle", "true_encoder_oracle", "true_encoder_fitted_utility",
33
+ "true_conditional_dispersion", "root_choice_mismatch"]
34
+
35
+
36
+ def sha(path):
37
+ return hashlib.sha256(Path(path).read_bytes()).hexdigest()
38
+
39
+
40
+ def dump(path, value):
41
+ Path(path).write_text(json.dumps(value, indent=2, allow_nan=False) + "\n")
42
+
43
+
44
+ def csv_write(path, rows):
45
+ rows = list(rows)
46
+ opener = gzip.open if str(path).endswith(".gz") else open
47
+ with opener(path, "wt", newline="") as f:
48
+ writer = csv.DictWriter(f, fieldnames=list(rows[0]))
49
+ writer.writeheader()
50
+ writer.writerows(rows)
51
+
52
+
53
+ def prepare(out, release, device):
54
+ assert not out.exists(), "The specification must use a fresh output directory."
55
+ out.mkdir(parents=True)
56
+ inputs = [release / "data/controlled/worlds.safetensors",
57
+ release / "data/controlled/worlds.json",
58
+ release / "data/controlled/manifest.json",
59
+ release / "postconfirmation.py",
60
+ release / "unionmind/decision_sufficiency/core.py",
61
+ release / "results/postconfirmation/controlled_worlds.csv.gz",
62
+ release / "results/postconfirmation/controlled_frontiers.csv.gz",
63
+ release / "results/postconfirmation/controlled_timing.json",
64
+ release / "provenance/controlled/confirmation_benchmark.json"]
65
+ spec = dict(created_utc=dt.datetime.now(dt.timezone.utc).isoformat(),
66
+ status="Post-review exploratory analysis of already opened worlds; not confirmation",
67
+ scientific_question="Why does joint minus action-value utility shrink with sample size?",
68
+ authorization="User authorized additional reviewer-response experiments; no new approval requested.",
69
+ source_release=str(release), code_sha256=sha(HERE), device=device,
70
+ input_sha256={str(p.relative_to(release)): sha(p) for p in inputs},
71
+ units="All 192 archived worlds, all six families, all three original nested sample streams",
72
+ grid=dict(availability=AVAILABILITY, records=SIZES, estimators=ESTIMATORS,
73
+ methods=METHODS, budgets=[0, 1, 2], costs=COSTS),
74
+ primary_aggregation="Average positive budgets, all costs and three streams within each world; equal worlds and six families",
75
+ interventions=[
76
+ "Native: estimated-law encoder, estimated encoded-law exact planner, true-law evaluation",
77
+ "Replan the identical native code under the pushed true law to separate encoding and model-estimation loss",
78
+ "Privileged diagnostic: construct each unchanged encoder under the true law, then plan under either true or fitted pushed law; never a deployable competitor"],
79
+ exact_decomposition="J-A = (encoding_loss_A-encoding_loss_J) + (estimation_loss_A-estimation_loss_J)",
80
+ encoding_loss="V(available true raw law)-V(true law pushed through native code)",
81
+ estimation_loss="V(true law pushed through native code)-realized utility of fitted policy",
82
+ alias_diagnostic="Evaluate all unordered raw-state pairs with true current-posterior difference <=1e-12 (exact) and <=0.01 (fixed near-alias threshold), weighted by 2 P(e)P(e'); compare true optimal root choices at each cost and budget and whether each native code merges disagreeing aliases",
83
+ intervals="Descriptive paired 95% percentile bootstrap, 10,000 draws, seed 18701, worlds resampled within fixed families after averaging streams; intervals are not corrected and are not new confirmatory tests",
84
+ practical_scale="Joint-action gain divided by remaining true available-oracle headroom of action-values; report signed gain, not just a percentage",
85
+ timing="Retain published complete-cost records. Fresh warm alternating-order 11-repeat 192-world primary-batch timings include count/estimate, profile, clustering/other encoding, push, exact planning, materialization and 1/32/1024/65536 recorded-event replays. One warmup charged and separately reported. Timing is descriptive on shared hardware.",
86
+ prohibition="No tuning, no model refits, no changed codes/iterations/costs, no dropped units, no original results or manuscript edits, no promotion",
87
+ planned_outputs=["scores.npz", "indices.json", "frontier_all_conditions.csv", "paired_by_condition.csv",
88
+ "sample_frontier.csv", "family_sample_frontier.csv", "native_worlds.csv.gz",
89
+ "alias_pairs.csv.gz", "alias_summary.csv", "privileged_geometry.csv",
90
+ "timing.json", "summary.json", "compute.json"])
91
+ dump(out / "specification.json", spec)
92
+ print(json.dumps(dict(status="SPECIFICATION_RECORDED_BEFORE_NEW_COMPUTATIONS", output=str(out))))
93
+
94
+
95
+ def sync():
96
+ torch.cuda.synchronize()
97
+
98
+
99
+ def timed(fn):
100
+ sync()
101
+ start = time.perf_counter()
102
+ result = fn()
103
+ sync()
104
+ return result, time.perf_counter() - start
105
+
106
+
107
+ def mismatch(policy, code, raw_policy, mass):
108
+ w, r = code.shape
109
+ table = policy["choices"].reshape(3, w, policy["shape"][1], 3, 9)[..., 0]
110
+ selected = table.gather(2, code[None, :, :, None].expand(3, w, r, 3))
111
+ reference = raw_policy["choices"].reshape(3, w, r, 3, 9)[..., 0]
112
+ return ((selected != reference).to(mass.dtype) * mass[None, :, :, None]).sum(2).permute(1, 0, 2)
113
+
114
+
115
+ def collect(out, release, device):
116
+ archive = load("data/controlled/worlds", device)
117
+ manifest = json.loads((release / "data/controlled/manifest.json").read_text())["worlds"]
118
+ world_ids = [r["id"] for r in manifest]
119
+ families = np.array([r["family"] for r in manifest])
120
+ assert len(world_ids) == 192 and len(set(families)) == 6
121
+ scores = np.empty((2, 4, 3, 3, 9, 192, 3, 3, len(METRICS)), dtype=np.float64)
122
+ alias_rows = []
123
+ alias_pair_rows = []
124
+ geometry_rows = []
125
+ progress = []
126
+ for ai, available in enumerate(AVAILABILITY):
127
+ truth = archive["joint"] if available == 6 else core.project(archive["joint"], 3)
128
+ oracle_policy = core.plan(truth)
129
+ oracle = oracle_policy["expected"]
130
+ true_mass = truth.sum((2, 3))
131
+ posterior = truth.sum(-1)[..., 1] / true_mass.clamp_min(1e-300)
132
+ r = truth.shape[1]
133
+ pi, pj = torch.triu_indices(r, r, 1, device=device)
134
+ pair_weight = 2 * true_mass[:, pi] * true_mass[:, pj]
135
+ gap = (posterior[:, pi] - posterior[:, pj]).abs()
136
+ raw_choices = oracle_policy["choices"].reshape(3, 192, r, 3, 9)[..., 0].permute(1, 2, 0, 3)
137
+ disagree = raw_choices[:, pi] != raw_choices[:, pj]
138
+ true_codes = {}
139
+ for method in METHODS:
140
+ code, k = encode(truth, method)
141
+ pushed = core.push(truth, code, k)
142
+ privileged = core.plan(pushed)["expected"]
143
+ true_codes[method] = (code, k, privileged)
144
+ for wi, world in enumerate(world_ids):
145
+ geometry_rows.append(dict(available=available, world=world, family=str(families[wi]), method=method,
146
+ utility=float(privileged[wi, 1:].mean()),
147
+ encoding_loss=float((oracle[wi, 1:] - privileged[wi, 1:]).mean())))
148
+ # Store every eligible pair and its exact law-derived statistics; no sampled pair subset.
149
+ for label, threshold in [("exact", 1e-12), ("near", 0.01)]:
150
+ mask = gap <= threshold
151
+ for wi in range(192):
152
+ inds = mask[wi].nonzero().flatten()
153
+ for ix in inds.tolist():
154
+ alias_pair_rows.append(dict(available=available, world=world_ids[wi], family=str(families[wi]),
155
+ threshold=label, raw_i=int(pi[ix]), raw_j=int(pj[ix]),
156
+ posterior_difference=float(gap[wi, ix]), pair_weight=float(pair_weight[wi, ix]),
157
+ disagreeing_positive_budget_cost_cells=int(disagree[wi, ix, 1:].sum())))
158
+ for ni, n in enumerate(SIZES):
159
+ for ei, estimator in enumerate(ESTIMATORS):
160
+ for stream in range(3):
161
+ fitted = core.estimate(core.count_samples(archive["samples"][:, stream, :n], available), estimator)
162
+ for mi, method in enumerate(METHODS):
163
+ code, k = encode(fitted, method)
164
+ compressed = core.push(fitted, code, k)
165
+ policy = core.plan(compressed)
166
+ pushed_truth = core.push(truth, code, k)
167
+ realized = core.realized(pushed_truth, policy)
168
+ encoded_oracle = core.plan(pushed_truth)["expected"]
169
+ true_code, true_k, privileged_oracle = true_codes[method]
170
+ privileged_fitted = core.plan(core.push(fitted, true_code, true_k))
171
+ privileged_actual = core.realized(core.push(truth, true_code, true_k), privileged_fitted)["utility"]
172
+ dispersion = core.dispersion(truth, code, pushed_truth)[:, None, None].expand_as(oracle)
173
+ root_mismatch = mismatch(policy, code, oracle_policy, true_mass)
174
+ metrics = [realized["utility"], policy["expected"], realized["queries"], realized["abstention"],
175
+ oracle, encoded_oracle, privileged_oracle, privileged_actual, dispersion, root_mismatch]
176
+ assert min(float((oracle - encoded_oracle).min()),
177
+ float((encoded_oracle - realized["utility"]).min())) > -1e-10
178
+ scores[ai, ni, ei, stream, mi] = torch.stack(metrics, -1).cpu().numpy()
179
+ # Exact expected-mass summaries across every eligible pair, all worlds and all cells.
180
+ same_code = code[:, pi] == code[:, pj]
181
+ for label, threshold in [("exact", 1e-12), ("near", 0.01)]:
182
+ weighted = pair_weight * (gap <= threshold)
183
+ denominator = weighted.sum(1)
184
+ den_disagree = (weighted[..., None, None] * disagree).sum(1)
185
+ merged_disagree = (weighted[..., None, None] * disagree * same_code[..., None, None]).sum(1)
186
+ for b in range(3):
187
+ for ci, cost in enumerate(COSTS):
188
+ alias_rows.append(dict(available=available, labels=n, estimator=estimator,
189
+ stream=stream, method=method, threshold=label, budget=b, cost=cost,
190
+ eligible_pair_mass=float(denominator.mean()),
191
+ disagreeing_pair_mass=float(den_disagree[:, b, ci].mean()),
192
+ merged_disagreeing_pair_mass=float(merged_disagree[:, b, ci].mean())))
193
+ progress.append(dict(available=available, labels=n, estimator=estimator))
194
+ print(json.dumps(dict(completed=progress[-1])), flush=True)
195
+ indices = dict(axes=["availability", "labels", "estimator", "stream", "method", "world", "budget", "cost", "metric"],
196
+ availability=AVAILABILITY, labels=SIZES, estimator=ESTIMATORS, stream=[0, 1, 2],
197
+ method=METHODS, world=world_ids, family=families.tolist(), budget=[0, 1, 2], cost=COSTS, metric=METRICS,
198
+ score_shape=list(scores.shape), all_units_retained=True)
199
+ np.savez_compressed(out / "scores.npz", scores=scores)
200
+ dump(out / "indices.json", indices)
201
+ csv_write(out / "alias_pairs.csv.gz", alias_pair_rows)
202
+ csv_write(out / "alias_summary.csv", alias_rows)
203
+ csv_write(out / "privileged_geometry.csv", geometry_rows)
204
+ return scores, indices, archive
205
+
206
+
207
+ def bootstrap_weights(families):
208
+ rng = np.random.default_rng(18701)
209
+ by_family = {}
210
+ combined = np.zeros((10000, len(families)))
211
+ groups = sorted(set(families))
212
+ for family in groups:
213
+ selected = np.flatnonzero(families == family)
214
+ draw = rng.integers(len(selected), size=(10000, len(selected)))
215
+ weights = np.stack([np.bincount(d, minlength=len(selected)) for d in draw]).astype(float) / len(selected)
216
+ combined[:, selected] = weights / len(groups)
217
+ by_family[family] = (selected, weights)
218
+ return combined, by_family
219
+
220
+
221
+ def summarize(out, release, scores, indices):
222
+ families = np.asarray(indices["family"])
223
+ # Average streams before any resampling. Cells and streams are not independent worlds.
224
+ worlds = scores.mean(3)
225
+ positive = worlds[..., 1:, :, :].mean((-3, -2))
226
+ metric = {name: i for i, name in enumerate(METRICS)}
227
+ ji, av = METHODS.index("query_joint"), METHODS.index("action_values")
228
+ utility = worlds[..., metric["utility"]]
229
+ encoding = worlds[..., metric["available_oracle"]] - worlds[..., metric["encoded_oracle"]]
230
+ estimation = worlds[..., metric["encoded_oracle"]] - utility
231
+ delta = utility[:, :, :, ji] - utility[:, :, :, av]
232
+ encoding_contribution = encoding[:, :, :, av] - encoding[:, :, :, ji]
233
+ estimation_contribution = estimation[:, :, :, av] - estimation[:, :, :, ji]
234
+ assert np.max(np.abs(delta - encoding_contribution - estimation_contribution)) < 1e-12
235
+ # Reproduce every saved native world score, not only a headline slice.
236
+ maximum_error = 0.0
237
+ primary_replay_error = 0.0
238
+ checked = 0
239
+ discrepancies = []
240
+ with gzip.open(release / "results/postconfirmation/controlled_worlds.csv.gz", "rt") as f:
241
+ world_index = {w: i for i, w in enumerate(indices["world"])}
242
+ for row in csv.DictReader(f):
243
+ key = (AVAILABILITY.index(int(row["available"])), SIZES.index(int(row["labels"])),
244
+ ESTIMATORS.index(row["estimator"]), int(row["stream"]), METHODS.index(row["method"]), world_index[row["world"]])
245
+ replayed = scores[key][1:, :, metric["utility"]].mean()
246
+ error = abs(float(row["utility"]) - replayed)
247
+ maximum_error = max(maximum_error, error)
248
+ if int(row["available"]) == 6 and row["estimator"] == "support_mixture" and row["method"] in ("identity", "query_joint", "action_values"):
249
+ primary_replay_error = max(primary_replay_error, error)
250
+ if error > 1e-10:
251
+ discrepancies.append(dict(**row, replayed_utility=float(replayed), difference=float(replayed)-float(row["utility"])))
252
+ checked += 1
253
+ # Preserve finite-precision/tie discrepancies outside the exact headline frontier.
254
+ # Never mix archived realized values with freshly recomputed code oracles.
255
+ assert checked == 124416 and primary_replay_error < 1e-10
256
+ if discrepancies:
257
+ csv_write(out / "replay_discrepancies.csv.gz", discrepancies)
258
+ native_rows = []
259
+ for ai, available in enumerate(AVAILABILITY):
260
+ for ni, n in enumerate(SIZES):
261
+ for ei, estimator in enumerate(ESTIMATORS):
262
+ for mi, method in enumerate(METHODS):
263
+ for wi, world in enumerate(indices["world"]):
264
+ values = positive[ai, ni, ei, mi, wi]
265
+ native_rows.append(dict(available=available, labels=n, estimator=estimator, method=method,
266
+ world=world, family=indices["family"][wi], **dict(zip(METRICS, values.tolist())),
267
+ encoding_loss=float(values[metric["available_oracle"]] - values[metric["encoded_oracle"]]),
268
+ estimation_loss=float(values[metric["encoded_oracle"]] - values[metric["utility"]])))
269
+ csv_write(out / "native_worlds.csv.gz", native_rows)
270
+ frontiers = []
271
+ for ai, available in enumerate(AVAILABILITY):
272
+ for ni, n in enumerate(SIZES):
273
+ for ei, estimator in enumerate(ESTIMATORS):
274
+ for mi, method in enumerate(METHODS):
275
+ for b in range(3):
276
+ for ci, cost in enumerate(COSTS):
277
+ v = worlds[ai, ni, ei, mi, :, b, ci].mean(0)
278
+ frontiers.append(dict(available=available, labels=n, estimator=estimator, method=method,
279
+ budget=b, cost=cost, **dict(zip(METRICS, v.tolist())),
280
+ encoding_loss=float(v[4] - v[5]), estimation_loss=float(v[5] - v[0])))
281
+ csv_write(out / "frontier_all_conditions.csv", frontiers)
282
+ weights, family_weights = bootstrap_weights(families)
283
+ # All conditions form one descriptive matrix, retaining every family/cost/budget.
284
+ flat_delta = delta.transpose(3, 0, 1, 2, 4, 5).reshape(192, -1)
285
+ paired_rows = []
286
+ for group in ["ALL"] + sorted(set(families)):
287
+ selected, w = (np.arange(192), weights) if group == "ALL" else family_weights[group]
288
+ boot = w @ flat_delta[selected]
289
+ interval = np.quantile(boot, [.025, .975], axis=0).reshape(2, 2, 4, 3, 3, 3)
290
+ for ai, available in enumerate(AVAILABILITY):
291
+ for ni, n in enumerate(SIZES):
292
+ for ei, estimator in enumerate(ESTIMATORS):
293
+ for b in range(3):
294
+ for ci, cost in enumerate(COSTS):
295
+ key = (ai, ni, ei)
296
+ oracle = worlds[key][ji, selected, b, ci, metric["available_oracle"]].mean()
297
+ a = utility[key][av, selected, b, ci].mean()
298
+ d = delta[key][selected, b, ci].mean()
299
+ paired_rows.append(dict(family=group, available=available, labels=n, estimator=estimator,
300
+ budget=b, cost=cost, worlds=len(selected), joint_minus_action=float(d),
301
+ paired_95_low=float(interval[0, ai, ni, ei, b, ci]),
302
+ paired_95_high=float(interval[1, ai, ni, ei, b, ci]),
303
+ encoding_contribution=float(encoding_contribution[key][selected, b, ci].mean()),
304
+ estimation_contribution=float(estimation_contribution[key][selected, b, ci].mean()),
305
+ action_remaining_headroom=float(oracle-a),
306
+ fraction_action_headroom_closed=float(d/(oracle-a)) if oracle-a > 1e-12 else None))
307
+ csv_write(out / "paired_by_condition.csv", paired_rows)
308
+ samples, family_samples = [], []
309
+ for group in ["ALL"] + sorted(set(families)):
310
+ selected, w = (np.arange(192), weights) if group == "ALL" else family_weights[group]
311
+ for ni, n in enumerate(SIZES):
312
+ v = positive[0, ni, 2]
313
+ d = v[ji, :, 0] - v[av, :, 0]
314
+ ci = np.quantile(w @ d[selected], [.025, .975])
315
+ econt = (v[av, :, 4]-v[av, :, 5]) - (v[ji, :, 4]-v[ji, :, 5])
316
+ mcont = (v[av, :, 5]-v[av, :, 0]) - (v[ji, :, 5]-v[ji, :, 0])
317
+ headroom = (v[av, selected, 4]-v[av, selected, 0]).mean()
318
+ row = dict(family=group, records=n, worlds=len(selected),
319
+ **{method: float(v[mi, selected, 0].mean()) for mi, method in enumerate(METHODS)},
320
+ true_oracle=float(v[0, selected, 4].mean()), joint_minus_action=float(d[selected].mean()),
321
+ paired_95_low=float(ci[0]), paired_95_high=float(ci[1]),
322
+ action_encoding_loss=float((v[av, selected, 4]-v[av, selected, 5]).mean()),
323
+ joint_encoding_loss=float((v[ji, selected, 4]-v[ji, selected, 5]).mean()),
324
+ action_estimation_loss=float((v[av, selected, 5]-v[av, selected, 0]).mean()),
325
+ joint_estimation_loss=float((v[ji, selected, 5]-v[ji, selected, 0]).mean()),
326
+ encoding_contribution=float(econt[selected].mean()), estimation_contribution=float(mcont[selected].mean()),
327
+ action_remaining_headroom=float(headroom),
328
+ fraction_action_headroom_closed=float(d[selected].mean()/headroom) if headroom > 1e-12 else None,
329
+ true_code_joint_oracle=float(v[ji, selected, 6].mean()),
330
+ true_code_action_oracle=float(v[av, selected, 6].mean()),
331
+ true_code_joint_fitted=float(v[ji, selected, 7].mean()),
332
+ true_code_action_fitted=float(v[av, selected, 7].mean()),
333
+ joint_root_choice_mismatch=float(v[ji, selected, 9].mean()),
334
+ action_root_choice_mismatch=float(v[av, selected, 9].mean()),
335
+ strongest_uncompressed_mean=max(float(positive[0, ni, ei, 0, selected, 0].mean()) for ei in range(3)))
336
+ (samples if group == "ALL" else family_samples).append(row)
337
+ csv_write(out / "sample_frontier.csv", samples)
338
+ csv_write(out / "family_sample_frontier.csv", family_samples)
339
+ summary = dict(status="POST_REVIEW_EXPLORATORY_OPENED_DATA_NO_NEW_CONFIRMATION", all_worlds=192,
340
+ native_saved_world_scores_checked=checked, native_saved_world_scores_within_1e_10=checked-len(discrepancies),
341
+ replay_discrepancy_count=len(discrepancies), maximum_absolute_replay_error=maximum_error,
342
+ exact_primary_mixture_joint_action_identity_frontier_maximum_replay_error=primary_replay_error,
343
+ exact_decomposition_max_error=float(np.max(np.abs(delta-encoding_contribution-estimation_contribution))),
344
+ primary_sample_frontier=samples, family_sample_frontier=family_samples,
345
+ all_conditions_reported=True, intervals="Descriptive paired within-family bootstrap, 10000 draws, seed 18701; no multiplicity correction or confirmatory interpretation",
346
+ limitations=["Constructed worlds share favorable finite-law structure; no external validity follows.",
347
+ "True-law encoders and true-law replanning have privileged access and are explanatory diagnostics only.",
348
+ "Exact additive loss separation identifies where the observed finite-grid gap arises; it does not identify a unique learning-theoretic cause or guarantee asymptotic dominance.",
349
+ "All algorithms retain the released deterministic initialization and 32 iterations; no attempt to optimize either clustering objective.",
350
+ "Headroom fractions are ratios of averages and not operational money or sensing savings.",
351
+ "Warm timing is a local batch observation on shared hardware; hardware startup and source data access are reported separately."])
352
+ if discrepancies:
353
+ summary["limitations"].append(f"{len(discrepancies)} archived native world means do not exactly replay in the expanded execution trace; all such rows are retained separately. Primary-mixture joint/action-value/identity world means at every sample budget replay with maximum error {primary_replay_error:.3g}. New full-grid diagnostics are recomputations, not replacements for frozen archival scores.")
354
+ dump(out / "summary.json", summary)
355
+ return summary
356
+
357
+
358
+ def benchmark(out, release, device, archive):
359
+ samples = archive["samples"][:, 0, :1024]
360
+ g = torch.Generator(device=device).manual_seed(18751)
361
+ ns = [1, 32, 1024, 65536]
362
+ def fixture():
363
+ selected = torch.randint(1024, (max(ns),), device=device, generator=g)
364
+ w = torch.arange(max(ns), device=device) % 192
365
+ events = samples[w, selected]
366
+ return (events//4) % 64, events % 4
367
+ (requests, future), fixture_seconds = timed(fixture)
368
+ records, warmups = [], []
369
+ for repetition in range(-1, 11):
370
+ order = METHODS if repetition % 2 == 0 else list(reversed(METHODS))
371
+ for method in order:
372
+ (counts, fitted), estimation_seconds = timed(lambda: (lambda counts: (counts, core.estimate(counts, "support_mixture")))(core.count_samples(samples, 6)))
373
+ profile_seconds = 0.0
374
+ if method in ("query_joint", "action_values"):
375
+ def profile():
376
+ mass = fitted.sum((2, 3))
377
+ features = action_profiles(fitted) if method == "action_values" else (fitted/mass[..., None, None]).flatten(2).sqrt()
378
+ return features, mass
379
+ (features, mass), profile_seconds = timed(profile)
380
+ code, encoding_seconds = timed(lambda: core.weighted_clusters(features, mass))
381
+ k = 8
382
+ else:
383
+ (code, k), encoding_seconds = timed(lambda: encode(fitted, method))
384
+ compressed, push_seconds = timed(lambda: core.push(fitted, code, k))
385
+ policy, planning_seconds = timed(lambda: core.plan(compressed))
386
+ def materialize():
387
+ deployed = dict(policy)
388
+ deployed["choices"] = policy["choices"].contiguous()
389
+ return (None if method == "identity" else code.to(torch.uint8).contiguous()), deployed
390
+ (deployed_code, deployed_policy), materialization_seconds = timed(materialize)
391
+ replay_times = {}
392
+ checksum = {}
393
+ for n in ns:
394
+ (actions, queries), elapsed = timed(lambda: replay(deployed_policy, deployed_code, requests[:n], future[:n]))
395
+ replay_times[str(n)] = elapsed
396
+ checksum[str(n)] = dict(actions=int(actions.sum()), queries=int(queries.sum()))
397
+ total = estimation_seconds+profile_seconds+encoding_seconds+push_seconds+planning_seconds+materialization_seconds
398
+ row = dict(repetition=repetition, method=method, estimation_seconds=estimation_seconds,
399
+ profile_seconds=profile_seconds, clustering_or_other_encoder_seconds=encoding_seconds,
400
+ push_seconds=push_seconds, exact_planning_seconds=planning_seconds,
401
+ materialization_seconds=materialization_seconds, complete_warm_construction_seconds=total,
402
+ replay_seconds=replay_times, checksums=checksum,
403
+ resident_law_bytes_per_world=compressed[0].numel()*compressed.element_size(),
404
+ resident_encoder_bytes_per_world=0 if deployed_code is None else deployed_code[0].numel()*deployed_code.element_size(),
405
+ deployed_choices_bytes_per_world=policy["choices"].numel()*policy["choices"].element_size()//192,
406
+ planner_values_bytes_per_world=policy["values"].numel()*policy["values"].element_size()//192,
407
+ count_state_bytes_per_world=counts[0].numel()*counts.element_size(),
408
+ source_events_bytes_per_world=samples[0].numel()*samples.element_size(),
409
+ temporary_profile_bytes_per_world=features[0].numel()*features.element_size() if method in ("query_joint", "action_values") else 0)
410
+ (warmups if repetition == -1 else records).append(row)
411
+ medians = {}
412
+ phases = ["estimation_seconds", "profile_seconds", "clustering_or_other_encoder_seconds", "push_seconds",
413
+ "exact_planning_seconds", "materialization_seconds", "complete_warm_construction_seconds"]
414
+ for method in METHODS:
415
+ rows = [r for r in records if r["method"] == method]
416
+ median = {p: float(np.median([r[p] for r in rows])) for p in phases}
417
+ warm = next(r for r in warmups if r["method"] == method)
418
+ median["replay_seconds"] = {str(n): float(np.median([r["replay_seconds"][str(n)] for r in rows])) for n in ns}
419
+ median["charged_warmup_seconds"] = warm["complete_warm_construction_seconds"] + sum(warm["replay_seconds"].values())
420
+ median["construction_plus_replay_seconds"] = {str(n): float(np.median([r["complete_warm_construction_seconds"]+r["replay_seconds"][str(n)] for r in rows])) for n in ns}
421
+ for k in warm:
422
+ if k.endswith("bytes_per_world"):
423
+ median[k] = warm[k]
424
+ medians[method] = median
425
+ original = json.loads((release / "provenance/controlled/confirmation_benchmark.json").read_text())
426
+ published = json.loads((release / "results/postconfirmation/controlled_timing.json").read_text())
427
+ old_medians = {m: float(np.median([r["complete_construction_seconds"] for r in published["repetitions"] if r["method"] == m])) for m in METHODS}
428
+ timing = dict(scope="192 worlds, 1024 records, stream 0, support mixture, float64 CUDA; 11 alternating-order warm repetitions plus one charged warmup",
429
+ shared_hardware=True, phase_synchronization_overhead_included=True, request_fixture_seconds=fixture_seconds,
430
+ fresh_medians=medians, fresh_records=records, fresh_warmups=warmups,
431
+ source_postconfirmation_medians_seconds=old_medians,
432
+ source_complete_benchmark={k: original[k] for k in ["scope", "cuda_startup_seconds", "cuda_warmup_seconds", "medians"]},
433
+ notes=["Phase synchronization can increase fresh construction latency relative to a fused production call.",
434
+ "The replay has fixed two-query budget and 0.01 query cost, matching the original transport benchmark; scientific utilities retain all nine cost/budget cells.",
435
+ "Retaining choices alone suffices for decisions; law/value-table storage is stated separately, not silently omitted.",
436
+ "Request fixtures are already recorded events, not sensor/backend access; operational sensing and data movement are not measured."])
437
+ dump(out / "timing.json", timing)
438
+ return timing
439
+
440
+
441
+ def numeric_diagnostic(out, release, device, archive):
442
+ """Keep encoder fixed; change only the mathematically irrelevant summation order.
443
+
444
+ This diagnoses finite-precision argmax sensitivity, without changing or
445
+ repairing the archived algorithms. It is never a utility competitor.
446
+ """
447
+ rows = []
448
+ values = np.empty((2, 4, 3, 3, 9, 192, 3, 3, 4))
449
+ manifest = json.loads((release / "data/controlled/manifest.json").read_text())["worlds"]
450
+ for ai, available in enumerate(AVAILABILITY):
451
+ truth = archive["joint"] if available == 6 else core.project(archive["joint"], 3)
452
+ for ni, n in enumerate(SIZES):
453
+ for ei, estimator in enumerate(ESTIMATORS):
454
+ for stream in range(3):
455
+ fitted = core.estimate(core.count_samples(archive["samples"][:, stream, :n], available), estimator)
456
+ for mi, method in enumerate(METHODS):
457
+ code, k = encode(fitted, method)
458
+ normal = core.push(fitted, code, k)
459
+ reversed_sum = core.push(fitted.flip(1), code.flip(1), k)
460
+ p, q = core.plan(normal), core.plan(reversed_sum)
461
+ gold = core.push(truth, code, k)
462
+ a, b = core.realized(gold, p)["utility"], core.realized(gold, q)["utility"]
463
+ law_difference = (normal-reversed_sum).abs().flatten(1).max(1).values
464
+ fitted_difference = (p["expected"]-q["expected"]).abs()
465
+ values[ai, ni, ei, stream, mi] = torch.stack([a, b, fitted_difference,
466
+ law_difference[:, None, None].expand_as(a)], -1).cpu().numpy()
467
+ for wi, world in enumerate(manifest):
468
+ rows.append(dict(available=available, labels=n, estimator=estimator, stream=stream,
469
+ method=method, world=world["id"], family=world["family"],
470
+ maximum_law_cell_difference=float(law_difference[wi]),
471
+ maximum_fitted_optimal_value_difference=float(fitted_difference[wi].max()),
472
+ maximum_realized_utility_difference=float((a[wi]-b[wi]).abs().max()),
473
+ positive_budget_average_utility_difference=float((b[wi, 1:]-a[wi, 1:]).mean())))
474
+ np.savez_compressed(out / "numeric_push_sensitivity.npz", scores=values)
475
+ csv_write(out / "numeric_push_sensitivity.csv.gz", rows)
476
+ changed = [r for r in rows if r["maximum_realized_utility_difference"] > 1e-10]
477
+ report = dict(status="POST_REVIEW_NUMERICAL_DIAGNOSTIC_NOT_A_COMPETITOR", code_sha256=sha(HERE),
478
+ operation="Hold each encoder fixed. Compare ordinary pushforward to reversal of raw-state accumulation order, which is the same probability law in exact arithmetic.",
479
+ metric_axis=["ordinary_realized_utility", "reversed_sum_realized_utility", "absolute_fitted_optimal_value_difference", "maximum_law_cell_difference"],
480
+ all_world_stream_method_rows=len(rows), changed_world_stream_method_rows=len(changed),
481
+ maximum_law_cell_difference=max(r["maximum_law_cell_difference"] for r in rows),
482
+ maximum_fitted_optimal_value_difference=max(r["maximum_fitted_optimal_value_difference"] for r in rows),
483
+ maximum_realized_utility_difference=max(r["maximum_realized_utility_difference"] for r in rows),
484
+ changed_by_estimator={e: sum(r["estimator"] == e for r in changed) for e in ESTIMATORS},
485
+ interpretation="Tiny law perturbations can select different fitted-optimal tied or nearly tied actions whose true-law utilities differ. This demonstrates numerical tie sensitivity; it does not uniquely reconstruct every archived discrepancy or justify replacing frozen scores.",
486
+ strongest_witnesses=sorted(changed, key=lambda r: r["maximum_realized_utility_difference"], reverse=True)[:10])
487
+ dump(out / "numeric_diagnostic.json", report)
488
+ print(json.dumps(report), flush=True)
489
+ return report
490
+
491
+
492
+ def main():
493
+ parser = argparse.ArgumentParser(description=__doc__)
494
+ parser.add_argument("--release", type=Path, default=DEFAULT_RELEASE)
495
+ parser.add_argument("--output", required=True, type=Path)
496
+ parser.add_argument("--device", default="cuda:4")
497
+ parser.add_argument("--prepare", action="store_true")
498
+ parser.add_argument("--resume-aggregation", action="store_true", help="Continue an explicitly recorded completed grid after a validation/reporting repair.")
499
+ parser.add_argument("--reference-only", action="store_true", help="Run the unmodified released postconfirmation.controlled function in a fresh reference_replay directory.")
500
+ parser.add_argument("--numeric-only", action="store_true", help="Only run the recorded fixed-code accumulation-order sensitivity diagnostic.")
501
+ args = parser.parse_args()
502
+ out, release = args.output.resolve(), args.release.resolve()
503
+ if args.prepare:
504
+ prepare(out, release, args.device)
505
+ return
506
+ spec = json.loads((out / "specification.json").read_text())
507
+ assert spec["code_sha256"] == sha(HERE), "Code changed after the recorded specification."
508
+ if not (args.resume_aggregation or args.reference_only or args.numeric_only):
509
+ assert not (out / "scores.npz").exists(), "Never overwrite a result."
510
+ for relative, expected in spec["input_sha256"].items():
511
+ assert sha(release / relative) == expected, relative
512
+ global np, torch, core, load, encode, action_profiles, replay
513
+ import numpy as np
514
+ import torch
515
+ sys.path.insert(0, str(release))
516
+ from reproduce import load, verify
517
+ from postconfirmation import encode, action_profiles
518
+ from unionmind.decision_sufficiency import core
519
+ from unionmind.decision_sufficiency.benchmark import replay
520
+ assert args.device.startswith("cuda") and torch.cuda.is_available()
521
+ torch.cuda.set_device(args.device)
522
+ torch.backends.cuda.matmul.allow_tf32 = False
523
+ torch.set_num_threads(4)
524
+ _, startup_seconds = timed(lambda: torch.cuda.init())
525
+ files_verified = verify()
526
+ analysis_start = time.perf_counter()
527
+ with torch.no_grad():
528
+ if args.reference_only:
529
+ from postconfirmation import controlled
530
+ reference = out / "reference_replay"
531
+ reference.mkdir()
532
+ controlled(reference, args.device)
533
+ print(json.dumps(dict(status="REFERENCE_REPLAY_COMPLETED", output=str(reference))))
534
+ return
535
+ if args.numeric_only:
536
+ assert not (out / "numeric_push_sensitivity.npz").exists()
537
+ numeric_diagnostic(out, release, args.device, load("data/controlled/worlds", args.device))
538
+ dump(out / "numeric_compute.json", dict(code_sha256=sha(HERE),
539
+ elapsed_analysis_seconds=time.perf_counter()-analysis_start,
540
+ elapsed_process_seconds=time.perf_counter()-PROCESS_START,
541
+ peak_cuda_allocated_bytes=torch.cuda.max_memory_allocated(), device=args.device,
542
+ original_files_modified=False, new_trainable_parameters=0))
543
+ return
544
+ if args.resume_aggregation:
545
+ scores = np.load(out / "scores.npz")["scores"]
546
+ indices = json.loads((out / "indices.json").read_text())
547
+ archive = load("data/controlled/worlds", args.device)
548
+ else:
549
+ scores, indices, archive = collect(out, release, args.device)
550
+ summary = summarize(out, release, scores, indices)
551
+ timing = benchmark(out, release, args.device, archive)
552
+ numeric_diagnostic(out, release, args.device, archive)
553
+ sync()
554
+ ledger = dict(created_utc=dt.datetime.now(dt.timezone.utc).isoformat(), status="COMPLETED",
555
+ source_release_files_hash_verified=files_verified, code_sha256=sha(HERE), specification_sha256=sha(out / "specification.json"),
556
+ command=sys.argv, device=args.device, gpu=torch.cuda.get_device_name(args.device),
557
+ python=sys.version, pytorch=torch.__version__, cuda=torch.version.cuda,
558
+ elapsed_analysis_seconds=time.perf_counter()-analysis_start,
559
+ elapsed_since_standard_library_import_seconds=time.perf_counter()-PROCESS_START,
560
+ first_cuda_noop_seconds=startup_seconds, peak_cuda_allocated_bytes=torch.cuda.max_memory_allocated(),
561
+ opened_worlds=True, new_backbone_executions=0, new_trainable_parameters=0, refitted_neural_models=0,
562
+ fitted_statistical_tables="Recomputed original estimators from archived counts, unchanged formulas",
563
+ original_files_modified=False, all_selected_units_retained=True,
564
+ all_statistics_descriptive=True, promotion_decisions_changed=False)
565
+ dump(out / "compute.json", ledger)
566
+ (out / "replay_code.py").write_bytes(HERE.read_bytes())
567
+ dump(out / "output_sha256.json", {p.name: sha(p) for p in sorted(out.iterdir()) if p.is_file() and p.name != "output_sha256.json"})
568
+ print(json.dumps(dict(status="COMPLETED", sample_frontier=summary["primary_sample_frontier"], compute=ledger), indent=2))
569
+
570
+
571
+ if __name__ == "__main__":
572
+ main()
revision_r1/controlled/sample_frontier.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ family,records,worlds,identity,query_joint,posterior_clusters,posterior_bins,future_marginal,raw_information,raw_entropy,hash_code,action_values,true_oracle,joint_minus_action,paired_95_low,paired_95_high,action_encoding_loss,joint_encoding_loss,action_estimation_loss,joint_estimation_loss,encoding_contribution,estimation_contribution,action_remaining_headroom,fraction_action_headroom_closed,true_code_joint_oracle,true_code_action_oracle,true_code_joint_fitted,true_code_action_fitted,joint_root_choice_mismatch,action_root_choice_mismatch,strongest_uncompressed_mean
2
+ ALL,64,192,0.2861019748560038,0.26584992657807344,0.2590280944592631,0.26074362507619164,0.2639323021727651,0.2722343061169504,0.21481644347086729,0.10239193116286986,0.26390613849214967,0.7139850169951533,0.001943788085923736,0.00021308171557470748,0.003647181976886664,0.23247242290469453,0.22743575439464836,0.21760645559830893,0.22069933602243133,0.005036668510046183,-0.0030928804241224478,0.4500788785030035,0.004318772061441592,0.7137990767560151,0.7138471222579433,0.33015852361105136,0.32977290874844095,0.5728323618858121,0.5747586230569347,0.33512680859485605
3
+ ALL,256,192,0.47464585889067984,0.46153544711762456,0.3821578365055907,0.37086161330575007,0.3117767655744309,0.39616931811795314,0.3287390155602434,0.15862081020012567,0.4421891239875439,0.7139850169951533,0.019346323130080677,0.016004547809026087,0.02272642638711617,0.11063576081059036,0.07695151696279438,0.16116013219701888,0.1754980529147342,0.033684243847795996,-0.014337920717715315,0.27179589300760926,0.07117960067755348,0.7137990767560151,0.7138471222579433,0.5040257394630626,0.5033361534654197,0.5138373564572506,0.532667401596573,0.5503906781894238
4
+ ALL,1024,192,0.6421842631007378,0.651568196746697,0.5045968970651954,0.4901829803145068,0.3289223833583756,0.5502229409756352,0.5192618455652109,0.20875256335562897,0.6375946140433673,0.7139850169951533,0.013973582703329749,0.012156191041507863,0.01579707133440269,0.020288801609760987,0.010064365748209498,0.05610160134202489,0.052352454500246635,0.010224435861551489,0.00374914684177826,0.07639040295178588,0.18292327522017698,0.7137990767560151,0.7138471222579433,0.6654659197644641,0.6648982429373541,0.33084781836365224,0.3441463140858656,0.6556393678743125
5
+ ALL,4096,192,0.6935868271710071,0.7032230168798058,0.5519797982991866,0.5318145046183101,0.34847248127428415,0.6062565500482736,0.5793404529052973,0.22923292167747147,0.7023595499794807,0.7139850169951533,0.0008634669003249706,0.0005238745996855586,0.001229058529941919,0.00203548792576397,0.0015186955544880705,0.009589979089908366,0.009243304560859296,0.0005167923712758993,0.00034667452904907126,0.011625467015672337,0.07427373878063802,0.7137990767560151,0.7138471222579433,0.7059254373130734,0.7056991595823129,0.19543563872713965,0.19527462898701206,0.6946993832338948
revision_r1/controlled/specification.json ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "created_utc": "2026-11-09T14:38:24.151548+00:00",
3
+ "status": "Post-review exploratory analysis of already opened worlds; not confirmation",
4
+ "scientific_question": "Why does joint minus action-value utility shrink with sample size?",
5
+ "authorization": "User authorized additional reviewer-response experiments; no new approval requested.",
6
+ "source_release": "/home/nurgaly/experiment/unionmind/articles/make_1/release/huggingface-revised",
7
+ "code_sha256": "4a300dfc769473f5afb08e15e877f6c53dfff5f6d3a0abfe17e0043798eecbab",
8
+ "device": "cuda:4",
9
+ "input_sha256": {
10
+ "data/controlled/worlds.safetensors": "e8649aeaed970fa78c3f402bbe9e775c486f3cf94ce2494cfe98a8099dbf3c7c",
11
+ "data/controlled/worlds.json": "860e154c4ffd0690080ab8c263756d4340cc119fffa66e1cc41d1b42ac3848cb",
12
+ "data/controlled/manifest.json": "2b50155e7f4d198449e2499dad21ba82d51e94423a1711eca8de4a01ab9ca156",
13
+ "postconfirmation.py": "3365341c056cee9b153cc864bad864965c918e11e0cec1819d49c8875c68dcbf",
14
+ "unionmind/decision_sufficiency/core.py": "fa2f7405a8d0dc6ec96ebae1bb2a2b810d03ef1fd2877f5ebef7df1722992cee",
15
+ "results/postconfirmation/controlled_worlds.csv.gz": "bdb15095eadf96dff2aa316619a838b74024eece87c57143bcd44b4710e9547f",
16
+ "results/postconfirmation/controlled_frontiers.csv.gz": "946ebf8764ee3dfed5a741fa3a45e0cb5f8476960453716d9abebed0aa617e57",
17
+ "results/postconfirmation/controlled_timing.json": "082e8b981d9e6b13a07bfd3136d6afe08681bb8c702ba2c0c2b9a2d4966b6eea",
18
+ "provenance/controlled/confirmation_benchmark.json": "133420e45f42ff5dbacc8d01700aeba56bf65c02f5bc0b30bb7aff45a88bc249"
19
+ },
20
+ "units": "All 192 archived worlds, all six families, all three original nested sample streams",
21
+ "grid": {
22
+ "availability": [
23
+ 6,
24
+ 3
25
+ ],
26
+ "records": [
27
+ 64,
28
+ 256,
29
+ 1024,
30
+ 4096
31
+ ],
32
+ "estimators": [
33
+ "joint_jeffreys",
34
+ "naive_bayes",
35
+ "support_mixture"
36
+ ],
37
+ "methods": [
38
+ "identity",
39
+ "query_joint",
40
+ "posterior_clusters",
41
+ "posterior_bins",
42
+ "future_marginal",
43
+ "raw_information",
44
+ "raw_entropy",
45
+ "hash_code",
46
+ "action_values"
47
+ ],
48
+ "budgets": [
49
+ 0,
50
+ 1,
51
+ 2
52
+ ],
53
+ "costs": [
54
+ 0.0,
55
+ 0.01,
56
+ 0.05
57
+ ]
58
+ },
59
+ "primary_aggregation": "Average positive budgets, all costs and three streams within each world; equal worlds and six families",
60
+ "interventions": [
61
+ "Native: estimated-law encoder, estimated encoded-law exact planner, true-law evaluation",
62
+ "Replan the identical native code under the pushed true law to separate encoding and model-estimation loss",
63
+ "Privileged diagnostic: construct each unchanged encoder under the true law, then plan under either true or fitted pushed law; never a deployable competitor"
64
+ ],
65
+ "exact_decomposition": "J-A = (encoding_loss_A-encoding_loss_J) + (estimation_loss_A-estimation_loss_J)",
66
+ "encoding_loss": "V(available true raw law)-V(true law pushed through native code)",
67
+ "estimation_loss": "V(true law pushed through native code)-realized utility of fitted policy",
68
+ "alias_diagnostic": "Evaluate all unordered raw-state pairs with true current-posterior difference <=1e-12 (exact) and <=0.01 (fixed near-alias threshold), weighted by 2 P(e)P(e'); compare true optimal root choices at each cost and budget and whether each native code merges disagreeing aliases",
69
+ "intervals": "Descriptive paired 95% percentile bootstrap, 10,000 draws, seed 18701, worlds resampled within fixed families after averaging streams; intervals are not corrected and are not new confirmatory tests",
70
+ "practical_scale": "Joint-action gain divided by remaining true available-oracle headroom of action-values; report signed gain, not just a percentage",
71
+ "timing": "Retain published complete-cost records. Fresh warm alternating-order 11-repeat 192-world primary-batch timings include count/estimate, profile, clustering/other encoding, push, exact planning, materialization and 1/32/1024/65536 recorded-event replays. One warmup charged and separately reported. Timing is descriptive on shared hardware.",
72
+ "prohibition": "No tuning, no model refits, no changed codes/iterations/costs, no dropped units, no original results or manuscript edits, no promotion",
73
+ "planned_outputs": [
74
+ "scores.npz",
75
+ "indices.json",
76
+ "frontier_all_conditions.csv",
77
+ "paired_by_condition.csv",
78
+ "sample_frontier.csv",
79
+ "family_sample_frontier.csv",
80
+ "native_worlds.csv.gz",
81
+ "alias_pairs.csv.gz",
82
+ "alias_summary.csv",
83
+ "privileged_geometry.csv",
84
+ "timing.json",
85
+ "summary.json",
86
+ "compute.json"
87
+ ],
88
+ "amendment_01": {
89
+ "created_utc": "2026-11-09T14:43:15.749761+00:00",
90
+ "reason": "The expanded execution trace completed and saved all grid cells, then strict whole-grid native replay validation failed. 946/124416 archived world means differ; primary-mixture joint/action-value/identity at all sample sizes match within 3.34e-16. Preserve all discrepancies and original code/spec/log; aggregate fresh diagnostics without altering source results. Run the released original full grid as a separate trace comparison to investigate tie/numerical sensitivity.",
91
+ "previous_code_sha256": "db7360ee28143a5b9f1e16f373d887354f4824826d8d9b5ed64425e6b732a5dc",
92
+ "completed_grid_code": "collection_code_01.py",
93
+ "completed_grid_specification": "specification_01.json",
94
+ "analysis_changes": "Validation/reporting repair only; no new encoders, thresholds, unit exclusions, tuning, or changed data. Add --resume-aggregation and --reference-only. All native values/oracles in each decomposition come from the same execution trace."
95
+ },
96
+ "amendment_02": {
97
+ "created_utc": "2026-11-09T14:47:08.492706+00:00",
98
+ "reason": "The unmodified released full-grid controlled replay also differs from archive in 968/124416 native world means while exactly reproducing the headline frontier. Test finite-precision argmax sensitivity directly.",
99
+ "operation": "For all grid cells and fixed freshly constructed encoders, compare normal pushforward to accumulation over reversed raw-state order. This is mathematically the same law; measure maximum cell-law perturbation, fitted-optimal expected-value difference, and true-law realized-utility difference.",
100
+ "no_scientific_comparator_change": true,
101
+ "new_outputs": [
102
+ "numeric_push_sensitivity.npz",
103
+ "numeric_push_sensitivity.csv.gz",
104
+ "numeric_diagnostic.json",
105
+ "numeric_compute.json"
106
+ ]
107
+ }
108
+ }
revision_r1/controlled/specification_01.json ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "created_utc": "2026-11-09T14:38:24.151548+00:00",
3
+ "status": "Post-review exploratory analysis of already opened worlds; not confirmation",
4
+ "scientific_question": "Why does joint minus action-value utility shrink with sample size?",
5
+ "authorization": "User authorized additional reviewer-response experiments; no new approval requested.",
6
+ "source_release": "/home/nurgaly/experiment/unionmind/articles/make_1/release/huggingface-revised",
7
+ "code_sha256": "db7360ee28143a5b9f1e16f373d887354f4824826d8d9b5ed64425e6b732a5dc",
8
+ "device": "cuda:4",
9
+ "input_sha256": {
10
+ "data/controlled/worlds.safetensors": "e8649aeaed970fa78c3f402bbe9e775c486f3cf94ce2494cfe98a8099dbf3c7c",
11
+ "data/controlled/worlds.json": "860e154c4ffd0690080ab8c263756d4340cc119fffa66e1cc41d1b42ac3848cb",
12
+ "data/controlled/manifest.json": "2b50155e7f4d198449e2499dad21ba82d51e94423a1711eca8de4a01ab9ca156",
13
+ "postconfirmation.py": "3365341c056cee9b153cc864bad864965c918e11e0cec1819d49c8875c68dcbf",
14
+ "unionmind/decision_sufficiency/core.py": "fa2f7405a8d0dc6ec96ebae1bb2a2b810d03ef1fd2877f5ebef7df1722992cee",
15
+ "results/postconfirmation/controlled_worlds.csv.gz": "bdb15095eadf96dff2aa316619a838b74024eece87c57143bcd44b4710e9547f",
16
+ "results/postconfirmation/controlled_frontiers.csv.gz": "946ebf8764ee3dfed5a741fa3a45e0cb5f8476960453716d9abebed0aa617e57",
17
+ "results/postconfirmation/controlled_timing.json": "082e8b981d9e6b13a07bfd3136d6afe08681bb8c702ba2c0c2b9a2d4966b6eea",
18
+ "provenance/controlled/confirmation_benchmark.json": "133420e45f42ff5dbacc8d01700aeba56bf65c02f5bc0b30bb7aff45a88bc249"
19
+ },
20
+ "units": "All 192 archived worlds, all six families, all three original nested sample streams",
21
+ "grid": {
22
+ "availability": [
23
+ 6,
24
+ 3
25
+ ],
26
+ "records": [
27
+ 64,
28
+ 256,
29
+ 1024,
30
+ 4096
31
+ ],
32
+ "estimators": [
33
+ "joint_jeffreys",
34
+ "naive_bayes",
35
+ "support_mixture"
36
+ ],
37
+ "methods": [
38
+ "identity",
39
+ "query_joint",
40
+ "posterior_clusters",
41
+ "posterior_bins",
42
+ "future_marginal",
43
+ "raw_information",
44
+ "raw_entropy",
45
+ "hash_code",
46
+ "action_values"
47
+ ],
48
+ "budgets": [
49
+ 0,
50
+ 1,
51
+ 2
52
+ ],
53
+ "costs": [
54
+ 0.0,
55
+ 0.01,
56
+ 0.05
57
+ ]
58
+ },
59
+ "primary_aggregation": "Average positive budgets, all costs and three streams within each world; equal worlds and six families",
60
+ "interventions": [
61
+ "Native: estimated-law encoder, estimated encoded-law exact planner, true-law evaluation",
62
+ "Replan the identical native code under the pushed true law to separate encoding and model-estimation loss",
63
+ "Privileged diagnostic: construct each unchanged encoder under the true law, then plan under either true or fitted pushed law; never a deployable competitor"
64
+ ],
65
+ "exact_decomposition": "J-A = (encoding_loss_A-encoding_loss_J) + (estimation_loss_A-estimation_loss_J)",
66
+ "encoding_loss": "V(available true raw law)-V(true law pushed through native code)",
67
+ "estimation_loss": "V(true law pushed through native code)-realized utility of fitted policy",
68
+ "alias_diagnostic": "Evaluate all unordered raw-state pairs with true current-posterior difference <=1e-12 (exact) and <=0.01 (fixed near-alias threshold), weighted by 2 P(e)P(e'); compare true optimal root choices at each cost and budget and whether each native code merges disagreeing aliases",
69
+ "intervals": "Descriptive paired 95% percentile bootstrap, 10,000 draws, seed 18701, worlds resampled within fixed families after averaging streams; intervals are not corrected and are not new confirmatory tests",
70
+ "practical_scale": "Joint-action gain divided by remaining true available-oracle headroom of action-values; report signed gain, not just a percentage",
71
+ "timing": "Retain published complete-cost records. Fresh warm alternating-order 11-repeat 192-world primary-batch timings include count/estimate, profile, clustering/other encoding, push, exact planning, materialization and 1/32/1024/65536 recorded-event replays. One warmup charged and separately reported. Timing is descriptive on shared hardware.",
72
+ "prohibition": "No tuning, no model refits, no changed codes/iterations/costs, no dropped units, no original results or manuscript edits, no promotion",
73
+ "planned_outputs": [
74
+ "scores.npz",
75
+ "indices.json",
76
+ "frontier_all_conditions.csv",
77
+ "paired_by_condition.csv",
78
+ "sample_frontier.csv",
79
+ "family_sample_frontier.csv",
80
+ "native_worlds.csv.gz",
81
+ "alias_pairs.csv.gz",
82
+ "alias_summary.csv",
83
+ "privileged_geometry.csv",
84
+ "timing.json",
85
+ "summary.json",
86
+ "compute.json"
87
+ ]
88
+ }
revision_r1/controlled/specification_02.json ADDED
@@ -0,0 +1,96 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "created_utc": "2026-11-09T14:38:24.151548+00:00",
3
+ "status": "Post-review exploratory analysis of already opened worlds; not confirmation",
4
+ "scientific_question": "Why does joint minus action-value utility shrink with sample size?",
5
+ "authorization": "User authorized additional reviewer-response experiments; no new approval requested.",
6
+ "source_release": "/home/nurgaly/experiment/unionmind/articles/make_1/release/huggingface-revised",
7
+ "code_sha256": "aa82ccdc6a1bb3c7d1e92b9ab0c4c88478c7d26e0442fe91523036a63260b2aa",
8
+ "device": "cuda:4",
9
+ "input_sha256": {
10
+ "data/controlled/worlds.safetensors": "e8649aeaed970fa78c3f402bbe9e775c486f3cf94ce2494cfe98a8099dbf3c7c",
11
+ "data/controlled/worlds.json": "860e154c4ffd0690080ab8c263756d4340cc119fffa66e1cc41d1b42ac3848cb",
12
+ "data/controlled/manifest.json": "2b50155e7f4d198449e2499dad21ba82d51e94423a1711eca8de4a01ab9ca156",
13
+ "postconfirmation.py": "3365341c056cee9b153cc864bad864965c918e11e0cec1819d49c8875c68dcbf",
14
+ "unionmind/decision_sufficiency/core.py": "fa2f7405a8d0dc6ec96ebae1bb2a2b810d03ef1fd2877f5ebef7df1722992cee",
15
+ "results/postconfirmation/controlled_worlds.csv.gz": "bdb15095eadf96dff2aa316619a838b74024eece87c57143bcd44b4710e9547f",
16
+ "results/postconfirmation/controlled_frontiers.csv.gz": "946ebf8764ee3dfed5a741fa3a45e0cb5f8476960453716d9abebed0aa617e57",
17
+ "results/postconfirmation/controlled_timing.json": "082e8b981d9e6b13a07bfd3136d6afe08681bb8c702ba2c0c2b9a2d4966b6eea",
18
+ "provenance/controlled/confirmation_benchmark.json": "133420e45f42ff5dbacc8d01700aeba56bf65c02f5bc0b30bb7aff45a88bc249"
19
+ },
20
+ "units": "All 192 archived worlds, all six families, all three original nested sample streams",
21
+ "grid": {
22
+ "availability": [
23
+ 6,
24
+ 3
25
+ ],
26
+ "records": [
27
+ 64,
28
+ 256,
29
+ 1024,
30
+ 4096
31
+ ],
32
+ "estimators": [
33
+ "joint_jeffreys",
34
+ "naive_bayes",
35
+ "support_mixture"
36
+ ],
37
+ "methods": [
38
+ "identity",
39
+ "query_joint",
40
+ "posterior_clusters",
41
+ "posterior_bins",
42
+ "future_marginal",
43
+ "raw_information",
44
+ "raw_entropy",
45
+ "hash_code",
46
+ "action_values"
47
+ ],
48
+ "budgets": [
49
+ 0,
50
+ 1,
51
+ 2
52
+ ],
53
+ "costs": [
54
+ 0.0,
55
+ 0.01,
56
+ 0.05
57
+ ]
58
+ },
59
+ "primary_aggregation": "Average positive budgets, all costs and three streams within each world; equal worlds and six families",
60
+ "interventions": [
61
+ "Native: estimated-law encoder, estimated encoded-law exact planner, true-law evaluation",
62
+ "Replan the identical native code under the pushed true law to separate encoding and model-estimation loss",
63
+ "Privileged diagnostic: construct each unchanged encoder under the true law, then plan under either true or fitted pushed law; never a deployable competitor"
64
+ ],
65
+ "exact_decomposition": "J-A = (encoding_loss_A-encoding_loss_J) + (estimation_loss_A-estimation_loss_J)",
66
+ "encoding_loss": "V(available true raw law)-V(true law pushed through native code)",
67
+ "estimation_loss": "V(true law pushed through native code)-realized utility of fitted policy",
68
+ "alias_diagnostic": "Evaluate all unordered raw-state pairs with true current-posterior difference <=1e-12 (exact) and <=0.01 (fixed near-alias threshold), weighted by 2 P(e)P(e'); compare true optimal root choices at each cost and budget and whether each native code merges disagreeing aliases",
69
+ "intervals": "Descriptive paired 95% percentile bootstrap, 10,000 draws, seed 18701, worlds resampled within fixed families after averaging streams; intervals are not corrected and are not new confirmatory tests",
70
+ "practical_scale": "Joint-action gain divided by remaining true available-oracle headroom of action-values; report signed gain, not just a percentage",
71
+ "timing": "Retain published complete-cost records. Fresh warm alternating-order 11-repeat 192-world primary-batch timings include count/estimate, profile, clustering/other encoding, push, exact planning, materialization and 1/32/1024/65536 recorded-event replays. One warmup charged and separately reported. Timing is descriptive on shared hardware.",
72
+ "prohibition": "No tuning, no model refits, no changed codes/iterations/costs, no dropped units, no original results or manuscript edits, no promotion",
73
+ "planned_outputs": [
74
+ "scores.npz",
75
+ "indices.json",
76
+ "frontier_all_conditions.csv",
77
+ "paired_by_condition.csv",
78
+ "sample_frontier.csv",
79
+ "family_sample_frontier.csv",
80
+ "native_worlds.csv.gz",
81
+ "alias_pairs.csv.gz",
82
+ "alias_summary.csv",
83
+ "privileged_geometry.csv",
84
+ "timing.json",
85
+ "summary.json",
86
+ "compute.json"
87
+ ],
88
+ "amendment_01": {
89
+ "created_utc": "2026-11-09T14:43:15.749761+00:00",
90
+ "reason": "The expanded execution trace completed and saved all grid cells, then strict whole-grid native replay validation failed. 946/124416 archived world means differ; primary-mixture joint/action-value/identity at all sample sizes match within 3.34e-16. Preserve all discrepancies and original code/spec/log; aggregate fresh diagnostics without altering source results. Run the released original full grid as a separate trace comparison to investigate tie/numerical sensitivity.",
91
+ "previous_code_sha256": "db7360ee28143a5b9f1e16f373d887354f4824826d8d9b5ed64425e6b732a5dc",
92
+ "completed_grid_code": "collection_code_01.py",
93
+ "completed_grid_specification": "specification_01.json",
94
+ "analysis_changes": "Validation/reporting repair only; no new encoders, thresholds, unit exclusions, tuning, or changed data. Add --resume-aggregation and --reference-only. All native values/oracles in each decomposition come from the same execution trace."
95
+ }
96
+ }
revision_r1/controlled/summary.json ADDED
@@ -0,0 +1,949 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "POST_REVIEW_EXPLORATORY_OPENED_DATA_NO_NEW_CONFIRMATION",
3
+ "all_worlds": 192,
4
+ "native_saved_world_scores_checked": 124416,
5
+ "native_saved_world_scores_within_1e_10": 123470,
6
+ "replay_discrepancy_count": 946,
7
+ "maximum_absolute_replay_error": 0.1456998408107582,
8
+ "exact_primary_mixture_joint_action_identity_frontier_maximum_replay_error": 3.3306690738754696e-16,
9
+ "exact_decomposition_max_error": 1.6653345369377348e-16,
10
+ "primary_sample_frontier": [
11
+ {
12
+ "family": "ALL",
13
+ "records": 64,
14
+ "worlds": 192,
15
+ "identity": 0.2861019748560038,
16
+ "query_joint": 0.26584992657807344,
17
+ "posterior_clusters": 0.2590280944592631,
18
+ "posterior_bins": 0.26074362507619164,
19
+ "future_marginal": 0.2639323021727651,
20
+ "raw_information": 0.2722343061169504,
21
+ "raw_entropy": 0.21481644347086729,
22
+ "hash_code": 0.10239193116286986,
23
+ "action_values": 0.26390613849214967,
24
+ "true_oracle": 0.7139850169951533,
25
+ "joint_minus_action": 0.001943788085923736,
26
+ "paired_95_low": 0.00021308171557470748,
27
+ "paired_95_high": 0.003647181976886664,
28
+ "action_encoding_loss": 0.23247242290469453,
29
+ "joint_encoding_loss": 0.22743575439464836,
30
+ "action_estimation_loss": 0.21760645559830893,
31
+ "joint_estimation_loss": 0.22069933602243133,
32
+ "encoding_contribution": 0.005036668510046183,
33
+ "estimation_contribution": -0.0030928804241224478,
34
+ "action_remaining_headroom": 0.4500788785030035,
35
+ "fraction_action_headroom_closed": 0.004318772061441592,
36
+ "true_code_joint_oracle": 0.7137990767560151,
37
+ "true_code_action_oracle": 0.7138471222579433,
38
+ "true_code_joint_fitted": 0.33015852361105136,
39
+ "true_code_action_fitted": 0.32977290874844095,
40
+ "joint_root_choice_mismatch": 0.5728323618858121,
41
+ "action_root_choice_mismatch": 0.5747586230569347,
42
+ "strongest_uncompressed_mean": 0.33512680859485605
43
+ },
44
+ {
45
+ "family": "ALL",
46
+ "records": 256,
47
+ "worlds": 192,
48
+ "identity": 0.47464585889067984,
49
+ "query_joint": 0.46153544711762456,
50
+ "posterior_clusters": 0.3821578365055907,
51
+ "posterior_bins": 0.37086161330575007,
52
+ "future_marginal": 0.3117767655744309,
53
+ "raw_information": 0.39616931811795314,
54
+ "raw_entropy": 0.3287390155602434,
55
+ "hash_code": 0.15862081020012567,
56
+ "action_values": 0.4421891239875439,
57
+ "true_oracle": 0.7139850169951533,
58
+ "joint_minus_action": 0.019346323130080677,
59
+ "paired_95_low": 0.016004547809026087,
60
+ "paired_95_high": 0.02272642638711617,
61
+ "action_encoding_loss": 0.11063576081059036,
62
+ "joint_encoding_loss": 0.07695151696279438,
63
+ "action_estimation_loss": 0.16116013219701888,
64
+ "joint_estimation_loss": 0.1754980529147342,
65
+ "encoding_contribution": 0.033684243847795996,
66
+ "estimation_contribution": -0.014337920717715315,
67
+ "action_remaining_headroom": 0.27179589300760926,
68
+ "fraction_action_headroom_closed": 0.07117960067755348,
69
+ "true_code_joint_oracle": 0.7137990767560151,
70
+ "true_code_action_oracle": 0.7138471222579433,
71
+ "true_code_joint_fitted": 0.5040257394630626,
72
+ "true_code_action_fitted": 0.5033361534654197,
73
+ "joint_root_choice_mismatch": 0.5138373564572506,
74
+ "action_root_choice_mismatch": 0.532667401596573,
75
+ "strongest_uncompressed_mean": 0.5503906781894238
76
+ },
77
+ {
78
+ "family": "ALL",
79
+ "records": 1024,
80
+ "worlds": 192,
81
+ "identity": 0.6421842631007378,
82
+ "query_joint": 0.651568196746697,
83
+ "posterior_clusters": 0.5045968970651954,
84
+ "posterior_bins": 0.4901829803145068,
85
+ "future_marginal": 0.3289223833583756,
86
+ "raw_information": 0.5502229409756352,
87
+ "raw_entropy": 0.5192618455652109,
88
+ "hash_code": 0.20875256335562897,
89
+ "action_values": 0.6375946140433673,
90
+ "true_oracle": 0.7139850169951533,
91
+ "joint_minus_action": 0.013973582703329749,
92
+ "paired_95_low": 0.012156191041507863,
93
+ "paired_95_high": 0.01579707133440269,
94
+ "action_encoding_loss": 0.020288801609760987,
95
+ "joint_encoding_loss": 0.010064365748209498,
96
+ "action_estimation_loss": 0.05610160134202489,
97
+ "joint_estimation_loss": 0.052352454500246635,
98
+ "encoding_contribution": 0.010224435861551489,
99
+ "estimation_contribution": 0.00374914684177826,
100
+ "action_remaining_headroom": 0.07639040295178588,
101
+ "fraction_action_headroom_closed": 0.18292327522017698,
102
+ "true_code_joint_oracle": 0.7137990767560151,
103
+ "true_code_action_oracle": 0.7138471222579433,
104
+ "true_code_joint_fitted": 0.6654659197644641,
105
+ "true_code_action_fitted": 0.6648982429373541,
106
+ "joint_root_choice_mismatch": 0.33084781836365224,
107
+ "action_root_choice_mismatch": 0.3441463140858656,
108
+ "strongest_uncompressed_mean": 0.6556393678743125
109
+ },
110
+ {
111
+ "family": "ALL",
112
+ "records": 4096,
113
+ "worlds": 192,
114
+ "identity": 0.6935868271710071,
115
+ "query_joint": 0.7032230168798058,
116
+ "posterior_clusters": 0.5519797982991866,
117
+ "posterior_bins": 0.5318145046183101,
118
+ "future_marginal": 0.34847248127428415,
119
+ "raw_information": 0.6062565500482736,
120
+ "raw_entropy": 0.5793404529052973,
121
+ "hash_code": 0.22923292167747147,
122
+ "action_values": 0.7023595499794807,
123
+ "true_oracle": 0.7139850169951533,
124
+ "joint_minus_action": 0.0008634669003249706,
125
+ "paired_95_low": 0.0005238745996855586,
126
+ "paired_95_high": 0.001229058529941919,
127
+ "action_encoding_loss": 0.00203548792576397,
128
+ "joint_encoding_loss": 0.0015186955544880705,
129
+ "action_estimation_loss": 0.009589979089908366,
130
+ "joint_estimation_loss": 0.009243304560859296,
131
+ "encoding_contribution": 0.0005167923712758993,
132
+ "estimation_contribution": 0.00034667452904907126,
133
+ "action_remaining_headroom": 0.011625467015672337,
134
+ "fraction_action_headroom_closed": 0.07427373878063802,
135
+ "true_code_joint_oracle": 0.7137990767560151,
136
+ "true_code_action_oracle": 0.7138471222579433,
137
+ "true_code_joint_fitted": 0.7059254373130734,
138
+ "true_code_action_fitted": 0.7056991595823129,
139
+ "joint_root_choice_mismatch": 0.19543563872713965,
140
+ "action_root_choice_mismatch": 0.19527462898701206,
141
+ "strongest_uncompressed_mean": 0.6946993832338948
142
+ }
143
+ ],
144
+ "family_sample_frontier": [
145
+ {
146
+ "family": "context",
147
+ "records": 64,
148
+ "worlds": 32,
149
+ "identity": 0.09339283465187889,
150
+ "query_joint": 0.0691595408487013,
151
+ "posterior_clusters": 0.05884192092177272,
152
+ "posterior_bins": 0.06287654697819856,
153
+ "future_marginal": 0.07187382385734353,
154
+ "raw_information": 0.07728978551777194,
155
+ "raw_entropy": 0.09455913414529662,
156
+ "hash_code": 0.10673495166633204,
157
+ "action_values": 0.06409177954684019,
158
+ "true_oracle": 0.7880888584118835,
159
+ "joint_minus_action": 0.005067761301861108,
160
+ "paired_95_low": 0.0015311813058195526,
161
+ "paired_95_high": 0.008840865316316502,
162
+ "action_encoding_loss": 0.3813222379944849,
163
+ "joint_encoding_loss": 0.3669534072130477,
164
+ "action_estimation_loss": 0.34267484087055833,
165
+ "joint_estimation_loss": 0.3519759103501344,
166
+ "encoding_contribution": 0.014368830781437176,
167
+ "estimation_contribution": -0.00930106947957607,
168
+ "action_remaining_headroom": 0.7239970788650432,
169
+ "fraction_action_headroom_closed": 0.00699969854823926,
170
+ "true_code_joint_oracle": 0.7880888584118835,
171
+ "true_code_action_oracle": 0.7880888584118835,
172
+ "true_code_joint_fitted": 0.12334932494026522,
173
+ "true_code_action_fitted": 0.12491782030181064,
174
+ "joint_root_choice_mismatch": 0.6849965145249336,
175
+ "action_root_choice_mismatch": 0.6875732689753824,
176
+ "strongest_uncompressed_mean": 0.2821009589784481
177
+ },
178
+ {
179
+ "family": "context",
180
+ "records": 256,
181
+ "worlds": 32,
182
+ "identity": 0.34655353787902665,
183
+ "query_joint": 0.3367188419663062,
184
+ "posterior_clusters": 0.16416722984358104,
185
+ "posterior_bins": 0.1463272094176618,
186
+ "future_marginal": 0.1871258673613156,
187
+ "raw_information": 0.26397328263687364,
188
+ "raw_entropy": 0.25243683714316034,
189
+ "hash_code": 0.15529673741167777,
190
+ "action_values": 0.31618507415861496,
191
+ "true_oracle": 0.7880888584118835,
192
+ "joint_minus_action": 0.0205337678076913,
193
+ "paired_95_low": 0.01188513381910629,
194
+ "paired_95_high": 0.03021118308609348,
195
+ "action_encoding_loss": 0.11119897518661376,
196
+ "joint_encoding_loss": 0.07603887002395061,
197
+ "action_estimation_loss": 0.3607048090666547,
198
+ "joint_estimation_loss": 0.37533114642162657,
199
+ "encoding_contribution": 0.035160105162663144,
200
+ "estimation_contribution": -0.014626337354971848,
201
+ "action_remaining_headroom": 0.47190378425326845,
202
+ "fraction_action_headroom_closed": 0.043512615267927005,
203
+ "true_code_joint_oracle": 0.7880888584118835,
204
+ "true_code_action_oracle": 0.7880888584118835,
205
+ "true_code_joint_fitted": 0.3856346205277096,
206
+ "true_code_action_fitted": 0.38320113863439054,
207
+ "joint_root_choice_mismatch": 0.629779568015901,
208
+ "action_root_choice_mismatch": 0.6496391850613675,
209
+ "strongest_uncompressed_mean": 0.5904656904258859
210
+ },
211
+ {
212
+ "family": "context",
213
+ "records": 1024,
214
+ "worlds": 32,
215
+ "identity": 0.6650919319515614,
216
+ "query_joint": 0.6814637624980395,
217
+ "posterior_clusters": 0.23779038974868094,
218
+ "posterior_bins": 0.17124022355036714,
219
+ "future_marginal": 0.36268052770898085,
220
+ "raw_information": 0.5145775504161055,
221
+ "raw_entropy": 0.5794729711288229,
222
+ "hash_code": 0.2170459709045936,
223
+ "action_values": 0.667719627246173,
224
+ "true_oracle": 0.7880888584118835,
225
+ "joint_minus_action": 0.013744135251866517,
226
+ "paired_95_low": 0.009521684888968644,
227
+ "paired_95_high": 0.01821770096658289,
228
+ "action_encoding_loss": 0.01430655071271671,
229
+ "joint_encoding_loss": 0.006186382463298001,
230
+ "action_estimation_loss": 0.10606268045299369,
231
+ "joint_estimation_loss": 0.10043871345054588,
232
+ "encoding_contribution": 0.008120168249418708,
233
+ "estimation_contribution": 0.005623967002447808,
234
+ "action_remaining_headroom": 0.12036923116571041,
235
+ "fraction_action_headroom_closed": 0.11418312735540517,
236
+ "true_code_joint_oracle": 0.7880888584118835,
237
+ "true_code_action_oracle": 0.7880888584118835,
238
+ "true_code_joint_fitted": 0.6988418074434672,
239
+ "true_code_action_fitted": 0.6964377598607214,
240
+ "joint_root_choice_mismatch": 0.3473903728720782,
241
+ "action_root_choice_mismatch": 0.3575560574608555,
242
+ "strongest_uncompressed_mean": 0.7340548744395632
243
+ },
244
+ {
245
+ "family": "context",
246
+ "records": 4096,
247
+ "worlds": 32,
248
+ "identity": 0.766972716874273,
249
+ "query_joint": 0.7766349438098745,
250
+ "posterior_clusters": 0.26807855851736373,
251
+ "posterior_bins": 0.1733365486789426,
252
+ "future_marginal": 0.5632767718308346,
253
+ "raw_information": 0.5928278598851834,
254
+ "raw_entropy": 0.6885210614964501,
255
+ "hash_code": 0.24707180576088097,
256
+ "action_values": 0.7768669245376707,
257
+ "true_oracle": 0.7880888584118835,
258
+ "joint_minus_action": -0.00023198072779632983,
259
+ "paired_95_low": -0.001376293766611737,
260
+ "paired_95_high": 0.0008712872445568171,
261
+ "action_encoding_loss": 0.0006891208483935686,
262
+ "joint_encoding_loss": 0.00018064349661515988,
263
+ "action_estimation_loss": 0.010532813025819078,
264
+ "joint_estimation_loss": 0.011273271105393817,
265
+ "encoding_contribution": 0.0005084773517784087,
266
+ "estimation_contribution": -0.0007404580795747386,
267
+ "action_remaining_headroom": 0.011221933874212647,
268
+ "fraction_action_headroom_closed": -0.02067208115790168,
269
+ "true_code_joint_oracle": 0.7880888584118835,
270
+ "true_code_action_oracle": 0.7880888584118835,
271
+ "true_code_joint_fitted": 0.7795383990383913,
272
+ "true_code_action_fitted": 0.7790645144622885,
273
+ "joint_root_choice_mismatch": 0.1620783588267283,
274
+ "action_root_choice_mismatch": 0.1554586985921155,
275
+ "strongest_uncompressed_mean": 0.7708361142534652
276
+ },
277
+ {
278
+ "family": "independent",
279
+ "records": 64,
280
+ "worlds": 32,
281
+ "identity": 0.909510548381619,
282
+ "query_joint": 0.9071402048294446,
283
+ "posterior_clusters": 0.9072370207266524,
284
+ "posterior_bins": 0.9093082065269007,
285
+ "future_marginal": 0.9074681824272653,
286
+ "raw_information": 0.8896605460004141,
287
+ "raw_entropy": 0.688211917829685,
288
+ "hash_code": 0.44151472051268514,
289
+ "action_values": 0.9068080779301189,
290
+ "true_oracle": 0.928933733504975,
291
+ "joint_minus_action": 0.00033212689932572356,
292
+ "paired_95_low": -0.0004962724151037125,
293
+ "paired_95_high": 0.0012075628763129809,
294
+ "action_encoding_loss": 0.009646569741400077,
295
+ "joint_encoding_loss": 0.00982195667832023,
296
+ "action_estimation_loss": 0.012479085833456041,
297
+ "joint_estimation_loss": 0.011971571997210165,
298
+ "encoding_contribution": -0.00017538693692015217,
299
+ "estimation_contribution": 0.0005075138362458757,
300
+ "action_remaining_headroom": 0.02212565557485612,
301
+ "fraction_action_headroom_closed": 0.015010940498556656,
302
+ "true_code_joint_oracle": 0.927818092070147,
303
+ "true_code_action_oracle": 0.9281063650817161,
304
+ "true_code_joint_fitted": 0.9185297572467013,
305
+ "true_code_action_fitted": 0.9192315821057243,
306
+ "joint_root_choice_mismatch": 0.23523002477266935,
307
+ "action_root_choice_mismatch": 0.23312046998815447,
308
+ "strongest_uncompressed_mean": 0.9136191430312471
309
+ },
310
+ {
311
+ "family": "independent",
312
+ "records": 256,
313
+ "worlds": 32,
314
+ "identity": 0.9116024194164163,
315
+ "query_joint": 0.9098670177155274,
316
+ "posterior_clusters": 0.9100779075741663,
317
+ "posterior_bins": 0.9126143882919777,
318
+ "future_marginal": 0.905503332137763,
319
+ "raw_information": 0.896961421482459,
320
+ "raw_entropy": 0.6703717514265709,
321
+ "hash_code": 0.4619766730060043,
322
+ "action_values": 0.9090004625042527,
323
+ "true_oracle": 0.928933733504975,
324
+ "joint_minus_action": 0.0008665552112746971,
325
+ "paired_95_low": -0.0004301648384077622,
326
+ "paired_95_high": 0.0021276027149003647,
327
+ "action_encoding_loss": 0.006369755441728155,
328
+ "joint_encoding_loss": 0.0065092495744983395,
329
+ "action_estimation_loss": 0.01356351555899413,
330
+ "joint_estimation_loss": 0.01255746621494925,
331
+ "encoding_contribution": -0.00013949413277018416,
332
+ "estimation_contribution": 0.0010060493440448813,
333
+ "action_remaining_headroom": 0.019933271000722286,
334
+ "fraction_action_headroom_closed": 0.04347280540375422,
335
+ "true_code_joint_oracle": 0.927818092070147,
336
+ "true_code_action_oracle": 0.9281063650817161,
337
+ "true_code_joint_fitted": 0.9170722759812633,
338
+ "true_code_action_fitted": 0.9171807569702725,
339
+ "joint_root_choice_mismatch": 0.23936604222246965,
340
+ "action_root_choice_mismatch": 0.23868371954799117,
341
+ "strongest_uncompressed_mean": 0.9251088162712999
342
+ },
343
+ {
344
+ "family": "independent",
345
+ "records": 1024,
346
+ "worlds": 32,
347
+ "identity": 0.9100079337524132,
348
+ "query_joint": 0.913013489675506,
349
+ "posterior_clusters": 0.915318662687657,
350
+ "posterior_bins": 0.9168213117935321,
351
+ "future_marginal": 0.8954121957914773,
352
+ "raw_information": 0.9031660787545921,
353
+ "raw_entropy": 0.6489306322062101,
354
+ "hash_code": 0.470594937948292,
355
+ "action_values": 0.9133526580360904,
356
+ "true_oracle": 0.928933733504975,
357
+ "joint_minus_action": -0.00033916836058426117,
358
+ "paired_95_low": -0.0011080639841925125,
359
+ "paired_95_high": 0.00048298344501220073,
360
+ "action_encoding_loss": 0.006376698269395811,
361
+ "joint_encoding_loss": 0.00708733789545923,
362
+ "action_estimation_loss": 0.009204377199488828,
363
+ "joint_estimation_loss": 0.00883290593400967,
364
+ "encoding_contribution": -0.0007106396260634189,
365
+ "estimation_contribution": 0.0003714712654791577,
366
+ "action_remaining_headroom": 0.01558107546888464,
367
+ "fraction_action_headroom_closed": -0.021767968537318196,
368
+ "true_code_joint_oracle": 0.927818092070147,
369
+ "true_code_action_oracle": 0.9281063650817161,
370
+ "true_code_joint_fitted": 0.9209817567018171,
371
+ "true_code_action_fitted": 0.9206480724749362,
372
+ "joint_root_choice_mismatch": 0.22556813929192912,
373
+ "action_root_choice_mismatch": 0.21870599191031026,
374
+ "strongest_uncompressed_mean": 0.9280646584591856
375
+ },
376
+ {
377
+ "family": "independent",
378
+ "records": 4096,
379
+ "worlds": 32,
380
+ "identity": 0.9190137606668483,
381
+ "query_joint": 0.9212120370356884,
382
+ "posterior_clusters": 0.9233281672038094,
383
+ "posterior_bins": 0.9234185604697762,
384
+ "future_marginal": 0.9120224134643348,
385
+ "raw_information": 0.9084610246510227,
386
+ "raw_entropy": 0.6461753611274292,
387
+ "hash_code": 0.47480680272329256,
388
+ "action_values": 0.9224258752130949,
389
+ "true_oracle": 0.928933733504975,
390
+ "joint_minus_action": -0.0012138381774065114,
391
+ "paired_95_low": -0.0017147112652704927,
392
+ "paired_95_high": -0.000726757585522743,
393
+ "action_encoding_loss": 0.0038865080122648353,
394
+ "joint_encoding_loss": 0.004790920676205797,
395
+ "action_estimation_loss": 0.002621350279615183,
396
+ "joint_estimation_loss": 0.002930775793080733,
397
+ "encoding_contribution": -0.0009044126639409615,
398
+ "estimation_contribution": -0.0003094255134655499,
399
+ "action_remaining_headroom": 0.0065078582918800185,
400
+ "fraction_action_headroom_closed": -0.18651883968048918,
401
+ "true_code_joint_oracle": 0.927818092070147,
402
+ "true_code_action_oracle": 0.9281063650817161,
403
+ "true_code_joint_fitted": 0.9258990135155555,
404
+ "true_code_action_fitted": 0.9259546622363166,
405
+ "joint_root_choice_mismatch": 0.21008579345280287,
406
+ "action_root_choice_mismatch": 0.19455413681065953,
407
+ "strongest_uncompressed_mean": 0.9287125775280298
408
+ },
409
+ {
410
+ "family": "orientation_alias",
411
+ "records": 64,
412
+ "worlds": 32,
413
+ "identity": 0.1635502787070287,
414
+ "query_joint": 0.12885644874690116,
415
+ "posterior_clusters": 0.11989069408297903,
416
+ "posterior_bins": 0.12129218066799444,
417
+ "future_marginal": 0.11565627135211173,
418
+ "raw_information": 0.1194667363872123,
419
+ "raw_entropy": 0.08451253981799893,
420
+ "hash_code": 0.062005774703194766,
421
+ "action_values": 0.12837855808157933,
422
+ "true_oracle": 0.828928467957698,
423
+ "joint_minus_action": 0.00047789066532179854,
424
+ "paired_95_low": -0.005027820336695715,
425
+ "paired_95_high": 0.005779686114251358,
426
+ "action_encoding_loss": 0.3951517175523762,
427
+ "joint_encoding_loss": 0.4017012049487419,
428
+ "action_estimation_loss": 0.3053981923237425,
429
+ "joint_estimation_loss": 0.29837081426205503,
430
+ "encoding_contribution": -0.0065494873963657056,
431
+ "estimation_contribution": 0.007027378061687502,
432
+ "action_remaining_headroom": 0.7005499098761186,
433
+ "fraction_action_headroom_closed": 0.0006821650514612244,
434
+ "true_code_joint_oracle": 0.828928467957698,
435
+ "true_code_action_oracle": 0.828928467957698,
436
+ "true_code_joint_fitted": 0.23582020614293137,
437
+ "true_code_action_fitted": 0.23582020614293134,
438
+ "joint_root_choice_mismatch": 0.6245539383534132,
439
+ "action_root_choice_mismatch": 0.6283048497052103,
440
+ "strongest_uncompressed_mean": 0.3744516848229764
441
+ },
442
+ {
443
+ "family": "orientation_alias",
444
+ "records": 256,
445
+ "worlds": 32,
446
+ "identity": 0.48105690410695323,
447
+ "query_joint": 0.45413249913205544,
448
+ "posterior_clusters": 0.3267375095162406,
449
+ "posterior_bins": 0.3044234970537981,
450
+ "future_marginal": 0.17946386454389426,
451
+ "raw_information": 0.3385633917456905,
452
+ "raw_entropy": 0.26942870751962467,
453
+ "hash_code": 0.14858964279293768,
454
+ "action_values": 0.4108957134778612,
455
+ "true_oracle": 0.828928467957698,
456
+ "joint_minus_action": 0.04323678565419422,
457
+ "paired_95_low": 0.03091808741028508,
458
+ "paired_95_high": 0.056253384606153346,
459
+ "action_encoding_loss": 0.22593360992407183,
460
+ "joint_encoding_loss": 0.15025128394784587,
461
+ "action_estimation_loss": 0.19209914455576496,
462
+ "joint_estimation_loss": 0.22454468487779672,
463
+ "encoding_contribution": 0.07568232597622598,
464
+ "estimation_contribution": -0.03244554032203176,
465
+ "action_remaining_headroom": 0.4180327544798368,
466
+ "fraction_action_headroom_closed": 0.1034291815434278,
467
+ "true_code_joint_oracle": 0.828928467957698,
468
+ "true_code_action_oracle": 0.828928467957698,
469
+ "true_code_joint_fitted": 0.5438758040593306,
470
+ "true_code_action_fitted": 0.5438758040593306,
471
+ "joint_root_choice_mismatch": 0.5760678352632569,
472
+ "action_root_choice_mismatch": 0.6058617660589638,
473
+ "strongest_uncompressed_mean": 0.6522145924562783
474
+ },
475
+ {
476
+ "family": "orientation_alias",
477
+ "records": 1024,
478
+ "worlds": 32,
479
+ "identity": 0.7375432253689826,
480
+ "query_joint": 0.749695480009097,
481
+ "posterior_clusters": 0.5557157182324963,
482
+ "posterior_bins": 0.5477364417189148,
483
+ "future_marginal": 0.18531714088359752,
484
+ "raw_information": 0.6134052546318305,
485
+ "raw_entropy": 0.6043586411911206,
486
+ "hash_code": 0.2567112571360594,
487
+ "action_values": 0.7156206511171621,
488
+ "true_oracle": 0.828928467957698,
489
+ "joint_minus_action": 0.03407482889193493,
490
+ "paired_95_low": 0.026490100551048997,
491
+ "paired_95_high": 0.0415008522799025,
492
+ "action_encoding_loss": 0.040254364563248254,
493
+ "joint_encoding_loss": 0.01143065598560973,
494
+ "action_estimation_loss": 0.07305345227728763,
495
+ "joint_estimation_loss": 0.06780233196299122,
496
+ "encoding_contribution": 0.028823708577638524,
497
+ "estimation_contribution": 0.00525112031429641,
498
+ "action_remaining_headroom": 0.1133078168405359,
499
+ "fraction_action_headroom_closed": 0.30072796248373795,
500
+ "true_code_joint_oracle": 0.828928467957698,
501
+ "true_code_action_oracle": 0.828928467957698,
502
+ "true_code_joint_fitted": 0.766961318774352,
503
+ "true_code_action_fitted": 0.7669613187743521,
504
+ "joint_root_choice_mismatch": 0.3803245810966434,
505
+ "action_root_choice_mismatch": 0.4051527363935008,
506
+ "strongest_uncompressed_mean": 0.767114986178982
507
+ },
508
+ {
509
+ "family": "orientation_alias",
510
+ "records": 4096,
511
+ "worlds": 32,
512
+ "identity": 0.804304431565352,
513
+ "query_joint": 0.8195916562097663,
514
+ "posterior_clusters": 0.6530579291374521,
515
+ "posterior_bins": 0.648285517967734,
516
+ "future_marginal": 0.18829678111200815,
517
+ "raw_information": 0.6964865610322947,
518
+ "raw_entropy": 0.706006590523646,
519
+ "hash_code": 0.3003141222527754,
520
+ "action_values": 0.819480859898903,
521
+ "true_oracle": 0.828928467957698,
522
+ "joint_minus_action": 0.00011079631086339387,
523
+ "paired_95_low": -2.228212253385705e-05,
524
+ "paired_95_high": 0.0002634376906898875,
525
+ "action_encoding_loss": 0.0002445290260795034,
526
+ "joint_encoding_loss": 0.00019810226223412192,
527
+ "action_estimation_loss": 0.009203079032715649,
528
+ "joint_estimation_loss": 0.009138709485697637,
529
+ "encoding_contribution": 4.64267638453815e-05,
530
+ "estimation_contribution": 6.436954701801237e-05,
531
+ "action_remaining_headroom": 0.009447608058795152,
532
+ "fraction_action_headroom_closed": 0.011727445737998115,
533
+ "true_code_joint_oracle": 0.828928467957698,
534
+ "true_code_action_oracle": 0.828928467957698,
535
+ "true_code_joint_fitted": 0.8198637628550727,
536
+ "true_code_action_fitted": 0.8198637628550727,
537
+ "joint_root_choice_mismatch": 0.21189316146075493,
538
+ "action_root_choice_mismatch": 0.21099546987915718,
539
+ "strongest_uncompressed_mean": 0.8096451880020169
540
+ },
541
+ {
542
+ "family": "parity",
543
+ "records": 64,
544
+ "worlds": 32,
545
+ "identity": 0.06273197484460194,
546
+ "query_joint": 0.02431126538233747,
547
+ "posterior_clusters": 0.011591763903878482,
548
+ "posterior_bins": 0.011981140143721973,
549
+ "future_marginal": 0.015523777582997956,
550
+ "raw_information": 0.011150427588187746,
551
+ "raw_entropy": 0.020184840684643574,
552
+ "hash_code": 0.024015757122275827,
553
+ "action_values": 0.02252062020077706,
554
+ "true_oracle": 0.8347957761429843,
555
+ "joint_minus_action": 0.0017906451815604108,
556
+ "paired_95_low": -0.0038315227161669556,
557
+ "paired_95_high": 0.006652954501555841,
558
+ "action_encoding_loss": 0.46294985993286397,
559
+ "joint_encoding_loss": 0.4524966576959601,
560
+ "action_estimation_loss": 0.34932529600934337,
561
+ "joint_estimation_loss": 0.35798785306468683,
562
+ "encoding_contribution": 0.010453202236903802,
563
+ "estimation_contribution": -0.008662557055343396,
564
+ "action_remaining_headroom": 0.8122751559422072,
565
+ "fraction_action_headroom_closed": 0.0022044810412591444,
566
+ "true_code_joint_oracle": 0.8347957761429844,
567
+ "true_code_action_oracle": 0.8347957761429844,
568
+ "true_code_joint_fitted": 0.12286533950243893,
569
+ "true_code_action_fitted": 0.12288347183177817,
570
+ "joint_root_choice_mismatch": 0.5209659718033554,
571
+ "action_root_choice_mismatch": 0.5285378271397623,
572
+ "strongest_uncompressed_mean": 0.385086063569183
573
+ },
574
+ {
575
+ "family": "parity",
576
+ "records": 256,
577
+ "worlds": 32,
578
+ "identity": 0.4475888678222731,
579
+ "query_joint": 0.41757388132172324,
580
+ "posterior_clusters": 0.2750026478906139,
581
+ "posterior_bins": 0.23783397003296353,
582
+ "future_marginal": 0.1298958512570368,
583
+ "raw_information": 0.20886862239855064,
584
+ "raw_entropy": 0.18373006939535583,
585
+ "hash_code": 0.11691960704302928,
586
+ "action_values": 0.38718031379931517,
587
+ "true_oracle": 0.8347957761429843,
588
+ "joint_minus_action": 0.030393567522408023,
589
+ "paired_95_low": 0.019497351412313457,
590
+ "paired_95_high": 0.041409246567013835,
591
+ "action_encoding_loss": 0.22704928735222724,
592
+ "joint_encoding_loss": 0.15105877295815492,
593
+ "action_estimation_loss": 0.220566174991442,
594
+ "joint_estimation_loss": 0.2661631218631063,
595
+ "encoding_contribution": 0.07599051439407231,
596
+ "estimation_contribution": -0.04559694687166428,
597
+ "action_remaining_headroom": 0.4476154623436692,
598
+ "fraction_action_headroom_closed": 0.06790106705266699,
599
+ "true_code_joint_oracle": 0.8347957761429844,
600
+ "true_code_action_oracle": 0.8347957761429844,
601
+ "true_code_joint_fitted": 0.4773303993370155,
602
+ "true_code_action_fitted": 0.4773303993370155,
603
+ "joint_root_choice_mismatch": 0.5556145869981621,
604
+ "action_root_choice_mismatch": 0.5840915047008,
605
+ "strongest_uncompressed_mean": 0.6622060304284867
606
+ },
607
+ {
608
+ "family": "parity",
609
+ "records": 1024,
610
+ "worlds": 32,
611
+ "identity": 0.7505863638432653,
612
+ "query_joint": 0.7643539645704318,
613
+ "posterior_clusters": 0.5785860682616655,
614
+ "posterior_bins": 0.5650067039347257,
615
+ "future_marginal": 0.17619358394784873,
616
+ "raw_information": 0.491345522934861,
617
+ "raw_entropy": 0.5428645204900009,
618
+ "hash_code": 0.20833025114024675,
619
+ "action_values": 0.749467269935558,
620
+ "true_oracle": 0.8347957761429843,
621
+ "joint_minus_action": 0.01488669463487384,
622
+ "paired_95_low": 0.010441901192385099,
623
+ "paired_95_high": 0.019552134986879535,
624
+ "action_encoding_loss": 0.021393038552985848,
625
+ "joint_encoding_loss": 0.0075410150709803,
626
+ "action_estimation_loss": 0.06393546765444057,
627
+ "joint_estimation_loss": 0.06290079650157226,
628
+ "encoding_contribution": 0.013852023482005548,
629
+ "estimation_contribution": 0.0010346711528682918,
630
+ "action_remaining_headroom": 0.08532850620742641,
631
+ "fraction_action_headroom_closed": 0.17446332177297866,
632
+ "true_code_joint_oracle": 0.8347957761429844,
633
+ "true_code_action_oracle": 0.8347957761429844,
634
+ "true_code_joint_fitted": 0.7761508394516206,
635
+ "true_code_action_fitted": 0.7761508394516206,
636
+ "joint_root_choice_mismatch": 0.3591226824418755,
637
+ "action_root_choice_mismatch": 0.35639638530427453,
638
+ "strongest_uncompressed_mean": 0.7761873106567354
639
+ },
640
+ {
641
+ "family": "parity",
642
+ "records": 4096,
643
+ "worlds": 32,
644
+ "identity": 0.811457808557315,
645
+ "query_joint": 0.824131626938427,
646
+ "posterior_clusters": 0.694169383413467,
647
+ "posterior_bins": 0.6900830814062695,
648
+ "future_marginal": 0.18472462130093503,
649
+ "raw_information": 0.6242484323345312,
650
+ "raw_entropy": 0.6210217172753493,
651
+ "hash_code": 0.2392919195668825,
652
+ "action_values": 0.824102588730724,
653
+ "true_oracle": 0.8347957761429843,
654
+ "joint_minus_action": 2.9038207703040858e-05,
655
+ "paired_95_low": -1.3477962034352808e-06,
656
+ "paired_95_high": 7.886248154405023e-05,
657
+ "action_encoding_loss": 9.112519989473711e-05,
658
+ "joint_encoding_loss": 6.378990475829419e-05,
659
+ "action_estimation_loss": 0.010602062212365704,
660
+ "joint_estimation_loss": 0.010600359299799106,
661
+ "encoding_contribution": 2.7335295136442922e-05,
662
+ "estimation_contribution": 1.7029125665979361e-06,
663
+ "action_remaining_headroom": 0.010693187412260441,
664
+ "fraction_action_headroom_closed": 0.0027155801711421096,
665
+ "true_code_joint_oracle": 0.8347957761429844,
666
+ "true_code_action_oracle": 0.8347957761429844,
667
+ "true_code_joint_fitted": 0.8242244796942859,
668
+ "true_code_action_fitted": 0.8242244796942859,
669
+ "joint_root_choice_mismatch": 0.199476969506594,
670
+ "action_root_choice_mismatch": 0.1989397435570667,
671
+ "strongest_uncompressed_mean": 0.8150734693187245
672
+ },
673
+ {
674
+ "family": "redundant",
675
+ "records": 64,
676
+ "worlds": 32,
677
+ "identity": 0.5363358739730864,
678
+ "query_joint": 0.5113855171174998,
679
+ "posterior_clusters": 0.5036569357456289,
680
+ "posterior_bins": 0.5085041397628068,
681
+ "future_marginal": 0.5000048354948977,
682
+ "raw_information": 0.5772794551090056,
683
+ "raw_entropy": 0.4124257595149454,
684
+ "hash_code": -0.04050167310864608,
685
+ "action_values": 0.5053617140765039,
686
+ "true_oracle": 0.8364283751534423,
687
+ "joint_minus_action": 0.006023803040995961,
688
+ "paired_95_low": 0.0009731619737643064,
689
+ "paired_95_high": 0.011502247775086136,
690
+ "action_encoding_loss": 0.14576415220704214,
691
+ "joint_encoding_loss": 0.13364129983182016,
692
+ "action_estimation_loss": 0.1853025088698963,
693
+ "joint_estimation_loss": 0.1914015582041223,
694
+ "encoding_contribution": 0.012122852375221975,
695
+ "estimation_contribution": -0.006099049334226014,
696
+ "action_remaining_headroom": 0.3310666610769384,
697
+ "fraction_action_headroom_closed": 0.018195136355321676,
698
+ "true_code_joint_oracle": 0.8364283751534423,
699
+ "true_code_action_oracle": 0.8364283751534423,
700
+ "true_code_joint_fitted": 0.5890955182088521,
701
+ "true_code_action_fitted": 0.5890955182088521,
702
+ "joint_root_choice_mismatch": 0.8539832767484425,
703
+ "action_root_choice_mismatch": 0.8572854915108054,
704
+ "strongest_uncompressed_mean": 0.5662414160281566
705
+ },
706
+ {
707
+ "family": "redundant",
708
+ "records": 256,
709
+ "worlds": 32,
710
+ "identity": 0.6453415693659241,
711
+ "query_joint": 0.6198034615018286,
712
+ "posterior_clusters": 0.5877161849929605,
713
+ "posterior_bins": 0.591621214981205,
714
+ "future_marginal": 0.42260265019739973,
715
+ "raw_information": 0.6312017727896639,
716
+ "raw_entropy": 0.5514639211048289,
717
+ "hash_code": 0.019364542901451464,
718
+ "action_values": 0.6050694499969314,
719
+ "true_oracle": 0.8364283751534423,
720
+ "joint_minus_action": 0.014734011504897247,
721
+ "paired_95_low": 0.009603019224582635,
722
+ "paired_95_high": 0.01995868486298478,
723
+ "action_encoding_loss": 0.0932629369589013,
724
+ "joint_encoding_loss": 0.07785092527231657,
725
+ "action_estimation_loss": 0.13809598819760954,
726
+ "joint_estimation_loss": 0.13877398837929703,
727
+ "encoding_contribution": 0.015412011686584716,
728
+ "estimation_contribution": -0.0006780001816874687,
729
+ "action_remaining_headroom": 0.23135892515651085,
730
+ "fraction_action_headroom_closed": 0.06368464711240299,
731
+ "true_code_joint_oracle": 0.8364283751534423,
732
+ "true_code_action_oracle": 0.8364283751534423,
733
+ "true_code_joint_fitted": 0.653629840527671,
734
+ "true_code_action_fitted": 0.653629840527671,
735
+ "joint_root_choice_mismatch": 0.763869962732947,
736
+ "action_root_choice_mismatch": 0.7826760570524698,
737
+ "strongest_uncompressed_mean": 0.7105456685441063
738
+ },
739
+ {
740
+ "family": "redundant",
741
+ "records": 1024,
742
+ "worlds": 32,
743
+ "identity": 0.7653668596783447,
744
+ "query_joint": 0.7545480306631399,
745
+ "posterior_clusters": 0.6989231534206255,
746
+ "posterior_bins": 0.6916977366851904,
747
+ "future_marginal": 0.29795754465378854,
748
+ "raw_information": 0.7244842637951285,
749
+ "raw_entropy": 0.6854194139838374,
750
+ "hash_code": 0.04484946990041256,
751
+ "action_values": 0.7412029739417946,
752
+ "true_oracle": 0.8364283751534423,
753
+ "joint_minus_action": 0.013345056721345335,
754
+ "paired_95_low": 0.010143917478369981,
755
+ "paired_95_high": 0.016472364541550773,
756
+ "action_encoding_loss": 0.039402157560219295,
757
+ "joint_encoding_loss": 0.02814080307390973,
758
+ "action_estimation_loss": 0.05582324365142841,
759
+ "joint_estimation_loss": 0.05373954141639264,
760
+ "encoding_contribution": 0.01126135448630956,
761
+ "estimation_contribution": 0.0020837022350357753,
762
+ "action_remaining_headroom": 0.0952254012116477,
763
+ "fraction_action_headroom_closed": 0.14014177468976635,
764
+ "true_code_joint_oracle": 0.8364283751534423,
765
+ "true_code_action_oracle": 0.8364283751534423,
766
+ "true_code_joint_fitted": 0.7721932591100672,
767
+ "true_code_action_fitted": 0.7721932591100673,
768
+ "joint_root_choice_mismatch": 0.4023677500023982,
769
+ "action_root_choice_mismatch": 0.43296693574774525,
770
+ "strongest_uncompressed_mean": 0.7863078402596235
771
+ },
772
+ {
773
+ "family": "redundant",
774
+ "records": 4096,
775
+ "worlds": 32,
776
+ "identity": 0.8155857149751549,
777
+ "query_joint": 0.8194726488422938,
778
+ "posterior_clusters": 0.7193415226940727,
779
+ "posterior_bins": 0.6960627204154617,
780
+ "future_marginal": 0.18000890013383308,
781
+ "raw_information": 0.7543788456935281,
782
+ "raw_entropy": 0.752804281570126,
783
+ "hash_code": 0.05292377176364181,
784
+ "action_values": 0.8165418628657128,
785
+ "true_oracle": 0.8364283751534423,
786
+ "joint_minus_action": 0.002930785976580884,
787
+ "paired_95_low": 0.0019304463049228831,
788
+ "paired_95_high": 0.004100380433294095,
789
+ "action_encoding_loss": 0.00730164446795117,
790
+ "joint_encoding_loss": 0.0038787169871150502,
791
+ "action_estimation_loss": 0.012584867819778103,
792
+ "joint_estimation_loss": 0.013077009324033339,
793
+ "encoding_contribution": 0.00342292748083612,
794
+ "estimation_contribution": -0.0004921415042552363,
795
+ "action_remaining_headroom": 0.019886512287729273,
796
+ "fraction_action_headroom_closed": 0.14737556461266912,
797
+ "true_code_joint_oracle": 0.8364283751534423,
798
+ "true_code_action_oracle": 0.8364283751534423,
799
+ "true_code_joint_fitted": 0.8226174136373886,
800
+ "true_code_action_fitted": 0.8226174136373885,
801
+ "joint_root_choice_mismatch": 0.1746054019842278,
802
+ "action_root_choice_mismatch": 0.18535970463320198,
803
+ "strongest_uncompressed_mean": 0.8207655371219653
804
+ },
805
+ {
806
+ "family": "uninformative",
807
+ "records": 64,
808
+ "worlds": 32,
809
+ "identity": -0.04890966142219218,
810
+ "query_joint": -0.04575341745644379,
811
+ "posterior_clusters": -0.047049768625332646,
812
+ "posterior_bins": -0.04950046362247272,
813
+ "future_marginal": -0.026933077678025653,
814
+ "raw_information": -0.041441113900889705,
815
+ "raw_entropy": -0.010995531167365718,
816
+ "hash_code": 0.02058205608137745,
817
+ "action_values": -0.043723918882921195,
818
+ "true_oracle": 0.0667348907999358,
819
+ "joint_minus_action": -0.0020294985735225856,
820
+ "paired_95_low": -0.0039889178805482635,
821
+ "paired_95_high": -0.000114955541232761,
822
+ "action_encoding_loss": -2.0599841277224584e-18,
823
+ "joint_encoding_loss": -9.75781955236954e-19,
824
+ "action_estimation_loss": 0.11045880968285701,
825
+ "joint_estimation_loss": 0.11248830825637958,
826
+ "encoding_contribution": -1.0842021724855044e-18,
827
+ "estimation_contribution": -0.002029498573522584,
828
+ "action_remaining_headroom": 0.11045880968285701,
829
+ "fraction_action_headroom_closed": -0.018373351834494372,
830
+ "true_code_joint_oracle": 0.0667348907999358,
831
+ "true_code_action_oracle": 0.0667348907999358,
832
+ "true_code_joint_fitted": -0.008709004374880524,
833
+ "true_code_action_fitted": -0.01331114610045102,
834
+ "joint_root_choice_mismatch": 0.5172644451120586,
835
+ "action_root_choice_mismatch": 0.5137298310222932,
836
+ "strongest_uncompressed_mean": -0.04890966142219218
837
+ },
838
+ {
839
+ "family": "uninformative",
840
+ "records": 256,
841
+ "worlds": 32,
842
+ "identity": 0.015731854753485836,
843
+ "query_joint": 0.03111698106830645,
844
+ "posterior_clusters": 0.02924553921598229,
845
+ "posterior_bins": 0.03234940005689452,
846
+ "future_marginal": 0.046069027949175986,
847
+ "raw_information": 0.03744741765448152,
848
+ "raw_entropy": 0.04500280677191939,
849
+ "hash_code": 0.04957765804565338,
850
+ "action_values": 0.024803729988287868,
851
+ "true_oracle": 0.0667348907999358,
852
+ "joint_minus_action": 0.006313251080018578,
853
+ "paired_95_low": 0.0035633612177853723,
854
+ "paired_95_high": 0.009296090920171755,
855
+ "action_encoding_loss": -1.1926223897340549e-18,
856
+ "joint_encoding_loss": 5.421010862427522e-19,
857
+ "action_estimation_loss": 0.04193116081164793,
858
+ "joint_estimation_loss": 0.03561790973162935,
859
+ "encoding_contribution": -1.734723475976807e-18,
860
+ "estimation_contribution": 0.006313251080018579,
861
+ "action_remaining_headroom": 0.04193116081164793,
862
+ "fraction_action_headroom_closed": 0.15056227773844122,
863
+ "true_code_joint_oracle": 0.0667348907999358,
864
+ "true_code_action_oracle": 0.0667348907999358,
865
+ "true_code_joint_fitted": 0.046611496345385736,
866
+ "true_code_action_fitted": 0.04479898126383822,
867
+ "joint_root_choice_mismatch": 0.318326143510767,
868
+ "action_root_choice_mismatch": 0.3350521771578457,
869
+ "strongest_uncompressed_mean": 0.026880985378230898
870
+ },
871
+ {
872
+ "family": "uninformative",
873
+ "records": 1024,
874
+ "worlds": 32,
875
+ "identity": 0.02450926400985953,
876
+ "query_joint": 0.046334453063967655,
877
+ "posterior_clusters": 0.041247390040046704,
878
+ "posterior_bins": 0.048595464204311184,
879
+ "future_marginal": 0.055973307164560476,
880
+ "raw_information": 0.05435897532129334,
881
+ "raw_entropy": 0.05452489439127397,
882
+ "hash_code": 0.05498349310416952,
883
+ "action_values": 0.038204503983425536,
884
+ "true_oracle": 0.0667348907999358,
885
+ "joint_minus_action": 0.008129949080542126,
886
+ "paired_95_low": 0.0048984844703365985,
887
+ "paired_95_high": 0.011408492793951204,
888
+ "action_encoding_loss": 4.553649124439119e-18,
889
+ "joint_encoding_loss": -3.3610267347050637e-18,
890
+ "action_estimation_loss": 0.028530386816510264,
891
+ "joint_estimation_loss": 0.020400437735968148,
892
+ "encoding_contribution": 7.914675859144182e-18,
893
+ "estimation_contribution": 0.008129949080542116,
894
+ "action_remaining_headroom": 0.028530386816510267,
895
+ "fraction_action_headroom_closed": 0.2849575483441185,
896
+ "true_code_joint_oracle": 0.0667348907999358,
897
+ "true_code_action_oracle": 0.0667348907999358,
898
+ "true_code_joint_fitted": 0.05766653710546027,
899
+ "true_code_action_fitted": 0.0569982079524269,
900
+ "joint_root_choice_mismatch": 0.27031338447698905,
901
+ "action_root_choice_mismatch": 0.29409977769850704,
902
+ "strongest_uncompressed_mean": 0.05871121561510939
903
+ },
904
+ {
905
+ "family": "uninformative",
906
+ "records": 4096,
907
+ "worlds": 32,
908
+ "identity": 0.044186530387099784,
909
+ "query_joint": 0.05829518844278467,
910
+ "posterior_clusters": 0.05390322882895525,
911
+ "posterior_bins": 0.05970059877167674,
912
+ "future_marginal": 0.06250539980375937,
913
+ "raw_information": 0.06113657669308175,
914
+ "raw_entropy": 0.061513705438783525,
915
+ "hash_code": 0.0609891079973558,
916
+ "action_values": 0.05473918863077931,
917
+ "true_oracle": 0.0667348907999358,
918
+ "joint_minus_action": 0.0035559998120053464,
919
+ "paired_95_low": 0.002404317887698644,
920
+ "paired_95_high": 0.004959589260649769,
921
+ "action_encoding_loss": 4.7704895589362195e-18,
922
+ "joint_encoding_loss": 5.421010862427522e-19,
923
+ "action_estimation_loss": 0.011995702169156477,
924
+ "joint_estimation_loss": 0.008439702357151135,
925
+ "encoding_contribution": 4.228388472693467e-18,
926
+ "estimation_contribution": 0.003555999812005342,
927
+ "action_remaining_headroom": 0.011995702169156483,
928
+ "fraction_action_headroom_closed": 0.2964394882317588,
929
+ "true_code_joint_oracle": 0.0667348907999358,
930
+ "true_code_action_oracle": 0.0667348907999358,
931
+ "true_code_joint_fitted": 0.06340955513774679,
932
+ "true_code_action_fitted": 0.0624701246085254,
933
+ "joint_root_choice_mismatch": 0.21447414713172985,
934
+ "action_root_choice_mismatch": 0.22634002044987156,
935
+ "strongest_uncompressed_mean": 0.06369397045613248
936
+ }
937
+ ],
938
+ "all_conditions_reported": true,
939
+ "intervals": "Descriptive paired within-family bootstrap, 10000 draws, seed 18701; no multiplicity correction or confirmatory interpretation",
940
+ "limitations": [
941
+ "Constructed worlds share favorable finite-law structure; no external validity follows.",
942
+ "True-law encoders and true-law replanning have privileged access and are explanatory diagnostics only.",
943
+ "Exact additive loss separation identifies where the observed finite-grid gap arises; it does not identify a unique learning-theoretic cause or guarantee asymptotic dominance.",
944
+ "All algorithms retain the released deterministic initialization and 32 iterations; no attempt to optimize either clustering objective.",
945
+ "Headroom fractions are ratios of averages and not operational money or sensing savings.",
946
+ "Warm timing is a local batch observation on shared hardware; hardware startup and source data access are reported separately.",
947
+ "946 archived native world means do not exactly replay in the expanded execution trace; all such rows are retained separately. Primary-mixture joint/action-value/identity world means at every sample budget replay with maximum error 3.33e-16. New full-grid diagnostics are recomputations, not replacements for frozen archival scores."
948
+ ]
949
+ }
revision_r1/controlled/timing.json ADDED
The diff for this file is too large to render. See raw diff
 
revision_r1/external/article_table.csv ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ analysis,common_current_joint_conditional_log_loss,fitted_compression_loss,gas_environments,gas_tasks,hydraulic_environments,hydraulic_tasks,joint_conditional_log_loss,labels_max,labels_min,noquery_utility,observation_given_h_log_loss,posterior_brier,posterior_log_loss,predicted_noquery_utility,predicted_query_increment,predicted_utility,queries,query_increment_optimism,realized_query_increment,tasks,utility
2
+ target_64_query_joint_on_target_evaluation_all57,0.9021537524971011,0.0014791213985141378,5,48,5,9,0.9202235047375056,64,64,0.4341858311122304,0.4713431260533678,0.14207717814233886,0.4488803786841378,0.5663140395782351,0.022600137737574737,0.5889141773158098,0.43706698485258466,0.028167567457767426,-0.005567429720192687,57,0.42861840139203766
3
+ target_64_posterior_clusters_on_target_evaluation_all57,0.9021537524971011,0.00206425583399923,5,48,5,9,0.9558448652595863,64,64,0.44368506699317145,0.5090721054033512,0.14116722448105434,0.4467727598562351,0.5668075955170042,0.021521447363320578,0.5883290428803247,0.4238130422310427,0.03794864053802975,-0.01642719317470917,57,0.42725787381846236
4
+ target_64_identity_on_target_evaluation_all57,0.9021537524971011,0.0,5,48,5,9,0.895850072130834,64,64,0.44420340770454436,0.4495203347543967,0.1413264442896859,0.4463297373764372,0.5674811876349388,0.02291211107938515,0.5903932987143239,0.42518703941463853,0.03604549979708882,-0.013133388717703665,57,0.4310700189868407
5
+ target_all_query_joint_on_target_evaluation_all57,0.6960366903855149,0.004799697896160156,5,48,5,9,0.7233527889368968,823,67,0.47670002693270863,0.32887130510623214,0.12778584484692165,0.3944814838306647,0.4952529448146883,0.01462258567675154,0.5098755304914399,0.4392621835902733,0.024546899602511985,-0.009924313925760445,57,0.4667757130069482
6
+ target_all_posterior_clusters_on_target_evaluation_all57,0.6960366903855149,0.003328144327637535,5,48,5,9,0.7876505767188415,823,67,0.47802205359215116,0.39681467189397057,0.12695844593296693,0.390835904824871,0.4996075681671639,0.011739515892798504,0.5113470840599624,0.43890637105567376,0.020765217143797324,-0.009025701250998821,57,0.4689963523411524
7
+ target_all_identity_on_target_evaluation_all57,0.6960366903855149,0.0,5,48,5,9,0.7078394813768716,823,67,0.4781249221284385,0.3163960529068561,0.1271812908421664,0.39144342847001556,0.5004346730166809,0.014240555370919127,0.5146752283876,0.44147345706354346,0.025157372572011803,-0.01091681720109267,57,0.46720810492734577
8
+ source_64_identity_on_source_held_all57,1.2938847482373521,0.0,5,48,5,9,1.3487109515234905,64,64,0.33440662650937636,0.8708436382289805,0.16025896218096947,0.47786731329451004,0.722022254088492,0.03006104296153668,0.7520832970500287,0.43304080035036646,-0.10170635547664639,0.13176739843818305,57,0.46617402494755944
9
+ source_64_identity_on_target_evaluation_all57,1.8532995603728606,0.0,5,48,5,9,1.9108555023828004,64,64,0.06673920740483243,1.052876482477108,0.2517142747070905,0.8579790199056924,0.722022254088492,0.03006104296153668,0.7520832970500287,0.483762359836609,-0.0035392117736777126,0.03360025473521439,57,0.10033946214004681
10
+ source_all_identity_on_source_held_all57,0.6404937484025411,0.0,5,48,5,9,0.7595273468083056,1430,400,0.731124166423397,0.5233765353700208,0.0719639418595227,0.23615081143828473,0.621897176545058,0.021625598675926755,0.6435227752209849,0.3886282868792795,0.05550853631963279,-0.033882937643706035,57,0.697241228779691
11
+ source_all_identity_on_target_evaluation_all57,1.3572171839606018,0.0,5,48,5,9,1.4448199908242385,1430,400,0.15937529013976298,0.7240851456651568,0.22750526451333272,0.7207348451590816,0.621897176545058,0.021625598675926755,0.6435227752209849,0.5231779414396178,-0.0766275842845273,0.09825318296045407,57,0.2576284731002171
12
+ target_64_query_joint_on_matched256eligible43,0.962944367808101,0.0016629665986308453,3,40,1,3,0.994952042783602,64,64,0.2933455394412311,0.4616198060448181,0.17695561042215174,0.533332236738784,0.4253933455778002,0.03476522133637247,0.46015856691417256,0.59982019504992,0.06686856540866529,-0.03210334407229281,43,0.2612421953689383
13
+ target_64_posterior_clusters_on_matched256eligible43,0.962944367808101,0.0017653999194657526,3,40,1,3,1.0543124169919997,64,64,0.29194760701280736,0.5212075436434427,0.17684661293020898,0.533104873348557,0.4261052889389456,0.03395084465439209,0.46005613359333775,0.5858980488484417,0.065767218097636,-0.031816373443243895,43,0.26013123356956347
14
+ target_64_identity_on_matched256eligible43,0.962944367808101,0.0,3,40,1,3,0.9812372654895479,64,64,0.2928115081984287,0.44856675165797866,0.17670638476417255,0.5326705138315693,0.4265949278730705,0.03522660563973295,0.4618215335128035,0.5999636913467485,0.06528806868106468,-0.030061463041331723,43,0.26275004515709705
15
+ target_256_query_joint_on_matched256eligible43,0.776763839151255,0.003964107645207928,3,40,1,3,0.7996329029836975,256,256,0.35585914250674133,0.2936103196675086,0.17152407443861453,0.5060225833161888,0.3512115814277825,0.023362338109154078,0.37457391953693664,0.564424019838484,0.027867767449796077,-0.0045054293406419955,43,0.3513537131660993
16
+ target_256_posterior_clusters_on_matched256eligible43,0.776763839151255,0.0028598850269441057,3,40,1,3,0.876625249191228,256,256,0.35662972148521666,0.371948609277301,0.1710897885778891,0.504676639913927,0.35368450489557457,0.021993637259625866,0.37567814215520046,0.548167026891073,0.02739402083024205,-0.00540038357061618,43,0.3512293379146005
17
+ target_256_identity_on_matched256eligible43,0.776763839151255,0.0,3,40,1,3,0.790112388208412,256,256,0.3565941836628872,0.28571420648711593,0.171004567859407,0.504398181721296,0.35440576895751386,0.024132258224630665,0.3785380271821446,0.5638333748702054,0.02748811340439971,-0.0033558551797690404,43,0.35323832848311815
18
+ empirical_alladaptation_query_joint_on_target_evaluation_all57,,6.807968017642793e-05,5,48,5,9,,823,67,0.46112566095566565,,,,0.6108515238336867,0.0013898851426896106,0.6122414089763762,0.02542517831387839,0.0013517400318673917,3.814511082221898e-05,57,0.4611638060664879
19
+ empirical_alladaptation_posterior_clusters_on_target_evaluation_all57,,2.4444444444429724e-05,5,48,5,9,,823,67,0.45795827338402395,,,,0.6109052872745468,0.0013797569375614061,0.6122850442121082,0.026571915606292424,0.001533767101958253,-0.00015401016439684685,57,0.4578042632196271
20
+ empirical_alladaptation_identity_on_target_evaluation_all57,,0.0,5,48,5,9,,823,67,0.4221120882269902,,,,0.6109052872745468,0.0014042013820058526,0.6123094886565527,0.025668404150831797,0.0015028618267391783,-9.866044473332557e-05,57,0.4220134277822568
revision_r1/external/execution.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "material_passport": {
3
+ "material_type": "experiment execution result",
4
+ "scientific_status": "post-review exploratory",
5
+ "created_utc": "2026-11-09T14:40:43.317363+00:00"
6
+ },
7
+ "status": "COMPLETED_FROM_SAVED_ARTIFACTS_AFTER_DIAGNOSTIC_DTYPE_CORRECTION",
8
+ "metric_rows": 13644,
9
+ "task_count": 57,
10
+ "original_fit_and_trace_seconds": 11.562392784981057,
11
+ "recovery_seconds": 9.258156107040122,
12
+ "maximum_replay_error": 6.661338147750939e-16,
13
+ "maximum_original_primary_expected_error": 2.220446049250313e-16,
14
+ "maximum_original_float32_query_charge_error": 1.4901161138336505e-09,
15
+ "gpu": "NVIDIA H200",
16
+ "device": "cuda:4",
17
+ "torch_version": "2.6.0+cu124",
18
+ "python_version": "3.13.11",
19
+ "peak_gpu_allocated_bytes": 63030784,
20
+ "peak_gpu_reserved_bytes": 81788928,
21
+ "script_sha256": "9739bca6e347271b91e4d7f9ffd4635b1bc7d58abe67d148f423b517fdea9d13",
22
+ "specification_sha256": "541dd0c1d4d7e242ee20c797bd9ef0928f7afdfffa4aacfeca5c3205cb212fab",
23
+ "original_results_and_gates_unchanged": true
24
+ }
revision_r1/external/implementation_correction.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "created_utc": "2026-11-09T14:40:20.960831+00:00",
3
+ "authorization": "Parent authorized routine numerical correction and artifact-only recovery after receiving failure report",
4
+ "failed_log_sha256": "9397cb2b52c3b201edd1b225736cef2f94060dcd026a2423e4765a7499b9ee49",
5
+ "original_specification_script_sha256": "18a499610df5927f78ae97af5bd4249b3c18beaef98ea39caecc5ad1303b7956",
6
+ "corrected_script_sha256": "9739bca6e347271b91e4d7f9ffd4635b1bc7d58abe67d148f423b517fdea9d13",
7
+ "failure": "Final equality assertion rejected float32 intermediate cost*integer_query_count in diagnostic batched replay utility; policy tables and traces unaffected",
8
+ "correction": "Multiply cost by queries.double(), matching original independent replay utility convention",
9
+ "estimators_rerun": 0,
10
+ "encoders_rerun": 0,
11
+ "planners_rerun": 0,
12
+ "policies_changed": 0,
13
+ "recovery": "Compute verified metrics and summaries from all 57 already saved model/policy/record-trace artifacts"
14
+ }
revision_r1/external/metrics.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9813ff30348ad76cb8b672e022a5f3fe852b8f0ee13625e7a5799400cfc01120
3
+ size 1003440
revision_r1/external/overall.json ADDED
The diff for this file is too large to render. See raw diff
 
revision_r1/external/paired_contrasts.json ADDED
@@ -0,0 +1,1338 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "scientific_status": "post-review exploratory; unadjusted descriptive intervals; no new confirmation",
3
+ "bootstrap_unit": "Target environment stratified by fixed dataset; tasks and seeds collapsed first",
4
+ "contrasts": [
5
+ {
6
+ "label": "joint_minus_posterior_target_64",
7
+ "field": "utility",
8
+ "overall": 0.0013605275735753635,
9
+ "by_dataset": {
10
+ "gas": -0.0026355068094112443,
11
+ "hydraulic": 0.005356561956561971
12
+ },
13
+ "by_environment": {
14
+ "gas:batch10": -0.006561130874027278,
15
+ "gas:batch6": 0.002410469681347993,
16
+ "gas:batch7": 0.010816431988927917,
17
+ "gas:batch8": -0.019843304843304852,
18
+ "gas:batch9": 0.0,
19
+ "hydraulic:context0=20": 0.0,
20
+ "hydraulic:context0=3": 0.0,
21
+ "hydraulic:context1=73": -0.0012910052910051577,
22
+ "hydraulic:context1=80": -0.0014615384615384401,
23
+ "hydraulic:context1=90": 0.029535353535353453
24
+ },
25
+ "tasks": 57,
26
+ "descriptive_paired_95_interval": [
27
+ -0.005097829104425849,
28
+ 0.008910461444960618
29
+ ],
30
+ "paired_fit_rows": 171,
31
+ "environment_resampling_only": true
32
+ },
33
+ {
34
+ "label": "target_minus_source_64_query_joint",
35
+ "field": "utility",
36
+ "overall": 0.35509999172132917,
37
+ "by_dataset": {
38
+ "gas": 0.473884680377134,
39
+ "hydraulic": 0.23631530306552437
40
+ },
41
+ "by_environment": {
42
+ "gas:batch10": 0.6827465125509808,
43
+ "gas:batch6": 0.5016533358279275,
44
+ "gas:batch7": 0.36589534226978987,
45
+ "gas:batch8": 0.5727454577218728,
46
+ "gas:batch9": 0.24638275351509886,
47
+ "hydraulic:context0=20": 0.019041394335511972,
48
+ "hydraulic:context0=3": 0.9375619275189165,
49
+ "hydraulic:context1=73": 0.06611111111111119,
50
+ "hydraulic:context1=80": 0.0742307692307691,
51
+ "hydraulic:context1=90": 0.084631313131313
52
+ },
53
+ "tasks": 57,
54
+ "descriptive_paired_95_interval": [
55
+ 0.21639674731918707,
56
+ 0.5385383430980829
57
+ ],
58
+ "paired_fit_rows": 171,
59
+ "environment_resampling_only": true
60
+ },
61
+ {
62
+ "label": "target_minus_source_64_posterior_clusters",
63
+ "field": "utility",
64
+ "overall": 0.374339618686491,
65
+ "by_dataset": {
66
+ "gas": 0.49935518424102254,
67
+ "hydraulic": 0.24932405313195946
68
+ },
69
+ "by_environment": {
70
+ "gas:batch10": 0.6940641487563329,
71
+ "gas:batch6": 0.5654931501290714,
72
+ "gas:batch7": 0.36169786854927394,
73
+ "gas:batch8": 0.6189817502553353,
74
+ "gas:batch9": 0.25653900351509884,
75
+ "hydraulic:context0=20": -0.07336819172113289,
76
+ "hydraulic:context0=3": 1.074090003982477,
77
+ "hydraulic:context1=73": 0.07578306878306873,
78
+ "hydraulic:context1=80": 0.08100427350427342,
79
+ "hydraulic:context1=90": 0.08911111111111107
80
+ },
81
+ "tasks": 57,
82
+ "descriptive_paired_95_interval": [
83
+ 0.21297235433542372,
84
+ 0.5883940113046724
85
+ ],
86
+ "paired_fit_rows": 171,
87
+ "environment_resampling_only": true
88
+ },
89
+ {
90
+ "label": "target_minus_source_64_identity",
91
+ "field": "utility",
92
+ "overall": 0.33073055684679387,
93
+ "by_dataset": {
94
+ "gas": 0.45630625630836985,
95
+ "hydraulic": 0.20515485738521785
96
+ },
97
+ "by_environment": {
98
+ "gas:batch10": 0.6827996391750814,
99
+ "gas:batch6": 0.4698706071131909,
100
+ "gas:batch7": 0.3612545191061068,
101
+ "gas:batch8": 0.521282790410149,
102
+ "gas:batch9": 0.24632372573732098,
103
+ "hydraulic:context0=20": -0.01192447349310094,
104
+ "hydraulic:context0=3": 0.9138932696136995,
105
+ "hydraulic:context1=73": 0.03314814814814826,
106
+ "hydraulic:context1=80": 0.026829059829059736,
107
+ "hydraulic:context1=90": 0.06382828282828272
108
+ },
109
+ "tasks": 57,
110
+ "descriptive_paired_95_interval": [
111
+ 0.19386722406937337,
112
+ 0.516573172190588
113
+ ],
114
+ "paired_fit_rows": 171,
115
+ "environment_resampling_only": true
116
+ },
117
+ {
118
+ "label": "joint_minus_posterior_target_256",
119
+ "field": "utility",
120
+ "overall": 0.0001243752514988102,
121
+ "by_dataset": {
122
+ "gas": 0.0002487505029976204,
123
+ "hydraulic": 0.0
124
+ },
125
+ "by_environment": {
126
+ "gas:batch10": -0.001310365636289264,
127
+ "gas:batch6": 2.2926454800525085e-06,
128
+ "gas:batch7": 0.0020543244998020727,
129
+ "hydraulic:context0=3": 0.0
130
+ },
131
+ "tasks": 43,
132
+ "descriptive_paired_95_interval": [
133
+ -0.0006551828181446319,
134
+ 0.0010271622499010364
135
+ ],
136
+ "paired_fit_rows": 129,
137
+ "environment_resampling_only": true
138
+ },
139
+ {
140
+ "label": "target_minus_source_256_query_joint",
141
+ "field": "utility",
142
+ "overall": 0.6972687563149934,
143
+ "by_dataset": {
144
+ "gas": 0.41733799052723897,
145
+ "hydraulic": 0.9771995221027479
146
+ },
147
+ "by_environment": {
148
+ "gas:batch10": 0.6106930751014984,
149
+ "gas:batch6": 0.26386387194882227,
150
+ "gas:batch7": 0.37745702453139623,
151
+ "hydraulic:context0=3": 0.9771995221027479
152
+ },
153
+ "tasks": 43,
154
+ "descriptive_paired_95_interval": [
155
+ 0.620531697025785,
156
+ 0.7939462986021231
157
+ ],
158
+ "paired_fit_rows": 129,
159
+ "environment_resampling_only": true
160
+ },
161
+ {
162
+ "label": "target_minus_source_256_posterior_clusters",
163
+ "field": "utility",
164
+ "overall": 0.7542922085489637,
165
+ "by_dataset": {
166
+ "gas": 0.45509855489163514,
167
+ "hydraulic": 1.0534858622062924
168
+ },
169
+ "by_environment": {
170
+ "gas:batch10": 0.6495552048920008,
171
+ "gas:batch6": 0.34364683092011294,
172
+ "gas:batch7": 0.37209362886279174,
173
+ "hydraulic:context0=3": 1.0534858622062924
174
+ },
175
+ "tasks": 43,
176
+ "descriptive_paired_95_interval": [
177
+ 0.6985663465632026,
178
+ 0.8515205335491466
179
+ ],
180
+ "paired_fit_rows": 129,
181
+ "environment_resampling_only": true
182
+ },
183
+ {
184
+ "label": "target_minus_source_256_identity",
185
+ "field": "utility",
186
+ "overall": 0.6795324016704808,
187
+ "by_dataset": {
188
+ "gas": 0.41381390728361395,
189
+ "hydraulic": 0.9452508960573477
190
+ },
191
+ "by_environment": {
192
+ "gas:batch10": 0.5945331914895884,
193
+ "gas:batch6": 0.2842302213368665,
194
+ "gas:batch7": 0.362678309024387,
195
+ "hydraulic:context0=3": 0.9452508960573477
196
+ },
197
+ "tasks": 43,
198
+ "descriptive_paired_95_interval": [
199
+ 0.6147405586971071,
200
+ 0.769892043773468
201
+ ],
202
+ "paired_fit_rows": 129,
203
+ "environment_resampling_only": true
204
+ },
205
+ {
206
+ "label": "joint_minus_posterior_target_all",
207
+ "field": "utility",
208
+ "overall": -0.0022206393342042033,
209
+ "by_dataset": {
210
+ "gas": -0.0042490486761783755,
211
+ "hydraulic": -0.00019222999223003078
212
+ },
213
+ "by_environment": {
214
+ "gas:batch10": 0.00128551606023862,
215
+ "gas:batch6": -0.00404940473759051,
216
+ "gas:batch7": 0.004146850424665106,
217
+ "gas:batch8": -0.022628205128205094,
218
+ "gas:batch9": 0.0,
219
+ "hydraulic:context0=20": 0.0,
220
+ "hydraulic:context0=3": 0.0,
221
+ "hydraulic:context1=73": -0.012555555555555542,
222
+ "hydraulic:context1=80": 0.011230769230769155,
223
+ "hydraulic:context1=90": 0.00036363636363623275
224
+ },
225
+ "tasks": 57,
226
+ "descriptive_paired_95_interval": [
227
+ -0.007971387996951206,
228
+ 0.0026320522784431004
229
+ ],
230
+ "paired_fit_rows": 171,
231
+ "environment_resampling_only": true
232
+ },
233
+ {
234
+ "label": "target_minus_source_all_query_joint",
235
+ "field": "utility",
236
+ "overall": 0.25648045155959887,
237
+ "by_dataset": {
238
+ "gas": 0.4477259480328195,
239
+ "hydraulic": 0.06523495508637825
240
+ },
241
+ "by_environment": {
242
+ "gas:batch10": 0.6156761617886918,
243
+ "gas:batch6": 0.23057048031017344,
244
+ "gas:batch7": 0.40794457305418264,
245
+ "gas:batch8": 0.5168382115787776,
246
+ "gas:batch9": 0.46760031343227215,
247
+ "hydraulic:context0=20": -0.5321677559912854,
248
+ "hydraulic:context0=3": 0.8422102747909199,
249
+ "hydraulic:context1=73": 0.01523015873015876,
250
+ "hydraulic:context1=80": 0.03553846153846152,
251
+ "hydraulic:context1=90": -0.034636363636363576
252
+ },
253
+ "tasks": 57,
254
+ "descriptive_paired_95_interval": [
255
+ 0.07132033952671643,
256
+ 0.4667324712229042
257
+ ],
258
+ "paired_fit_rows": 171,
259
+ "environment_resampling_only": true
260
+ },
261
+ {
262
+ "label": "target_minus_source_all_posterior_clusters",
263
+ "field": "utility",
264
+ "overall": 0.26156087109101617,
265
+ "by_dataset": {
266
+ "gas": 0.4580375849238713,
267
+ "hydraulic": 0.06508415725816101
268
+ },
269
+ "by_environment": {
270
+ "gas:batch10": 0.6322578156867729,
271
+ "gas:batch6": 0.2463184991774343,
272
+ "gas:batch7": 0.40254274286375746,
273
+ "gas:batch8": 0.5360102201257863,
274
+ "gas:batch9": 0.4730586467656055,
275
+ "hydraulic:context0=20": -0.5141677559912854,
276
+ "hydraulic:context0=3": 0.7874157706093189,
277
+ "hydraulic:context1=73": 0.06619841269841265,
278
+ "hydraulic:context1=80": 0.04314102564102562,
279
+ "hydraulic:context1=90": -0.0571666666666667
280
+ },
281
+ "tasks": 57,
282
+ "descriptive_paired_95_interval": [
283
+ 0.08312789207608386,
284
+ 0.45910042894694186
285
+ ],
286
+ "paired_fit_rows": 171,
287
+ "environment_resampling_only": true
288
+ },
289
+ {
290
+ "label": "target_minus_source_all_identity",
291
+ "field": "utility",
292
+ "overall": 0.20957963182712874,
293
+ "by_dataset": {
294
+ "gas": 0.4421509491140644,
295
+ "hydraulic": -0.02299168545980692
296
+ },
297
+ "by_environment": {
298
+ "gas:batch10": 0.6093241353610174,
299
+ "gas:batch6": 0.23886951796836567,
300
+ "gas:batch7": 0.38617590540115165,
301
+ "gas:batch8": 0.508784873407515,
302
+ "gas:batch9": 0.46760031343227215,
303
+ "hydraulic:context0=20": -0.6987799564270153,
304
+ "hydraulic:context0=3": 0.5693249701314217,
305
+ "hydraulic:context1=73": 0.036039682539682705,
306
+ "hydraulic:context1=80": 0.03548717948717939,
307
+ "hydraulic:context1=90": -0.0570303030303031
308
+ },
309
+ "tasks": 57,
310
+ "descriptive_paired_95_interval": [
311
+ 0.02266356074770446,
312
+ 0.3862194467643001
313
+ ],
314
+ "paired_fit_rows": 171,
315
+ "environment_resampling_only": true
316
+ },
317
+ {
318
+ "label": "target_256_minus_64_query_joint",
319
+ "field": "utility",
320
+ "overall": 0.09011151779716105,
321
+ "by_dataset": {
322
+ "gas": 0.06971407502084541,
323
+ "hydraulic": 0.11050896057347669
324
+ },
325
+ "by_environment": {
326
+ "gas:batch10": 0.0756806253273861,
327
+ "gas:batch6": 0.09056258626038896,
328
+ "gas:batch7": 0.04289901347476115,
329
+ "hydraulic:context0=3": 0.11050896057347669
330
+ },
331
+ "tasks": 43,
332
+ "descriptive_paired_95_interval": [
333
+ 0.07670398702411892,
334
+ 0.10053577341693282
335
+ ],
336
+ "paired_fit_rows": 129,
337
+ "environment_resampling_only": true
338
+ },
339
+ {
340
+ "label": "target_256_minus_64_query_joint",
341
+ "field": "noquery_utility",
342
+ "overall": 0.06251360306551022,
343
+ "by_dataset": {
344
+ "gas": 0.05334261831740041,
345
+ "hydraulic": 0.07168458781362004
346
+ },
347
+ "by_environment": {
348
+ "gas:batch10": 0.08115054963730205,
349
+ "gas:batch6": 0.0496728948353739,
350
+ "gas:batch7": 0.029204410479525288,
351
+ "hydraulic:context0=3": 0.07168458781362004
352
+ },
353
+ "tasks": 43,
354
+ "descriptive_paired_95_interval": [
355
+ 0.05044449914657266,
356
+ 0.07641756872546104
357
+ ],
358
+ "paired_fit_rows": 129,
359
+ "environment_resampling_only": true
360
+ },
361
+ {
362
+ "label": "target_256_minus_64_query_joint",
363
+ "field": "realized_query_increment",
364
+ "overall": 0.02759791473165081,
365
+ "by_dataset": {
366
+ "gas": 0.016371456703444996,
367
+ "hydraulic": 0.038824372759856625
368
+ },
369
+ "by_environment": {
370
+ "gas:batch10": -0.00546992430991594,
371
+ "gas:batch6": 0.040889691425015064,
372
+ "gas:batch7": 0.013694602995235868,
373
+ "hydraulic:context0=3": 0.038824372759856625
374
+ },
375
+ "tasks": 43,
376
+ "descriptive_paired_95_interval": [
377
+ 0.016677224224970342,
378
+ 0.039857032092435844
379
+ ],
380
+ "paired_fit_rows": 129,
381
+ "environment_resampling_only": true
382
+ },
383
+ {
384
+ "label": "target_256_minus_64_query_joint",
385
+ "field": "joint_conditional_log_loss",
386
+ "overall": -0.1953191397999047,
387
+ "by_dataset": {
388
+ "gas": -0.2657428747277448,
389
+ "hydraulic": -0.12489540487206459
390
+ },
391
+ "by_environment": {
392
+ "gas:batch10": -0.3102340868093479,
393
+ "gas:batch6": -0.21311841464661282,
394
+ "gas:batch7": -0.27387612272727374,
395
+ "hydraulic:context0=3": -0.12489540487206459
396
+ },
397
+ "tasks": 43,
398
+ "descriptive_paired_95_interval": [
399
+ -0.21756474584070623,
400
+ -0.16900690975933869
401
+ ],
402
+ "paired_fit_rows": 129,
403
+ "environment_resampling_only": true
404
+ },
405
+ {
406
+ "label": "target_all_minus_64_query_joint",
407
+ "field": "utility",
408
+ "overall": 0.038157311614910504,
409
+ "by_dataset": {
410
+ "gas": 0.05676291617893625,
411
+ "hydraulic": 0.019551707050884753
412
+ },
413
+ "by_environment": {
414
+ "gas:batch10": 0.11750666905912108,
415
+ "gas:batch6": 0.10113160947997442,
416
+ "gas:batch7": 0.07350602580668636,
417
+ "gas:batch8": -0.008274740633231198,
418
+ "gas:batch9": -5.498281786941268e-05,
419
+ "hydraulic:context0=20": -0.14266521423384168,
420
+ "hydraulic:context0=3": 0.11050896057347669,
421
+ "hydraulic:context1=73": -0.0011481481481481943,
422
+ "hydraulic:context1=80": 0.0032649572649571614,
423
+ "hydraulic:context1=90": 0.1277979797979798
424
+ },
425
+ "tasks": 57,
426
+ "descriptive_paired_95_interval": [
427
+ -0.01084189535725685,
428
+ 0.0850130341338505
429
+ ],
430
+ "paired_fit_rows": 171,
431
+ "environment_resampling_only": true
432
+ },
433
+ {
434
+ "label": "target_all_minus_64_query_joint",
435
+ "field": "noquery_utility",
436
+ "overall": 0.04251419582047826,
437
+ "by_dataset": {
438
+ "gas": 0.04728120243266674,
439
+ "hydraulic": 0.03774718920828978
440
+ },
441
+ "by_environment": {
442
+ "gas:batch10": 0.11361625735686477,
443
+ "gas:batch6": 0.056744070821711144,
444
+ "gas:batch7": 0.05963542757450136,
445
+ "gas:batch8": 0.0064102564102564005,
446
+ "gas:batch9": 0.0,
447
+ "hydraulic:context0=20": -0.1305010893246187,
448
+ "hydraulic:context0=3": 0.07168458781362004,
449
+ "hydraulic:context1=73": -1.4802973661668753e-16,
450
+ "hydraulic:context1=80": 0.005128205128205295,
451
+ "hydraulic:context1=90": 0.24242424242424243
452
+ },
453
+ "tasks": 57,
454
+ "descriptive_paired_95_interval": [
455
+ -0.009599849722433367,
456
+ 0.10225396410031547
457
+ ],
458
+ "paired_fit_rows": 171,
459
+ "environment_resampling_only": true
460
+ },
461
+ {
462
+ "label": "target_all_minus_64_query_joint",
463
+ "field": "realized_query_increment",
464
+ "overall": -0.004356884205567757,
465
+ "by_dataset": {
466
+ "gas": 0.00948171374626951,
467
+ "hydraulic": -0.018195482157405025
468
+ },
469
+ "by_environment": {
470
+ "gas:batch10": 0.0038904117022563103,
471
+ "gas:batch6": 0.04438753865826327,
472
+ "gas:batch7": 0.013870598232184988,
473
+ "gas:batch8": -0.014684997043487599,
474
+ "gas:batch9": -5.498281786941268e-05,
475
+ "hydraulic:context0=20": -0.012164124909222947,
476
+ "hydraulic:context0=3": 0.038824372759856625,
477
+ "hydraulic:context1=73": -0.0011481481481480464,
478
+ "hydraulic:context1=80": -0.0018632478632481335,
479
+ "hydraulic:context1=90": -0.11462626262626262
480
+ },
481
+ "tasks": 57,
482
+ "descriptive_paired_95_interval": [
483
+ -0.030791613659059625,
484
+ 0.01679233011427237
485
+ ],
486
+ "paired_fit_rows": 171,
487
+ "environment_resampling_only": true
488
+ },
489
+ {
490
+ "label": "target_all_minus_64_query_joint",
491
+ "field": "joint_conditional_log_loss",
492
+ "overall": -0.19687071580060886,
493
+ "by_dataset": {
494
+ "gas": -0.23411345652960858,
495
+ "hydraulic": -0.15962797507160914
496
+ },
497
+ "by_environment": {
498
+ "gas:batch10": -0.44018328456321953,
499
+ "gas:batch6": -0.2833204085585301,
500
+ "gas:batch7": -0.4036770926283303,
501
+ "gas:batch8": -0.02071081915566557,
502
+ "gas:batch9": -0.022675677742297333,
503
+ "hydraulic:context0=20": -0.1815063570268768,
504
+ "hydraulic:context0=3": -0.1292211545357918,
505
+ "hydraulic:context1=73": -0.08545253886246955,
506
+ "hydraulic:context1=80": -0.13971554729893834,
507
+ "hydraulic:context1=90": -0.2622442776339691
508
+ },
509
+ "tasks": 57,
510
+ "descriptive_paired_95_interval": [
511
+ -0.2781864450355927,
512
+ -0.1120529765173951
513
+ ],
514
+ "paired_fit_rows": 171,
515
+ "environment_resampling_only": true
516
+ },
517
+ {
518
+ "label": "target_256_minus_64_posterior_clusters",
519
+ "field": "utility",
520
+ "overall": 0.09109810434503701,
521
+ "by_dataset": {
522
+ "gas": 0.07168724811659732,
523
+ "hydraulic": 0.11050896057347669
524
+ },
525
+ "by_environment": {
526
+ "gas:batch10": 0.07042986008964809,
527
+ "gas:batch6": 0.09297076329625689,
528
+ "gas:batch7": 0.051661120963886996,
529
+ "hydraulic:context0=3": 0.11050896057347669
530
+ },
531
+ "tasks": 43,
532
+ "descriptive_paired_95_interval": [
533
+ 0.08108504076868184,
534
+ 0.10173986193486678
535
+ ],
536
+ "paired_fit_rows": 129,
537
+ "environment_resampling_only": true
538
+ },
539
+ {
540
+ "label": "target_256_minus_64_posterior_clusters",
541
+ "field": "noquery_utility",
542
+ "overall": 0.06468211447240929,
543
+ "by_dataset": {
544
+ "gas": 0.05767964113119853,
545
+ "hydraulic": 0.07168458781362004
546
+ },
547
+ "by_environment": {
548
+ "gas:batch10": 0.08343377230673357,
549
+ "gas:batch6": 0.05454737389719194,
550
+ "gas:batch7": 0.03505777718967007,
551
+ "hydraulic:context0=3": 0.07168458781362004
552
+ },
553
+ "tasks": 43,
554
+ "descriptive_paired_95_interval": [
555
+ 0.053371182501645055,
556
+ 0.07755918006017681
557
+ ],
558
+ "paired_fit_rows": 129,
559
+ "environment_resampling_only": true
560
+ },
561
+ {
562
+ "label": "target_256_minus_64_posterior_clusters",
563
+ "field": "realized_query_increment",
564
+ "overall": 0.026415989872627714,
565
+ "by_dataset": {
566
+ "gas": 0.014007606985398803,
567
+ "hydraulic": 0.038824372759856625
568
+ },
569
+ "by_environment": {
570
+ "gas:batch10": -0.01300391221708548,
571
+ "gas:batch6": 0.038423389399064964,
572
+ "gas:batch7": 0.016603343774216926,
573
+ "hydraulic:context0=3": 0.038824372759856625
574
+ },
575
+ "tasks": 43,
576
+ "descriptive_paired_95_interval": [
577
+ 0.012910230271385574,
578
+ 0.0386238810794608
579
+ ],
580
+ "paired_fit_rows": 129,
581
+ "environment_resampling_only": true
582
+ },
583
+ {
584
+ "label": "target_256_minus_64_posterior_clusters",
585
+ "field": "joint_conditional_log_loss",
586
+ "overall": -0.1776871678007718,
587
+ "by_dataset": {
588
+ "gas": -0.23048284772806546,
589
+ "hydraulic": -0.1248914878734781
590
+ },
591
+ "by_environment": {
592
+ "gas:batch10": -0.2824511247837857,
593
+ "gas:batch6": -0.1741531575627679,
594
+ "gas:batch7": -0.23484426083764282,
595
+ "hydraulic:context0=3": -0.1248914878734781
596
+ },
597
+ "tasks": 43,
598
+ "descriptive_paired_95_interval": [
599
+ -0.20367130632863187,
600
+ -0.149522322718123
601
+ ],
602
+ "paired_fit_rows": 129,
603
+ "environment_resampling_only": true
604
+ },
605
+ {
606
+ "label": "target_all_minus_64_posterior_clusters",
607
+ "field": "utility",
608
+ "overall": 0.04173847852269007,
609
+ "by_dataset": {
610
+ "gas": 0.05837645804570337,
611
+ "hydraulic": 0.025100498999676757
612
+ },
613
+ "by_environment": {
614
+ "gas:batch10": 0.10966002212485516,
615
+ "gas:batch6": 0.1075914838989129,
616
+ "gas:batch7": 0.08017560737094916,
617
+ "gas:batch8": -0.0054898403483309545,
618
+ "gas:batch9": -5.498281786941268e-05,
619
+ "hydraulic:context0=20": -0.14266521423384168,
620
+ "hydraulic:context0=3": 0.11050896057347669,
621
+ "hydraulic:context1=73": 0.01011640211640219,
622
+ "hydraulic:context1=80": -0.009427350427350434,
623
+ "hydraulic:context1=90": 0.156969696969697
624
+ },
625
+ "tasks": 57,
626
+ "descriptive_paired_95_interval": [
627
+ -0.009828601087338618,
628
+ 0.09135077312892757
629
+ ],
630
+ "paired_fit_rows": 171,
631
+ "environment_resampling_only": true
632
+ },
633
+ {
634
+ "label": "target_all_minus_64_posterior_clusters",
635
+ "field": "noquery_utility",
636
+ "overall": 0.03433698659897971,
637
+ "by_dataset": {
638
+ "gas": 0.05032072338360901,
639
+ "hydraulic": 0.018353249814350416
640
+ },
641
+ "by_environment": {
642
+ "gas:batch10": 0.11928082949136254,
643
+ "gas:batch6": 0.061912046117019635,
644
+ "gas:batch7": 0.07041074130966289,
645
+ "gas:batch8": -1.2335811384723961e-17,
646
+ "gas:batch9": 0.0,
647
+ "hydraulic:context0=20": -0.1305010893246187,
648
+ "hydraulic:context0=3": 0.07168458781362004,
649
+ "hydraulic:context1=73": -7.401486830834377e-17,
650
+ "hydraulic:context1=80": 0.005128205128205184,
651
+ "hydraulic:context1=90": 0.14545454545454564
652
+ },
653
+ "tasks": 57,
654
+ "descriptive_paired_95_interval": [
655
+ -0.011377864439774938,
656
+ 0.07841280648218549
657
+ ],
658
+ "paired_fit_rows": 171,
659
+ "environment_resampling_only": true
660
+ },
661
+ {
662
+ "label": "target_all_minus_64_posterior_clusters",
663
+ "field": "realized_query_increment",
664
+ "overall": 0.007401491923710351,
665
+ "by_dataset": {
666
+ "gas": 0.008055734662094364,
667
+ "hydraulic": 0.006747249185326339
668
+ },
669
+ "by_environment": {
670
+ "gas:batch10": -0.009620807366507366,
671
+ "gas:batch6": 0.045679437781893276,
672
+ "gas:batch7": 0.009764866061286272,
673
+ "gas:batch8": -0.005489840348330943,
674
+ "gas:batch9": -5.498281786941268e-05,
675
+ "hydraulic:context0=20": -0.012164124909222947,
676
+ "hydraulic:context0=3": 0.038824372759856625,
677
+ "hydraulic:context1=73": 0.010116402116402265,
678
+ "hydraulic:context1=80": -0.014555555555555618,
679
+ "hydraulic:context1=90": 0.011515151515151367
680
+ },
681
+ "tasks": 57,
682
+ "descriptive_paired_95_interval": [
683
+ -0.0038314389064033304,
684
+ 0.020036524552761063
685
+ ],
686
+ "paired_fit_rows": 171,
687
+ "environment_resampling_only": true
688
+ },
689
+ {
690
+ "label": "target_all_minus_64_posterior_clusters",
691
+ "field": "joint_conditional_log_loss",
692
+ "overall": -0.1681942885407447,
693
+ "by_dataset": {
694
+ "gas": -0.2319239590935684,
695
+ "hydraulic": -0.10446461798792102
696
+ },
697
+ "by_environment": {
698
+ "gas:batch10": -0.4418217882384215,
699
+ "gas:batch6": -0.29143101306951796,
700
+ "gas:batch7": -0.3907604992191855,
701
+ "gas:batch8": -0.012722694648025131,
702
+ "gas:batch9": -0.02288380029269186,
703
+ "hydraulic:context0=20": -0.1815063570268768,
704
+ "hydraulic:context0=3": -0.12923307115226043,
705
+ "hydraulic:context1=73": -0.04543127935882071,
706
+ "hydraulic:context1=80": 0.0002010261398282346,
707
+ "hydraulic:context1=90": -0.1663534085414754
708
+ },
709
+ "tasks": 57,
710
+ "descriptive_paired_95_interval": [
711
+ -0.25084335229608745,
712
+ -0.08113264067417841
713
+ ],
714
+ "paired_fit_rows": 171,
715
+ "environment_resampling_only": true
716
+ },
717
+ {
718
+ "label": "target_256_minus_64_identity",
719
+ "field": "utility",
720
+ "overall": 0.09048828332602109,
721
+ "by_dataset": {
722
+ "gas": 0.07046760607856549,
723
+ "hydraulic": 0.11050896057347669
724
+ },
725
+ "by_environment": {
726
+ "gas:batch10": 0.07419248136066033,
727
+ "gas:batch6": 0.0904448025664073,
728
+ "gas:batch7": 0.04676553430862883,
729
+ "hydraulic:context0=3": 0.11050896057347669
730
+ },
731
+ "tasks": 43,
732
+ "descriptive_paired_95_interval": [
733
+ 0.07863724744105276,
734
+ 0.100476881569942
735
+ ],
736
+ "paired_fit_rows": 129,
737
+ "environment_resampling_only": true
738
+ },
739
+ {
740
+ "label": "target_256_minus_64_identity",
741
+ "field": "noquery_utility",
742
+ "overall": 0.0637826754644584,
743
+ "by_dataset": {
744
+ "gas": 0.055880763115296744,
745
+ "hydraulic": 0.07168458781362005
746
+ },
747
+ "by_environment": {
748
+ "gas:batch10": 0.08491540788816963,
749
+ "gas:batch6": 0.0473874535281433,
750
+ "gas:batch7": 0.0353394279295773,
751
+ "hydraulic:context0=3": 0.07168458781362005
752
+ },
753
+ "tasks": 43,
754
+ "descriptive_paired_95_interval": [
755
+ 0.053512007871598676,
756
+ 0.07829999785089484
757
+ ],
758
+ "paired_fit_rows": 129,
759
+ "environment_resampling_only": true
760
+ },
761
+ {
762
+ "label": "target_256_minus_64_identity",
763
+ "field": "realized_query_increment",
764
+ "overall": 0.026705607861562685,
765
+ "by_dataset": {
766
+ "gas": 0.01458684296326875,
767
+ "hydraulic": 0.03882437275985662
768
+ },
769
+ "by_environment": {
770
+ "gas:batch10": -0.010722926527509303,
771
+ "gas:batch6": 0.04305734903826401,
772
+ "gas:batch7": 0.011426106379051539,
773
+ "hydraulic:context0=3": 0.03882437275985662
774
+ },
775
+ "tasks": 43,
776
+ "descriptive_paired_95_interval": [
777
+ 0.014050723116173659,
778
+ 0.04094086089906031
779
+ ],
780
+ "paired_fit_rows": 129,
781
+ "environment_resampling_only": true
782
+ },
783
+ {
784
+ "label": "target_256_minus_64_identity",
785
+ "field": "joint_conditional_log_loss",
786
+ "overall": -0.19112487728113603,
787
+ "by_dataset": {
788
+ "gas": -0.2573433341697351,
789
+ "hydraulic": -0.12490642039253695
790
+ },
791
+ "by_environment": {
792
+ "gas:batch10": -0.30373009969425085,
793
+ "gas:batch6": -0.20457846206949787,
794
+ "gas:batch7": -0.26372144074545656,
795
+ "hydraulic:context0=3": -0.12490642039253695
796
+ },
797
+ "tasks": 43,
798
+ "descriptive_paired_95_interval": [
799
+ -0.21431826004339388,
800
+ -0.1647424412310174
801
+ ],
802
+ "paired_fit_rows": 129,
803
+ "environment_resampling_only": true
804
+ },
805
+ {
806
+ "label": "target_all_minus_64_identity",
807
+ "field": "utility",
808
+ "overall": 0.03613808594050513,
809
+ "by_dataset": {
810
+ "gas": 0.05583580164146228,
811
+ "hydraulic": 0.01644037023954799
812
+ },
813
+ "by_environment": {
814
+ "gas:batch10": 0.12010778243557424,
815
+ "gas:batch6": 0.10662597263758453,
816
+ "gas:batch7": 0.0705773270126037,
817
+ "gas:batch8": -0.018077091060581647,
818
+ "gas:batch9": -5.498281786941268e-05,
819
+ "hydraulic:context0=20": -0.14266521423384168,
820
+ "hydraulic:context0=3": 0.11050896057347669,
821
+ "hydraulic:context1=73": -0.002232804232804253,
822
+ "hydraulic:context1=80": -0.009388888888888891,
823
+ "hydraulic:context1=90": 0.12597979797979808
824
+ },
825
+ "tasks": 57,
826
+ "descriptive_paired_95_interval": [
827
+ -0.013789338141264412,
828
+ 0.08366247811554169
829
+ ],
830
+ "paired_fit_rows": 171,
831
+ "environment_resampling_only": true
832
+ },
833
+ {
834
+ "label": "target_all_minus_64_identity",
835
+ "field": "noquery_utility",
836
+ "overall": 0.03392151442389414,
837
+ "by_dataset": {
838
+ "gas": 0.04948977903343789,
839
+ "hydraulic": 0.01835324981435039
840
+ },
841
+ "by_environment": {
842
+ "gas:batch10": 0.11999143590250941,
843
+ "gas:batch6": 0.060364321921130296,
844
+ "gas:batch7": 0.06709313734354976,
845
+ "gas:batch8": 0.0,
846
+ "gas:batch9": 0.0,
847
+ "hydraulic:context0=20": -0.1305010893246187,
848
+ "hydraulic:context0=3": 0.07168458781362005,
849
+ "hydraulic:context1=73": 0.0,
850
+ "hydraulic:context1=80": 0.00512820512820511,
851
+ "hydraulic:context1=90": 0.1454545454545455
852
+ },
853
+ "tasks": 57,
854
+ "descriptive_paired_95_interval": [
855
+ -0.011763300302199731,
856
+ 0.07763983304671877
857
+ ],
858
+ "paired_fit_rows": 171,
859
+ "environment_resampling_only": true
860
+ },
861
+ {
862
+ "label": "target_all_minus_64_identity",
863
+ "field": "realized_query_increment",
864
+ "overall": 0.002216571516610994,
865
+ "by_dataset": {
866
+ "gas": 0.006346022608024389,
867
+ "hydraulic": -0.0019128795748024011
868
+ },
869
+ "by_environment": {
870
+ "gas:batch10": 0.00011634653306482027,
871
+ "gas:batch6": 0.04626165071645424,
872
+ "gas:batch7": 0.003484189669053949,
873
+ "gas:batch8": -0.018077091060581647,
874
+ "gas:batch9": -5.498281786941268e-05,
875
+ "hydraulic:context0=20": -0.012164124909222947,
876
+ "hydraulic:context0=3": 0.03882437275985662,
877
+ "hydraulic:context1=73": -0.002232804232804253,
878
+ "hydraulic:context1=80": -0.014517094017094001,
879
+ "hydraulic:context1=90": -0.019474747474747423
880
+ },
881
+ "tasks": 57,
882
+ "descriptive_paired_95_interval": [
883
+ -0.009652936741599463,
884
+ 0.016161801817197174
885
+ ],
886
+ "paired_fit_rows": 171,
887
+ "environment_resampling_only": true
888
+ },
889
+ {
890
+ "label": "target_all_minus_64_identity",
891
+ "field": "joint_conditional_log_loss",
892
+ "overall": -0.18801059075396234,
893
+ "by_dataset": {
894
+ "gas": -0.22465903697780676,
895
+ "hydraulic": -0.15136214453011793
896
+ },
897
+ "by_environment": {
898
+ "gas:batch10": -0.42647144343500426,
899
+ "gas:batch6": -0.27447906953799683,
900
+ "gas:batch7": -0.38825936559980373,
901
+ "gas:batch8": -0.011590157952489915,
902
+ "gas:batch9": -0.022495148363738955,
903
+ "hydraulic:context0=20": -0.1815063570268768,
904
+ "hydraulic:context0=3": -0.1292322133249814,
905
+ "hydraulic:context1=73": -0.09410779468813202,
906
+ "hydraulic:context1=80": -0.10823786750704219,
907
+ "hydraulic:context1=90": -0.24372649010355726
908
+ },
909
+ "tasks": 57,
910
+ "descriptive_paired_95_interval": [
911
+ -0.26639647181970033,
912
+ -0.10573772708136289
913
+ ],
914
+ "paired_fit_rows": 171,
915
+ "environment_resampling_only": true
916
+ },
917
+ {
918
+ "label": "source_fixed_policy_target_minus_sourceheld_64",
919
+ "field": "utility",
920
+ "overall": -0.3658345628075126,
921
+ "by_dataset": {
922
+ "gas": -0.5610563133386619,
923
+ "hydraulic": -0.1706128122763632
924
+ },
925
+ "by_environment": {
926
+ "gas:batch10": -0.9396158433078271,
927
+ "gas:batch6": -0.5596432280059105,
928
+ "gas:batch7": -0.6502933684535855,
929
+ "gas:batch8": -0.19518422367138946,
930
+ "gas:batch9": -0.46054490325459685,
931
+ "hydraulic:context0=20": -0.5833520872736558,
932
+ "hydraulic:context0=3": -0.8878212699287965,
933
+ "hydraulic:context1=73": 0.2936963399334534,
934
+ "hydraulic:context1=80": 0.2983467265838401,
935
+ "hydraulic:context1=90": 0.02606622930334268
936
+ },
937
+ "tasks": 57,
938
+ "descriptive_paired_95_interval": [
939
+ -0.6094731109065757,
940
+ -0.14076501702032948
941
+ ],
942
+ "paired_fit_rows": 171,
943
+ "environment_resampling_only": true
944
+ },
945
+ {
946
+ "label": "source_fixed_policy_target_minus_sourceheld_64",
947
+ "field": "noquery_utility",
948
+ "overall": -0.2676674191045439,
949
+ "by_dataset": {
950
+ "gas": -0.5513721366349306,
951
+ "hydraulic": 0.016037298425842818
952
+ },
953
+ "by_environment": {
954
+ "gas:batch10": -0.9166768695714907,
955
+ "gas:batch6": -0.5431060684074404,
956
+ "gas:batch7": -0.630188728970858,
957
+ "gas:batch8": -0.19767658909014943,
958
+ "gas:batch9": -0.4692124271347146,
959
+ "hydraulic:context0=20": -0.7049630108453638,
960
+ "hydraulic:context0=3": -0.951136318878254,
961
+ "hydraulic:context1=73": 0.705187367043037,
962
+ "hydraulic:context1=80": 0.7126354744911446,
963
+ "hydraulic:context1=90": 0.31846298031865033
964
+ },
965
+ "tasks": 57,
966
+ "descriptive_paired_95_interval": [
967
+ -0.6011976930988733,
968
+ 0.041075856067274444
969
+ ],
970
+ "paired_fit_rows": 171,
971
+ "environment_resampling_only": true
972
+ },
973
+ {
974
+ "label": "source_fixed_policy_target_minus_sourceheld_64",
975
+ "field": "realized_query_increment",
976
+ "overall": -0.09816714370296867,
977
+ "by_dataset": {
978
+ "gas": -0.009684176703731236,
979
+ "hydraulic": -0.1866501107022061
980
+ },
981
+ "by_environment": {
982
+ "gas:batch10": -0.02293897373633634,
983
+ "gas:batch6": -0.0165371595984701,
984
+ "gas:batch7": -0.020104639482727473,
985
+ "gas:batch8": 0.0024923654187599905,
986
+ "gas:batch9": 0.00866752388011774,
987
+ "hydraulic:context0=20": 0.12161092357170783,
988
+ "hydraulic:context0=3": 0.06331504894945743,
989
+ "hydraulic:context1=73": -0.4114910271095837,
990
+ "hydraulic:context1=80": -0.4142887479073045,
991
+ "hydraulic:context1=90": -0.2923967510153076
992
+ },
993
+ "tasks": 57,
994
+ "descriptive_paired_95_interval": [
995
+ -0.197981945594568,
996
+ 0.003237513471611564
997
+ ],
998
+ "paired_fit_rows": 171,
999
+ "environment_resampling_only": true
1000
+ },
1001
+ {
1002
+ "label": "source_fixed_policy_target_minus_sourceheld_64",
1003
+ "field": "joint_conditional_log_loss",
1004
+ "overall": 0.5621445508593099,
1005
+ "by_dataset": {
1006
+ "gas": 0.5800953291658606,
1007
+ "hydraulic": 0.5441937725527592
1008
+ },
1009
+ "by_environment": {
1010
+ "gas:batch10": 0.9486170445415449,
1011
+ "gas:batch6": 0.4437875424241321,
1012
+ "gas:batch7": 0.7164416200807845,
1013
+ "gas:batch8": 0.426021559021634,
1014
+ "gas:batch9": 0.36560887976120765,
1015
+ "hydraulic:context0=20": 1.6799077884463471,
1016
+ "hydraulic:context0=3": 1.682189335498656,
1017
+ "hydraulic:context1=73": -0.4180981671810516,
1018
+ "hydraulic:context1=80": -0.19134701166286475,
1019
+ "hydraulic:context1=90": -0.03168308233729015
1020
+ },
1021
+ "tasks": 57,
1022
+ "descriptive_paired_95_interval": [
1023
+ 0.15401369963728354,
1024
+ 0.9865531222258157
1025
+ ],
1026
+ "paired_fit_rows": 171,
1027
+ "environment_resampling_only": true
1028
+ },
1029
+ {
1030
+ "label": "source_fixed_policy_target_minus_sourceheld_64",
1031
+ "field": "common_current_joint_conditional_log_loss",
1032
+ "overall": 0.5594148121355085,
1033
+ "by_dataset": {
1034
+ "gas": 0.20223821826777272,
1035
+ "hydraulic": 0.9165914060032442
1036
+ },
1037
+ "by_environment": {
1038
+ "gas:batch10": 0.6142603038782647,
1039
+ "gas:batch6": 0.02758507900012181,
1040
+ "gas:batch7": 0.5621359751087247,
1041
+ "gas:batch8": 0.2504521509982081,
1042
+ "gas:batch9": -0.4432424176464558,
1043
+ "hydraulic:context0=20": 2.0467262627943366,
1044
+ "hydraulic:context0=3": 2.4213909252713104,
1045
+ "hydraulic:context1=73": 0.017684862182114403,
1046
+ "hydraulic:context1=80": 0.12330671842806,
1047
+ "hydraulic:context1=90": -0.02615173865960056
1048
+ },
1049
+ "tasks": 57,
1050
+ "descriptive_paired_95_interval": [
1051
+ 0.08425091300955856,
1052
+ 1.0753570244945965
1053
+ ],
1054
+ "paired_fit_rows": 171,
1055
+ "environment_resampling_only": true
1056
+ },
1057
+ {
1058
+ "label": "source_fixed_policy_target_minus_sourceheld_256",
1059
+ "field": "utility",
1060
+ "overall": -0.45219716166405916,
1061
+ "by_dataset": {
1062
+ "gas": -0.6613710559765582,
1063
+ "hydraulic": -0.24302326735156013
1064
+ },
1065
+ "by_environment": {
1066
+ "gas:batch10": -0.9162104276816436,
1067
+ "gas:batch6": -0.45642588117381144,
1068
+ "gas:batch7": -0.7440051374832067,
1069
+ "gas:batch8": -0.4961445700097846,
1070
+ "gas:batch9": -0.6940692635343448,
1071
+ "hydraulic:context0=20": -0.3936751897340132,
1072
+ "hydraulic:context0=3": -0.9409959431249754,
1073
+ "hydraulic:context1=73": 0.05824164075710466,
1074
+ "hydraulic:context1=80": 0.044280509295973235,
1075
+ "hydraulic:context1=90": 0.017032646048109983
1076
+ },
1077
+ "tasks": 57,
1078
+ "descriptive_paired_95_interval": [
1079
+ -0.6480843443520771,
1080
+ -0.286414985093672
1081
+ ],
1082
+ "paired_fit_rows": 171,
1083
+ "environment_resampling_only": true
1084
+ },
1085
+ {
1086
+ "label": "source_fixed_policy_target_minus_sourceheld_256",
1087
+ "field": "noquery_utility",
1088
+ "overall": -0.49481529547920394,
1089
+ "by_dataset": {
1090
+ "gas": -0.6612623846504923,
1091
+ "hydraulic": -0.32836820630791563
1092
+ },
1093
+ "by_environment": {
1094
+ "gas:batch10": -0.9064315822975246,
1095
+ "gas:batch6": -0.4365394381842272,
1096
+ "gas:batch7": -0.7362261658674275,
1097
+ "gas:batch8": -0.551893963728601,
1098
+ "gas:batch9": -0.6752207731746808,
1099
+ "hydraulic:context0=20": -0.6481936364289306,
1100
+ "hydraulic:context0=3": -1.2564443919282626,
1101
+ "hydraulic:context1=73": 0.11600392734413338,
1102
+ "hydraulic:context1=80": 0.09659000793021395,
1103
+ "hydraulic:context1=90": 0.05020306154326767
1104
+ },
1105
+ "tasks": 57,
1106
+ "descriptive_paired_95_interval": [
1107
+ -0.755367243841711,
1108
+ -0.270041361311064
1109
+ ],
1110
+ "paired_fit_rows": 171,
1111
+ "environment_resampling_only": true
1112
+ },
1113
+ {
1114
+ "label": "source_fixed_policy_target_minus_sourceheld_256",
1115
+ "field": "realized_query_increment",
1116
+ "overall": 0.042618133815144746,
1117
+ "by_dataset": {
1118
+ "gas": -0.00010867132606601146,
1119
+ "hydraulic": 0.0853449389563555
1120
+ },
1121
+ "by_environment": {
1122
+ "gas:batch10": -0.00977884538411901,
1123
+ "gas:batch6": -0.019886442989584283,
1124
+ "gas:batch7": -0.007778971615779189,
1125
+ "gas:batch8": 0.055749393718816466,
1126
+ "gas:batch9": -0.01884849035966404,
1127
+ "hydraulic:context0=20": 0.25451844669491724,
1128
+ "hydraulic:context0=3": 0.3154484488032874,
1129
+ "hydraulic:context1=73": -0.05776228658702872,
1130
+ "hydraulic:context1=80": -0.052309498634240725,
1131
+ "hydraulic:context1=90": -0.03317041549515768
1132
+ },
1133
+ "tasks": 57,
1134
+ "descriptive_paired_95_interval": [
1135
+ -0.027765450120834784,
1136
+ 0.11627422347851524
1137
+ ],
1138
+ "paired_fit_rows": 171,
1139
+ "environment_resampling_only": true
1140
+ },
1141
+ {
1142
+ "label": "source_fixed_policy_target_minus_sourceheld_256",
1143
+ "field": "joint_conditional_log_loss",
1144
+ "overall": 0.6138085845900283,
1145
+ "by_dataset": {
1146
+ "gas": 0.6632822137991161,
1147
+ "hydraulic": 0.5643349553809406
1148
+ },
1149
+ "by_environment": {
1150
+ "gas:batch10": 0.8991049325233872,
1151
+ "gas:batch6": 0.37054835953278614,
1152
+ "gas:batch7": 0.8245430070724563,
1153
+ "gas:batch8": 0.6995219189865814,
1154
+ "gas:batch9": 0.5226928508803698,
1155
+ "hydraulic:context0=20": 1.7382918974288823,
1156
+ "hydraulic:context0=3": 1.7819212294649285,
1157
+ "hydraulic:context1=73": -0.38166563484429133,
1158
+ "hydraulic:context1=80": -0.21327311004018704,
1159
+ "hydraulic:context1=90": -0.10359960510462929
1160
+ },
1161
+ "tasks": 57,
1162
+ "descriptive_paired_95_interval": [
1163
+ 0.19329377403326248,
1164
+ 1.053042905163758
1165
+ ],
1166
+ "paired_fit_rows": 171,
1167
+ "environment_resampling_only": true
1168
+ },
1169
+ {
1170
+ "label": "source_fixed_policy_target_minus_sourceheld_256",
1171
+ "field": "common_current_joint_conditional_log_loss",
1172
+ "overall": 0.6519535658720754,
1173
+ "by_dataset": {
1174
+ "gas": 0.38500890988983894,
1175
+ "hydraulic": 0.9188982218543119
1176
+ },
1177
+ "by_environment": {
1178
+ "gas:batch10": 0.7156680549909323,
1179
+ "gas:batch6": 0.16467203783818304,
1180
+ "gas:batch7": 0.7492018890832239,
1181
+ "gas:batch8": 0.36610723064655865,
1182
+ "gas:batch9": -0.07060466310970315,
1183
+ "hydraulic:context0=20": 2.128697318307584,
1184
+ "hydraulic:context0=3": 2.451537531866698,
1185
+ "hydraulic:context1=73": 0.011443334798441415,
1186
+ "hydraulic:context1=80": 0.07249372345870368,
1187
+ "hydraulic:context1=90": -0.06968079915986725
1188
+ },
1189
+ "tasks": 57,
1190
+ "descriptive_paired_95_interval": [
1191
+ 0.16244540247836492,
1192
+ 1.1709003990409212
1193
+ ],
1194
+ "paired_fit_rows": 171,
1195
+ "environment_resampling_only": true
1196
+ },
1197
+ {
1198
+ "label": "source_fixed_policy_target_minus_sourceheld_all",
1199
+ "field": "utility",
1200
+ "overall": -0.4396127556794738,
1201
+ "by_dataset": {
1202
+ "gas": -0.7350152367869054,
1203
+ "hydraulic": -0.14421027457204222
1204
+ },
1205
+ "by_environment": {
1206
+ "gas:batch10": -0.9367731818344003,
1207
+ "gas:batch6": -0.460957579508446,
1208
+ "gas:batch7": -0.795378052512238,
1209
+ "gas:batch8": -0.5596553210223444,
1210
+ "gas:batch9": -0.9223120490570984,
1211
+ "hydraulic:context0=20": -0.21553340994517464,
1212
+ "hydraulic:context0=3": -0.6091156012446335,
1213
+ "hydraulic:context1=73": 0.04251243658975612,
1214
+ "hydraulic:context1=80": 0.03424015331747299,
1215
+ "hydraulic:context1=90": 0.026845048422368012
1216
+ },
1217
+ "tasks": 57,
1218
+ "descriptive_paired_95_interval": [
1219
+ -0.5862730773581387,
1220
+ -0.30774058981316016
1221
+ ],
1222
+ "paired_fit_rows": 171,
1223
+ "environment_resampling_only": true
1224
+ },
1225
+ {
1226
+ "label": "source_fixed_policy_target_minus_sourceheld_all",
1227
+ "field": "noquery_utility",
1228
+ "overall": -0.5717488762836339,
1229
+ "by_dataset": {
1230
+ "gas": -0.7452046483002596,
1231
+ "hydraulic": -0.39829310426700826
1232
+ },
1233
+ "by_environment": {
1234
+ "gas:batch10": -0.8891339260517567,
1235
+ "gas:batch6": -0.4653948925826942,
1236
+ "gas:batch7": -0.8336366769520721,
1237
+ "gas:batch8": -0.5831796742059052,
1238
+ "gas:batch9": -0.9546780717088698,
1239
+ "hydraulic:context0=20": -0.7205774617539324,
1240
+ "hydraulic:context0=3": -1.2731812989877502,
1241
+ "hydraulic:context1=73": 0.01183112420225818,
1242
+ "hydraulic:context1=80": 0.015860428231562196,
1243
+ "hydraulic:context1=90": -0.025398313027178987
1244
+ },
1245
+ "tasks": 57,
1246
+ "descriptive_paired_95_interval": [
1247
+ -0.8251105760935765,
1248
+ -0.3498471014539569
1249
+ ],
1250
+ "paired_fit_rows": 171,
1251
+ "environment_resampling_only": true
1252
+ },
1253
+ {
1254
+ "label": "source_fixed_policy_target_minus_sourceheld_all",
1255
+ "field": "realized_query_increment",
1256
+ "overall": 0.13213612060416013,
1257
+ "by_dataset": {
1258
+ "gas": 0.010189411513354157,
1259
+ "hydraulic": 0.2540828296949661
1260
+ },
1261
+ "by_environment": {
1262
+ "gas:batch10": -0.0476392557826436,
1263
+ "gas:batch6": 0.004437313074248212,
1264
+ "gas:batch7": 0.038258624439834056,
1265
+ "gas:batch8": 0.02352435318356087,
1266
+ "gas:batch9": 0.03236602265177125,
1267
+ "hydraulic:context0=20": 0.5050440518087577,
1268
+ "hydraulic:context0=3": 0.664065697743117,
1269
+ "hydraulic:context1=73": 0.03068131238749794,
1270
+ "hydraulic:context1=80": 0.018379725085910792,
1271
+ "hydraulic:context1=90": 0.052243361449547
1272
+ },
1273
+ "tasks": 57,
1274
+ "descriptive_paired_95_interval": [
1275
+ 0.01925016215607177,
1276
+ 0.2549383828614076
1277
+ ],
1278
+ "paired_fit_rows": 171,
1279
+ "environment_resampling_only": true
1280
+ },
1281
+ {
1282
+ "label": "source_fixed_policy_target_minus_sourceheld_all",
1283
+ "field": "joint_conditional_log_loss",
1284
+ "overall": 0.685292644015933,
1285
+ "by_dataset": {
1286
+ "gas": 0.7914155648485723,
1287
+ "hydraulic": 0.5791697231832938
1288
+ },
1289
+ "by_environment": {
1290
+ "gas:batch10": 0.9767006328121373,
1291
+ "gas:batch6": 0.3787905554315623,
1292
+ "gas:batch7": 0.932105359957967,
1293
+ "gas:batch8": 0.9789245055964885,
1294
+ "gas:batch9": 0.6905567704447064,
1295
+ "hydraulic:context0=20": 1.7437744440617207,
1296
+ "hydraulic:context0=3": 1.8285083029868723,
1297
+ "hydraulic:context1=73": -0.39084329451424943,
1298
+ "hydraulic:context1=80": -0.19394298606297666,
1299
+ "hydraulic:context1=90": -0.09164785055489755
1300
+ },
1301
+ "tasks": 57,
1302
+ "descriptive_paired_95_interval": [
1303
+ 0.25785834074167685,
1304
+ 1.1380925405286935
1305
+ ],
1306
+ "paired_fit_rows": 171,
1307
+ "environment_resampling_only": true
1308
+ },
1309
+ {
1310
+ "label": "source_fixed_policy_target_minus_sourceheld_all",
1311
+ "field": "common_current_joint_conditional_log_loss",
1312
+ "overall": 0.7167234355580608,
1313
+ "by_dataset": {
1314
+ "gas": 0.531140981710618,
1315
+ "hydraulic": 0.9023058894055037
1316
+ },
1317
+ "by_environment": {
1318
+ "gas:batch10": 0.8650012114531468,
1319
+ "gas:batch6": 0.23007087447055435,
1320
+ "gas:batch7": 0.8912368902514587,
1321
+ "gas:batch8": 0.541883922758298,
1322
+ "gas:batch9": 0.12751200961963197,
1323
+ "hydraulic:context0=20": 2.0541632239252148,
1324
+ "hydraulic:context0=3": 2.3804864627490447,
1325
+ "hydraulic:context1=73": 0.008144612778454824,
1326
+ "hydraulic:context1=80": 0.05183122892122599,
1327
+ "hydraulic:context1=90": 0.016903918653577987
1328
+ },
1329
+ "tasks": 57,
1330
+ "descriptive_paired_95_interval": [
1331
+ 0.2463337707830208,
1332
+ 1.221176584780543
1333
+ ],
1334
+ "paired_fit_rows": 171,
1335
+ "environment_resampling_only": true
1336
+ }
1337
+ ]
1338
+ }
revision_r1/external/run.py ADDED
@@ -0,0 +1,515 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Post-review external failure analysis; original data, policies and gates stay frozen.
2
+
3
+ All scientific tensor operations run on CUDA. JSON/CSV, hashes, split membership,
4
+ and archive serialization are CPU bookkeeping. No neural model is introduced.
5
+ """
6
+ from __future__ import annotations
7
+
8
+ import argparse
9
+ import csv
10
+ from datetime import datetime, timezone
11
+ import gzip
12
+ import hashlib
13
+ import json
14
+ from pathlib import Path
15
+ import platform
16
+ import sys
17
+ import time
18
+ from collections import defaultdict
19
+
20
+ ROOT = Path(__file__).resolve().parents[3]
21
+ _bootstrap = argparse.ArgumentParser(add_help=False)
22
+ _bootstrap.add_argument("--release", type=Path, default=ROOT / "articles/make_1/release/huggingface-revised")
23
+ _boot_args, _ = _bootstrap.parse_known_args()
24
+ RELEASE = _boot_args.release.resolve()
25
+ sys.path.insert(0, str(RELEASE))
26
+ import torch
27
+ from unionmind.decision_sufficiency.core import (
28
+ CONTEXTS, calibration, count_samples, encode, estimate, plan, push, realized,
29
+ )
30
+ from unionmind.external_sufficiency.runtime import replay, select_samples
31
+
32
+ INPUT = ROOT / "artifacts/external_decision_sufficiency_01"
33
+ DEFAULT_OUTPUT = ROOT / "articles/make_1/metadata/review_r1/external"
34
+ SEEDS = [20111, 20112, 20113]
35
+ METHODS = ["query_joint", "posterior_clusters", "identity"]
36
+ MODELS = ["support_mixture", "joint_jeffreys", "naive_bayes"]
37
+
38
+
39
+ def sha(path):
40
+ return hashlib.sha256(Path(path).read_bytes()).hexdigest()
41
+
42
+
43
+ def write(path, obj):
44
+ path.parent.mkdir(parents=True, exist_ok=True)
45
+ path.write_text(json.dumps(obj, indent=2, allow_nan=False) + "\n")
46
+
47
+
48
+ def utc():
49
+ return datetime.now(timezone.utc).isoformat()
50
+
51
+
52
+ def specification(out, device):
53
+ manifest = json.loads((INPUT / "prepared/manifest.json").read_text())
54
+ spec = dict(
55
+ material_passport=dict(material_type="post-review exploratory experiment specification",
56
+ provenance="Previously opened external sensor archives and original eligible task manifest",
57
+ license="Hydraulic UCI CC BY 4.0; gas legacy research-only restriction retained; local research use only",
58
+ created_utc=utc(), scientific_status="exploratory; does not replace original confirmation"),
59
+ purpose="Distinguish small within-fitted-law encoding headroom, finite labeled support, and observed source-target transfer failure without claiming an identifiable population decomposition",
60
+ device=device, dtype="float64", backbone_executions=0, trainable_parameters=0,
61
+ task_indices=[t["index"] for t in manifest["tasks"]], seeds=SEEDS,
62
+ methods=METHODS, estimators=MODELS,
63
+ interventions=[
64
+ "Original target64 exact seed selection, nested target256 when available, and all original target adaptation records",
65
+ "Source group-disjoint train/held split: SHA256('review_r1_source_held:' + original_group) mod 5 == 0 held; source64/256/all train only",
66
+ "Each fixed policy replayed on original target evaluation and, for source-trained policies, source held records",
67
+ "All-target-adaptation unsmoothed empirical reference law: encoders and exact policy fitted without target evaluation access, evaluated on original independent-group target evaluation; not a population oracle",
68
+ ],
69
+ source_preprocessing="Original source-only binary thresholds kept; source-held records contributed unlabeled preprocessing before this post-review split, never labeled model fitting",
70
+ metrics=["paired realized utility and no-query utility", "model-predicted versus held-realized query increment", "fitted encoding loss", "posterior/joint/observation conditional log losses", "same-policy source-held versus shifted-target scores", "finite-sample empirical conditional TV on overlapping current support", "target mass outside source-held current support"],
71
+ aggregation="Original weighting: seeds, then tasks within environment, then environments within dataset, then the two datasets equally. Sample-size contrasts restricted to identical tasks. Costs 0,.01,.05 and budgets 1,2 averaged; no-query average from budget0.",
72
+ inference="Descriptive 95% paired bootstrap by target environment within fixed dataset; collapse tasks/seeds first. Few clusters; not population transfer inference; no confirmatory gate rescues.",
73
+ limitations=["True external law is unknown; histogram reference and TV have finite-sample error", "Source-target contrast also includes differences in physical groups and sample composition; no causal attribution percentages", "Labels-all support varies by task; 256 contrasts require available support and use matched task subsets", "Source group split introduces a new held partition; original target adaptation/evaluation split unchanged", "Three seeds are repeated fitting samples, not independent environmental replications", "Empirical-reference policies are exploratory benchmarks and their NLL is not reported because zero fitted cells yield uninformative infinite log scores"],
74
+ input_hashes={str(p.relative_to(ROOT) if p.is_relative_to(ROOT) else p): sha(p) for p in [INPUT / "prepared/manifest.json", RELEASE / "protocols/external_sufficiency_protocol_01.json", RELEASE / "unionmind/decision_sufficiency/core.py", RELEASE / "unionmind/external_sufficiency/runtime.py"]},
75
+ script_sha256=sha(__file__), no_recipe_selection=True, original_results_and_gates_unchanged=True,
76
+ )
77
+ assert not (out / "specification.json").exists(), "Specification already exists; never overwrite it."
78
+ write(out / "specification.json", spec)
79
+ return spec
80
+
81
+
82
+ def normalized_counts(events):
83
+ c = count_samples(events[None], 6)
84
+ return c / c.sum()
85
+
86
+
87
+ def scores(model, truth):
88
+ result = calibration(model, truth)
89
+ cond = model / model.sum((2, 3), keepdim=True).clamp_min(1e-300)
90
+ givenh = model / model.sum(3, keepdim=True).clamp_min(1e-300)
91
+ result["joint_conditional_log_loss"] = -(truth * cond.clamp_min(1e-300).log()).sum((1, 2, 3))
92
+ result["observation_given_h_log_loss"] = -(truth * givenh.clamp_min(1e-300).log()).sum((1, 2, 3))
93
+ return result
94
+
95
+
96
+ def batched_replay(policy, code, events, budget, ci):
97
+ """Same paid-bit access convention as original replay; batch independent policies."""
98
+ b, n = len(code), len(events)
99
+ r = policy["shape"][1]
100
+ table = policy["choices"].reshape(3, b, r, 3, 9)
101
+ z = code[:, (events // 4) % 64].long()
102
+ row = torch.arange(b, device=events.device)[:, None].expand(b, n)
103
+ context = torch.zeros((b, n), device=events.device, dtype=torch.long)
104
+ action, queries, seen = context.clone(), context.clone(), context.clone()
105
+ live = torch.ones_like(context, dtype=torch.bool)
106
+ child = torch.zeros((9, 2, 2), device=events.device, dtype=torch.long)
107
+ for t, c in enumerate(CONTEXTS):
108
+ for q in (0, 1):
109
+ if c[q] == -1:
110
+ for o in (0, 1):
111
+ child[t, q, o] = CONTEXTS.index(tuple(o if i == q else v for i, v in enumerate(c)))
112
+ for left in range(budget, -1, -1):
113
+ choice = table[left, row, z, ci, context]
114
+ querying = live & (choice >= 3)
115
+ q = (choice - 3).clamp(0, 1)
116
+ selected = querying.nonzero()
117
+ revealed = torch.zeros_like(context)
118
+ rr, nn = selected[:, 0], selected[:, 1]
119
+ revealed[rr, nn] = (events[nn] >> (1 - q[rr, nn])) & 1
120
+ context = torch.where(querying, child[context, q, revealed], context)
121
+ action = torch.where(live & ~querying, choice, action)
122
+ queries += querying.long()
123
+ seen |= torch.where(querying, 1 << q, 0)
124
+ live = querying
125
+ assert not bool(live.any())
126
+ utility = torch.where(action == 2, 0., torch.where(action == events[None] // 256, 1., -2.)).double()
127
+ utility -= policy["costs"][ci] * queries.double()
128
+ return torch.stack((action, queries, seen), -1).to(torch.uint8), utility.mean(1)
129
+
130
+
131
+ def get_samples(archive, provenance, seed):
132
+ device = archive["source"].device
133
+ # Exact original selection consumes source permutation before adaptation.
134
+ g = torch.Generator(device=device).manual_seed(seed)
135
+ torch.randperm(len(archive["source"]), device=device, generator=g)
136
+ target_order = torch.randperm(len(archive["adaptation"]), device=device, generator=g)
137
+ held_mask = torch.tensor([int(hashlib.sha256(("review_r1_source_held:" + x["group"]).encode()).hexdigest(), 16) % 5 == 0 for x in provenance["source"]], device=device)
138
+ source_held = held_mask.nonzero().flatten()
139
+ source_train = (~held_mask).nonzero().flatten()
140
+ sg = torch.Generator(device=device).manual_seed(seed)
141
+ source_order = source_train[torch.randperm(len(source_train), device=device, generator=sg)]
142
+ assert len(source_order) >= 64 and len(source_held) > 0
143
+ values = []
144
+ indices = {}
145
+ for domain, order in [("target", target_order), ("source", source_order)]:
146
+ source_key = "adaptation" if domain == "target" else "source"
147
+ for size in [64, 256, "all"]:
148
+ if size != "all" and len(order) < size:
149
+ continue
150
+ ix = order if size == "all" else order[:size]
151
+ events = archive[source_key][ix]
152
+ values.append((domain, str(size), events))
153
+ indices[f"{domain}:{size}:{seed}"] = ix
154
+ original = select_samples(archive, "target_64", seed)
155
+ assert torch.equal(archive["adaptation"][target_order[:64]], original)
156
+ return values, indices, source_held, source_train
157
+
158
+
159
+ def evaluate(out, device):
160
+ spec = json.loads((out / "specification.json").read_text())
161
+ assert spec["script_sha256"] == sha(__file__)
162
+ for path, digest in spec["input_hashes"].items():
163
+ assert sha(ROOT / path) == digest
164
+ assert not (out / "metrics.csv.gz").exists(), "Never overwrite an analysis run."
165
+ manifest = json.loads((INPUT / "prepared/manifest.json").read_text())
166
+ torch.cuda.set_device(device)
167
+ torch.backends.cuda.matmul.allow_tf32 = False
168
+ torch.backends.cudnn.allow_tf32 = False
169
+ torch.cuda.reset_peak_memory_stats(device)
170
+ start = time.perf_counter()
171
+ rows, shifts = [], []
172
+ maximum_replay_error, maximum_original_error = 0., 0.
173
+ for task in manifest["tasks"]:
174
+ ti = task["index"]
175
+ archive = torch.load(INPUT / "prepared" / task["file"], map_location=device, weights_only=True)
176
+ prov = json.loads((INPUT / "prepared" / f"task_{ti:03}_provenance.json").read_text())
177
+ meta, counts, indices = [], [], {}
178
+ for seed in SEEDS:
179
+ samples, ix, source_held, source_train = get_samples(archive, prov, seed)
180
+ indices.update(ix)
181
+ for domain, size, events in samples:
182
+ c = count_samples(events[None], 6)[0]
183
+ for modelname in MODELS:
184
+ meta.append(dict(training=domain, size=size, sample_seed=seed, model=modelname, sample_count=len(events)))
185
+ counts.append(c)
186
+ # Nondeployable population-oracle claims are avoided: only independent
187
+ # adaptation labels define this empirical finite-sample reference.
188
+ meta.append(dict(training="target", size="all", sample_seed=-1, model="empirical_reference", sample_count=len(archive["adaptation"])))
189
+ counts.append(count_samples(archive["adaptation"][None], 6)[0])
190
+ allcounts = torch.stack(counts)
191
+ model = torch.empty_like(allcounts)
192
+ for name in MODELS:
193
+ ix = torch.tensor([i for i, m in enumerate(meta) if m["model"] == name], device=device)
194
+ model[ix] = estimate(allcounts[ix], name)
195
+ model[-1] = allcounts[-1] / allcounts[-1].sum()
196
+ full_policy = plan(model)
197
+ source_events = archive["source"][source_held]
198
+ laws = {"target_evaluation": normalized_counts(archive["evaluation"]), "source_held": normalized_counts(source_events)}
199
+ events_by_domain = {"target_evaluation": archive["evaluation"], "source_held": source_events}
200
+ saved = dict(metadata=meta, indices=indices, source_held_indices=source_held,
201
+ source_train_indices=source_train, source_held_events=source_events,
202
+ target_evaluation_events=archive["evaluation"], fitted_models=model,
203
+ input_sha256=sha(INPUT / "prepared" / task["file"]), methods={})
204
+ # Descriptive empirical shift; never use these laws to build a policy.
205
+ ps, pt = laws["source_held"], laws["target_evaluation"]
206
+ ms, mt = ps.sum((2, 3)), pt.sum((2, 3))
207
+ overlap = torch.minimum(ms, mt)
208
+ cs, ct = ps / ms[..., None, None].clamp_min(1e-300), pt / mt[..., None, None].clamp_min(1e-300)
209
+ conditional_tv = (.5 * (cs - ct).abs().sum((2, 3)) * overlap).sum() / overlap.sum().clamp_min(1e-300)
210
+ posterior_tv = (.5 * (cs.sum(3) - ct.sum(3)).abs().sum(2) * overlap).sum() / overlap.sum().clamp_min(1e-300)
211
+ shift = dict(task_index=ti, dataset=task["dataset"], environment=task["environment"], task=task["task"],
212
+ empirical_joint_tv=float(.5 * (ps-pt).abs().sum()), overlap_current_mass=float(overlap.sum()),
213
+ empirical_conditional_tv_overlap=float(conditional_tv), empirical_posterior_tv_overlap=float(posterior_tv),
214
+ target_mass_absent_source_held=float((mt * (ms == 0)).sum()), source_held_records=len(source_events), target_evaluation_records=len(archive["evaluation"]))
215
+ shifts.append(shift)
216
+ for method in METHODS:
217
+ code, k = encode(model, method)
218
+ compact = push(model, code, k)
219
+ policy = plan(compact)
220
+ fitted_loss = full_policy["expected"][:, 1:].mean((1, 2)) - policy["expected"][:, 1:].mean((1, 2))
221
+ state = dict(code=code.to(torch.uint8), compact=compact, choices=policy["choices"], values=policy["values"], expected=policy["expected"], traces={})
222
+ for domain, law in laws.items():
223
+ truth = push(law.expand(len(model), -1, -1, -1), code, k)
224
+ actual = realized(truth, policy)
225
+ sc = scores(compact, truth)
226
+ traces = []
227
+ for budget in [1, 2]:
228
+ for ci in range(3):
229
+ trace, utility = batched_replay(policy, code, events_by_domain[domain], budget, ci)
230
+ maximum_replay_error = max(maximum_replay_error, float((utility-actual["utility"][:, budget, ci]).abs().max()))
231
+ traces.append(trace)
232
+ state["traces"][domain] = torch.stack(traces, 1)
233
+ # Original independent replay API confirms batched access and aggregation.
234
+ check_i = next(i for i, m in enumerate(meta) if m["training"] == "target" and m["size"] == "64" and m["model"] == "support_mixture")
235
+ single = {**policy, "shape": (1, k), "choices": policy["choices"].reshape(3, len(meta), k, 3, 9)[:, check_i:check_i+1].reshape(3, k, 3, 9)}
236
+ a, q, seen = replay(single, code[check_i:check_i+1].to(torch.uint8), (events_by_domain[domain]//4)%64, events_by_domain[domain]%4, 2, 1)
237
+ assert torch.equal(torch.stack((a, q, seen), -1).to(torch.uint8), state["traces"][domain][check_i, 4])
238
+ for i, m in enumerate(meta):
239
+ if domain == "source_held" and m["training"] != "source":
240
+ continue
241
+ pred0 = float(policy["expected"][i, 0].mean())
242
+ predq = float(policy["expected"][i, 1:].mean())
243
+ act0 = float(actual["utility"][i, 0].mean())
244
+ actq = float(actual["utility"][i, 1:].mean())
245
+ row = dict(task_index=ti, dataset=task["dataset"], environment=task["environment"], task=task["task"], **m,
246
+ representation=method, evaluation=domain, evaluation_records=len(events_by_domain[domain]),
247
+ utility=actq, noquery_utility=act0, realized_query_increment=actq-act0,
248
+ predicted_utility=predq, predicted_noquery_utility=pred0, predicted_query_increment=predq-pred0,
249
+ query_increment_optimism=(predq-pred0)-(actq-act0), queries=float(actual["queries"][i, 1:].mean()),
250
+ fitted_compression_loss=float(fitted_loss[i]))
251
+ if m["model"] != "empirical_reference":
252
+ row.update({name:float(v[i]) for name, v in sc.items()})
253
+ raw_cond = model[i] / model[i].sum((1, 2), keepdim=True).clamp_min(1e-300)
254
+ loss_by_raw = -(law[0] * raw_cond.clamp_min(1e-300).log()).sum((1, 2))
255
+ # Same current-pattern weighting in both domains isolates
256
+ # conditional predictive mismatch from current-mass shifts.
257
+ row["common_current_joint_conditional_log_loss"] = float((loss_by_raw/law[0].sum((1,2)).clamp_min(1e-300)*overlap[0]).sum()/overlap.sum().clamp_min(1e-300))
258
+ rows.append(row)
259
+ # Exact same trained primary models/encoders/predictions as frozen study.
260
+ old = torch.load(INPUT / "evaluation" / f"primary_task_{ti:03}.pt", map_location=device, weights_only=True)
261
+ for i, m in enumerate(meta):
262
+ if m["training"] == "target" and m["size"] == "64" and (m["model"] == "support_mixture" or method == "identity"):
263
+ original = old[f'{m["sample_seed"]}:{m["model"]}:{method}']
264
+ assert torch.equal(code[i], original["code"][0].long())
265
+ maximum_original_error = max(maximum_original_error, float((policy["expected"][i]-original["expected"][0]).abs().max()))
266
+ saved["methods"][method] = state
267
+ torch.save(saved, out / "predictions" / f"task_{ti:03}.pt")
268
+ print(json.dumps(dict(task=ti, finished=len(shifts), tasks=len(manifest["tasks"]), elapsed_seconds=time.perf_counter()-start)), flush=True)
269
+ assert maximum_replay_error < 1e-12 and maximum_original_error < 1e-12
270
+ fields = sorted(set().union(*(r.keys() for r in rows)))
271
+ with gzip.open(out / "metrics.csv.gz", "wt", newline="") as f:
272
+ writer = csv.DictWriter(f, fieldnames=fields)
273
+ writer.writeheader(); writer.writerows(rows)
274
+ write(out / "shift_by_task.json", shifts)
275
+ torch.cuda.synchronize(device)
276
+ write(out / "execution.json", dict(material_passport=dict(material_type="experiment execution result", scientific_status="post-review exploratory", created_utc=utc()),
277
+ status="COMPLETED", metric_rows=len(rows), task_count=len(shifts), elapsed_seconds=time.perf_counter()-start,
278
+ maximum_replay_error=maximum_replay_error, maximum_original_primary_expected_error=maximum_original_error,
279
+ gpu=torch.cuda.get_device_name(device), device=device, torch_version=torch.__version__, python_version=platform.python_version(),
280
+ peak_gpu_allocated_bytes=torch.cuda.max_memory_allocated(device), peak_gpu_reserved_bytes=torch.cuda.max_memory_reserved(device),
281
+ script_sha256=sha(__file__), specification_sha256=sha(out / "specification.json")))
282
+
283
+
284
+ def aggregate(rows, field):
285
+ tasks = defaultdict(list)
286
+ for r in rows:
287
+ if field in r and r[field] != "":
288
+ tasks[(r["dataset"], r["environment"], r["task"])].append(float(r[field]))
289
+ envs = defaultdict(list)
290
+ for key, vals in tasks.items():
291
+ envs[key[:2]].append(sum(vals)/len(vals))
292
+ datasets = defaultdict(list)
293
+ for (d, e), vals in envs.items():
294
+ datasets[d].append(sum(vals)/len(vals))
295
+ ds = {d:sum(vals)/len(vals) for d, vals in datasets.items()}
296
+ return dict(overall=sum(ds.values())/len(ds), by_dataset=ds,
297
+ by_environment={d+":"+e:sum(v)/len(v) for (d,e),v in envs.items()}, tasks=len(tasks))
298
+
299
+
300
+ def recover_saved(out, device):
301
+ """Export saved fits/traces after correcting only diagnostic query-charge dtype."""
302
+ assert not (out / "metrics.csv.gz").exists()
303
+ spec = json.loads((out / "specification.json").read_text())
304
+ correction = dict(created_utc=utc(), authorization="Parent authorized routine numerical correction and artifact-only recovery after receiving failure report", failed_log_sha256=sha(out / "run.log"),
305
+ original_specification_script_sha256=spec["script_sha256"], corrected_script_sha256=sha(__file__),
306
+ failure="Final equality assertion rejected float32 intermediate cost*integer_query_count in diagnostic batched replay utility; policy tables and traces unaffected",
307
+ correction="Multiply cost by queries.double(), matching original independent replay utility convention",
308
+ estimators_rerun=0, encoders_rerun=0, planners_rerun=0, policies_changed=0,
309
+ recovery="Compute verified metrics and summaries from all 57 already saved model/policy/record-trace artifacts")
310
+ write(out / "implementation_correction.json", correction)
311
+ for path, digest in spec["input_hashes"].items():
312
+ assert sha(ROOT / path) == digest
313
+ torch.cuda.set_device(device)
314
+ torch.backends.cuda.matmul.allow_tf32 = False
315
+ torch.cuda.reset_peak_memory_stats(device)
316
+ start = time.perf_counter()
317
+ rows, shifts = [], []
318
+ maxerr, maxoriginal, maxfloat32 = 0., 0., 0.
319
+ manifest = json.loads((INPUT / "prepared/manifest.json").read_text())
320
+ for task in manifest["tasks"]:
321
+ ti = task["index"]
322
+ saved = torch.load(out / "predictions" / f"task_{ti:03}.pt", map_location=device, weights_only=True)
323
+ assert saved["input_sha256"] == sha(INPUT / "prepared" / task["file"])
324
+ meta, model = saved["metadata"], saved["fitted_models"]
325
+ full_expected = saved["methods"]["identity"]["expected"]
326
+ events_by_domain = {"target_evaluation":saved["target_evaluation_events"],"source_held":saved["source_held_events"]}
327
+ laws = {d:normalized_counts(e) for d,e in events_by_domain.items()}
328
+ ps, pt = laws["source_held"], laws["target_evaluation"]
329
+ ms, mt = ps.sum((2,3)),pt.sum((2,3))
330
+ overlap = torch.minimum(ms,mt)
331
+ cs,ct = ps/ms[...,None,None].clamp_min(1e-300),pt/mt[...,None,None].clamp_min(1e-300)
332
+ shifts.append(dict(task_index=ti,dataset=task["dataset"],environment=task["environment"],task=task["task"],
333
+ empirical_joint_tv=float(.5*(ps-pt).abs().sum()),overlap_current_mass=float(overlap.sum()),
334
+ empirical_conditional_tv_overlap=float((.5*(cs-ct).abs().sum((2,3))*overlap).sum()/overlap.sum().clamp_min(1e-300)),
335
+ empirical_posterior_tv_overlap=float((.5*(cs.sum(3)-ct.sum(3)).abs().sum(2)*overlap).sum()/overlap.sum().clamp_min(1e-300)),
336
+ target_mass_absent_source_held=float((mt*(ms==0)).sum()),source_held_records=len(events_by_domain["source_held"]),target_evaluation_records=len(events_by_domain["target_evaluation"])))
337
+ original = torch.load(INPUT / "evaluation" / f"primary_task_{ti:03}.pt", map_location=device, weights_only=True)
338
+ for method,state in saved["methods"].items():
339
+ code,compact = state["code"].long(),state["compact"]
340
+ k=compact.shape[1]
341
+ policy=dict(choices=state["choices"],values=state["values"],expected=state["expected"],costs=(0.,.01,.05),shape=(len(meta),k))
342
+ fitted_loss=full_expected[:,1:].mean((1,2))-policy["expected"][:,1:].mean((1,2))
343
+ for domain,law in laws.items():
344
+ events=events_by_domain[domain]
345
+ truth=push(law.expand(len(model),-1,-1,-1),code,k)
346
+ actual=realized(truth,policy)
347
+ sc=scores(compact,truth)
348
+ # Check every saved action/count/mask trace against newly evaluated
349
+ # fixed tables, and every double-precision charge against exact law replay.
350
+ for bi,budget in enumerate([1,2]):
351
+ for ci,cost in enumerate(policy["costs"]):
352
+ recorded=state["traces"][domain][:,bi*3+ci]
353
+ recomputed,utility=batched_replay(policy,code,events,budget,ci)
354
+ assert torch.equal(recorded,recomputed)
355
+ maxerr=max(maxerr,float((utility-actual["utility"][:,budget,ci]).abs().max()))
356
+ # Retain quantitative evidence for original failure.
357
+ q=recorded[:,:,1].long()
358
+ maxfloat32=max(maxfloat32,float((cost*q-cost*q.double()).abs().max()))
359
+ check_i=next(i for i,m in enumerate(meta) if m["training"]=="target" and m["size"]=="64" and m["model"]=="support_mixture")
360
+ single={**policy,"shape":(1,k),"choices":policy["choices"].reshape(3,len(meta),k,3,9)[:,check_i:check_i+1].reshape(3,k,3,9)}
361
+ a,q,seen=replay(single,code[check_i:check_i+1].to(torch.uint8),(events//4)%64,events%4,2,1)
362
+ assert torch.equal(torch.stack((a,q,seen),-1).to(torch.uint8),state["traces"][domain][check_i,4])
363
+ independent_u=torch.where(a==2,0.,torch.where(a==events//256,1.,-2.)).double()-.01*q.double()
364
+ maxerr=max(maxerr,abs(float(independent_u.mean())-float(actual["utility"][check_i,2,1])))
365
+ for i,m in enumerate(meta):
366
+ if domain=="source_held" and m["training"]!="source":continue
367
+ p0,pq=float(policy["expected"][i,0].mean()),float(policy["expected"][i,1:].mean())
368
+ a0,aq=float(actual["utility"][i,0].mean()),float(actual["utility"][i,1:].mean())
369
+ row=dict(task_index=ti,dataset=task["dataset"],environment=task["environment"],task=task["task"],**m,
370
+ representation=method,evaluation=domain,evaluation_records=len(events),utility=aq,noquery_utility=a0,
371
+ realized_query_increment=aq-a0,predicted_utility=pq,predicted_noquery_utility=p0,predicted_query_increment=pq-p0,
372
+ query_increment_optimism=(pq-p0)-(aq-a0),queries=float(actual["queries"][i,1:].mean()),fitted_compression_loss=float(fitted_loss[i]))
373
+ if m["model"]!="empirical_reference":
374
+ row.update({name:float(v[i]) for name,v in sc.items()})
375
+ rawcond=model[i]/model[i].sum((1,2),keepdim=True).clamp_min(1e-300)
376
+ loss=-(law[0]*rawcond.clamp_min(1e-300).log()).sum((1,2))
377
+ row["common_current_joint_conditional_log_loss"]=float((loss/law[0].sum((1,2)).clamp_min(1e-300)*overlap[0]).sum()/overlap.sum().clamp_min(1e-300))
378
+ rows.append(row)
379
+ for i,m in enumerate(meta):
380
+ if m["training"]=="target" and m["size"]=="64" and (m["model"]=="support_mixture" or method=="identity"):
381
+ old=original[f'{m["sample_seed"]}:{m["model"]}:{method}']
382
+ assert torch.equal(code[i],old["code"][0].long())
383
+ maxoriginal=max(maxoriginal,float((policy["expected"][i]-old["expected"][0]).abs().max()))
384
+ print(json.dumps(dict(recovered_task=ti,finished=len(shifts),elapsed_seconds=time.perf_counter()-start)),flush=True)
385
+ assert maxerr<1e-12 and maxoriginal<1e-12
386
+ fields=sorted(set().union(*(r.keys() for r in rows)))
387
+ with gzip.open(out/"metrics.csv.gz","wt",newline="") as f:
388
+ writer=csv.DictWriter(f,fieldnames=fields);writer.writeheader();writer.writerows(rows)
389
+ write(out/"shift_by_task.json",shifts)
390
+ write(out/"execution.json",dict(material_passport=dict(material_type="experiment execution result",scientific_status="post-review exploratory",created_utc=utc()),
391
+ status="COMPLETED_FROM_SAVED_ARTIFACTS_AFTER_DIAGNOSTIC_DTYPE_CORRECTION",metric_rows=len(rows),task_count=len(shifts),
392
+ original_fit_and_trace_seconds=11.562392784981057,recovery_seconds=time.perf_counter()-start,
393
+ maximum_replay_error=maxerr,maximum_original_primary_expected_error=maxoriginal,maximum_original_float32_query_charge_error=maxfloat32,
394
+ gpu=torch.cuda.get_device_name(device),device=device,torch_version=torch.__version__,python_version=platform.python_version(),
395
+ peak_gpu_allocated_bytes=torch.cuda.max_memory_allocated(device),peak_gpu_reserved_bytes=torch.cuda.max_memory_reserved(device),
396
+ script_sha256=sha(__file__),specification_sha256=sha(out/"specification.json"),original_results_and_gates_unchanged=True))
397
+
398
+
399
+ def summarize(out, device):
400
+ with gzip.open(out / "metrics.csv.gz", "rt") as f:
401
+ rows = list(csv.DictReader(f))
402
+ keys = ["training", "size", "model", "representation", "evaluation"]
403
+ fields = ["utility", "noquery_utility", "realized_query_increment", "predicted_utility", "predicted_noquery_utility", "predicted_query_increment", "query_increment_optimism", "queries", "fitted_compression_loss", "posterior_brier", "posterior_log_loss", "joint_conditional_log_loss", "observation_given_h_log_loss", "common_current_joint_conditional_log_loss"]
404
+ groups = defaultdict(list)
405
+ for r in rows:
406
+ groups[tuple(r[k] for k in keys)].append(r)
407
+ overall = []
408
+ for key, rs in groups.items():
409
+ obj = dict(zip(keys, key))
410
+ for f in fields:
411
+ if any(r.get(f, "") != "" for r in rs):
412
+ obj[f] = aggregate(rs, f)
413
+ overall.append(obj)
414
+ write(out / "overall.json", overall)
415
+ # Contrasts pair at task and seed, then resample target environments only.
416
+ contrasts = []
417
+ def contrast(label, left, right, field="utility"):
418
+ a = { (r["dataset"],r["environment"],r["task"],r["sample_seed"]):r for r in rows if left(r) }
419
+ b = { (r["dataset"],r["environment"],r["task"],r["sample_seed"]):r for r in rows if right(r) }
420
+ pairs = []
421
+ for key in sorted(a.keys() & b.keys()):
422
+ pairs.append(dict(dataset=key[0],environment=key[1],task=key[2],difference=float(a[key][field])-float(b[key][field])))
423
+ if not pairs:
424
+ return
425
+ agg = aggregate(pairs, "difference")
426
+ g = torch.Generator(device=device).manual_seed(20261109)
427
+ boots = []
428
+ for d in sorted(agg["by_dataset"]):
429
+ vals = torch.tensor([v for name,v in sorted(agg["by_environment"].items()) if name.startswith(d+":")], device=device, dtype=torch.float64)
430
+ ix = torch.randint(len(vals), (10000,len(vals)), generator=g,device=device)
431
+ boots.append(vals[ix].mean(1))
432
+ low, high = torch.quantile(torch.stack(boots).mean(0),torch.tensor([.025,.975],device=device,dtype=torch.float64)).tolist()
433
+ contrasts.append(dict(label=label, field=field, **agg, descriptive_paired_95_interval=[low,high], paired_fit_rows=len(pairs), environment_resampling_only=True))
434
+ base = lambda r:r["model"] == "support_mixture" and r["evaluation"] == "target_evaluation"
435
+ for size in ["64","256","all"]:
436
+ contrast("joint_minus_posterior_target_"+size,
437
+ lambda r:base(r) and r["training"]=="target" and r["size"]==size and r["representation"]=="query_joint",
438
+ lambda r:base(r) and r["training"]=="target" and r["size"]==size and r["representation"]=="posterior_clusters")
439
+ for method in METHODS:
440
+ contrast("target_minus_source_"+size+"_"+method,
441
+ lambda r:base(r) and r["training"]=="target" and r["size"]==size and r["representation"]==method,
442
+ lambda r:base(r) and r["training"]=="source" and r["size"]==size and r["representation"]==method)
443
+ for method in METHODS:
444
+ for size in ["256","all"]:
445
+ for field in ["utility","noquery_utility","realized_query_increment","joint_conditional_log_loss"]:
446
+ contrast("target_"+size+"_minus_64_"+method,
447
+ lambda r:base(r) and r["training"]=="target" and r["size"]==size and r["representation"]==method,
448
+ lambda r:base(r) and r["training"]=="target" and r["size"]=="64" and r["representation"]==method, field)
449
+ for size in ["64","256","all"]:
450
+ for field in ["utility","noquery_utility","realized_query_increment","joint_conditional_log_loss","common_current_joint_conditional_log_loss"]:
451
+ contrast("source_fixed_policy_target_minus_sourceheld_"+size,
452
+ lambda r:r["model"]=="support_mixture" and r["training"]=="source" and r["size"]==size and r["representation"]=="identity" and r["evaluation"]=="target_evaluation",
453
+ lambda r:r["model"]=="support_mixture" and r["training"]=="source" and r["size"]==size and r["representation"]=="identity" and r["evaluation"]=="source_held", field)
454
+ write(out / "paired_contrasts.json",dict(scientific_status="post-review exploratory; unadjusted descriptive intervals; no new confirmation", bootstrap_unit="Target environment stratified by fixed dataset; tasks and seeds collapsed first", contrasts=contrasts))
455
+ shifts = json.loads((out / "shift_by_task.json").read_text())
456
+ write(out / "shift_overall.json", {f:aggregate(shifts,f) for f in ["empirical_joint_tv","overlap_current_mass","empirical_conditional_tv_overlap","empirical_posterior_tv_overlap","target_mass_absent_source_held"]})
457
+ # Compact tables retain the original weighting and explicitly flag subsets.
458
+ table=[]
459
+ def table_row(label,selected):
460
+ taskset={(r["dataset"],r["environment"],r["task"]) for r in selected}
461
+ envset={(r["dataset"],r["environment"]) for r in selected}
462
+ samples=[int(r["sample_count"]) for r in selected]
463
+ table.append(dict(analysis=label,tasks=len(taskset),
464
+ hydraulic_tasks=sum(x[0]=="hydraulic" for x in taskset),gas_tasks=sum(x[0]=="gas" for x in taskset),
465
+ hydraulic_environments=sum(x[0]=="hydraulic" for x in envset),gas_environments=sum(x[0]=="gas" for x in envset),
466
+ labels_min=min(samples),labels_max=max(samples),
467
+ **{f:aggregate(selected,f)["overall"] for f in fields if any(r.get(f,"")!="" for r in selected)}))
468
+ for training in ["target","source"]:
469
+ for size in ["64","all"]:
470
+ for method in (METHODS if training=="target" else ["identity"]):
471
+ for evaluation in (["target_evaluation"] if training=="target" else ["source_held","target_evaluation"]):
472
+ table_row(f"{training}_{size}_{method}_on_{evaluation}_all57",[r for r in rows if r["training"]==training and r["size"]==size and r["model"]=="support_mixture" and r["representation"]==method and r["evaluation"]==evaluation])
473
+ eligible256={r["task_index"] for r in rows if r["training"]=="target" and r["size"]=="256"}
474
+ for size in ["64","256"]:
475
+ for method in METHODS:
476
+ table_row(f"target_{size}_{method}_on_matched256eligible43",[r for r in rows if r["task_index"] in eligible256 and r["training"]=="target" and r["size"]==size and r["model"]=="support_mixture" and r["representation"]==method and r["evaluation"]=="target_evaluation"])
477
+ for method in METHODS:
478
+ table_row(f"empirical_alladaptation_{method}_on_target_evaluation_all57",[r for r in rows if r["model"]=="empirical_reference" and r["representation"]==method])
479
+ tablefields=sorted(set().union(*(r.keys() for r in table)))
480
+ with (out/"article_table.csv").open("w",newline="") as f:
481
+ writer=csv.DictWriter(f,fieldnames=tablefields);writer.writeheader();writer.writerows(table)
482
+ write(out/"table_provenance.json",dict(created_utc=utc(),script_sha256=sha(__file__),
483
+ input_metrics_sha256=sha(out/"metrics.csv.gz"),table_sha256=sha(out/"article_table.csv"),
484
+ design_changed=False,scientific_status="post-review exploratory",weighting="As specification; each table row lists support and independent environment counts",
485
+ caution="The 256-label subset has only one hydraulic environment; empirical TV contains finite-count noise and is not a causal decomposition. Gas record artifacts remain research-only local data."))
486
+ print(json.dumps(dict(summary_groups=len(overall), contrasts=len(contrasts))), flush=True)
487
+
488
+
489
+ def main():
490
+ global INPUT
491
+ parser = argparse.ArgumentParser(description=__doc__)
492
+ parser.add_argument("stage", choices=["specify","run","recover","summarize"])
493
+ parser.add_argument("--output", type=Path, default=DEFAULT_OUTPUT)
494
+ parser.add_argument("--device", default="cuda:4")
495
+ parser.add_argument("--release", type=Path, default=RELEASE, help="Unmodified release containing unionmind/ and protocols/")
496
+ parser.add_argument("--input", type=Path, default=INPUT, help="Previously opened external study artifacts: prepared/ and evaluation/")
497
+ args = parser.parse_args()
498
+ INPUT = args.input.resolve()
499
+ out = args.output.resolve()
500
+ if args.stage == "specify":
501
+ out.mkdir(parents=True, exist_ok=True)
502
+ (out / "predictions").mkdir(exist_ok=True)
503
+ specification(out,args.device)
504
+ elif args.stage == "run":
505
+ evaluate(out,args.device)
506
+ summarize(out,args.device)
507
+ elif args.stage == "recover":
508
+ recover_saved(out,args.device)
509
+ summarize(out,args.device)
510
+ else:
511
+ summarize(out,args.device)
512
+
513
+
514
+ if __name__ == "__main__":
515
+ main()
revision_r1/external/shift_by_task.json ADDED
@@ -0,0 +1,743 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "task_index": 0,
4
+ "dataset": "hydraulic",
5
+ "environment": "context1=73",
6
+ "task": "cooler",
7
+ "empirical_joint_tv": 0.5891016200294551,
8
+ "overlap_current_mass": 0.4204221894943544,
9
+ "empirical_conditional_tv_overlap": 0.022652965903783285,
10
+ "empirical_posterior_tv_overlap": 0.0,
11
+ "target_mass_absent_source_held": 0.41904761904761906,
12
+ "source_held_records": 97,
13
+ "target_evaluation_records": 210
14
+ },
15
+ {
16
+ "task_index": 1,
17
+ "dataset": "hydraulic",
18
+ "environment": "context1=80",
19
+ "task": "cooler",
20
+ "empirical_joint_tv": 0.6455194290245836,
21
+ "overlap_current_mass": 0.4006344171292625,
22
+ "empirical_conditional_tv_overlap": 0.11520190023752969,
23
+ "empirical_posterior_tv_overlap": 0.0,
24
+ "target_mass_absent_source_held": 0.4,
25
+ "source_held_records": 97,
26
+ "target_evaluation_records": 130
27
+ },
28
+ {
29
+ "task_index": 2,
30
+ "dataset": "hydraulic",
31
+ "environment": "context1=90",
32
+ "task": "cooler",
33
+ "empirical_joint_tv": 0.4949390815370197,
34
+ "overlap_current_mass": 0.5638238050609184,
35
+ "empirical_conditional_tv_overlap": 0.13436391843971632,
36
+ "empirical_posterior_tv_overlap": 2.86413918586809e-17,
37
+ "target_mass_absent_source_held": 0.3090909090909091,
38
+ "source_held_records": 97,
39
+ "target_evaluation_records": 110
40
+ },
41
+ {
42
+ "task_index": 3,
43
+ "dataset": "hydraulic",
44
+ "environment": "context0=3",
45
+ "task": "valve",
46
+ "empirical_joint_tv": 0.9999999999999999,
47
+ "overlap_current_mass": 0.3076923076923077,
48
+ "empirical_conditional_tv_overlap": 1.0,
49
+ "empirical_posterior_tv_overlap": 0.34560439560439554,
50
+ "target_mass_absent_source_held": 0.30107526881720426,
51
+ "source_held_records": 91,
52
+ "target_evaluation_records": 465
53
+ },
54
+ {
55
+ "task_index": 4,
56
+ "dataset": "hydraulic",
57
+ "environment": "context0=20",
58
+ "task": "valve",
59
+ "empirical_joint_tv": 1.0,
60
+ "overlap_current_mass": 0.3076923076923077,
61
+ "empirical_conditional_tv_overlap": 1.0,
62
+ "empirical_posterior_tv_overlap": 0.3014660012476606,
63
+ "target_mass_absent_source_held": 0.10196078431372549,
64
+ "source_held_records": 91,
65
+ "target_evaluation_records": 510
66
+ },
67
+ {
68
+ "task_index": 5,
69
+ "dataset": "hydraulic",
70
+ "environment": "context0=3",
71
+ "task": "pump",
72
+ "empirical_joint_tv": 0.9999999999999999,
73
+ "overlap_current_mass": 0.30769230769230765,
74
+ "empirical_conditional_tv_overlap": 1.0,
75
+ "empirical_posterior_tv_overlap": 0.8646153846153845,
76
+ "target_mass_absent_source_held": 0.30107526881720426,
77
+ "source_held_records": 91,
78
+ "target_evaluation_records": 465
79
+ },
80
+ {
81
+ "task_index": 6,
82
+ "dataset": "hydraulic",
83
+ "environment": "context0=20",
84
+ "task": "pump",
85
+ "empirical_joint_tv": 1.0,
86
+ "overlap_current_mass": 0.30769230769230765,
87
+ "empirical_conditional_tv_overlap": 1.0,
88
+ "empirical_posterior_tv_overlap": 0.6157205240174672,
89
+ "target_mass_absent_source_held": 0.10196078431372549,
90
+ "source_held_records": 91,
91
+ "target_evaluation_records": 510
92
+ },
93
+ {
94
+ "task_index": 7,
95
+ "dataset": "hydraulic",
96
+ "environment": "context0=3",
97
+ "task": "accumulator",
98
+ "empirical_joint_tv": 0.9999999999999999,
99
+ "overlap_current_mass": 0.30769230769230765,
100
+ "empirical_conditional_tv_overlap": 1.0,
101
+ "empirical_posterior_tv_overlap": 0.7753846153846153,
102
+ "target_mass_absent_source_held": 0.30107526881720426,
103
+ "source_held_records": 91,
104
+ "target_evaluation_records": 465
105
+ },
106
+ {
107
+ "task_index": 8,
108
+ "dataset": "hydraulic",
109
+ "environment": "context0=20",
110
+ "task": "accumulator",
111
+ "empirical_joint_tv": 1.0,
112
+ "overlap_current_mass": 0.30769230769230765,
113
+ "empirical_conditional_tv_overlap": 0.9999999999999998,
114
+ "empirical_posterior_tv_overlap": 0.08951965065502185,
115
+ "target_mass_absent_source_held": 0.10196078431372549,
116
+ "source_held_records": 91,
117
+ "target_evaluation_records": 510
118
+ },
119
+ {
120
+ "task_index": 9,
121
+ "dataset": "gas",
122
+ "environment": "batch6",
123
+ "task": "gas1_vs_2",
124
+ "empirical_joint_tv": 0.6525147928994083,
125
+ "overlap_current_mass": 0.37140039447731754,
126
+ "empirical_conditional_tv_overlap": 0.0967938892854072,
127
+ "empirical_posterior_tv_overlap": 0.04700992883347823,
128
+ "target_mass_absent_source_held": 0.022916666666666665,
129
+ "source_held_records": 338,
130
+ "target_evaluation_records": 480
131
+ },
132
+ {
133
+ "task_index": 10,
134
+ "dataset": "gas",
135
+ "environment": "batch7",
136
+ "task": "gas1_vs_2",
137
+ "empirical_joint_tv": 0.49545848565936756,
138
+ "overlap_current_mass": 0.5531602155805978,
139
+ "empirical_conditional_tv_overlap": 0.17988007041133974,
140
+ "empirical_posterior_tv_overlap": 0.07768418217812881,
141
+ "target_mass_absent_source_held": 0.2372611464968153,
142
+ "source_held_records": 338,
143
+ "target_evaluation_records": 628
144
+ },
145
+ {
146
+ "task_index": 13,
147
+ "dataset": "gas",
148
+ "environment": "batch10",
149
+ "task": "gas1_vs_2",
150
+ "empirical_joint_tv": 0.7252295449908184,
151
+ "overlap_current_mass": 0.4283411548663538,
152
+ "empirical_conditional_tv_overlap": 0.5057037565792537,
153
+ "empirical_posterior_tv_overlap": 0.497653454097474,
154
+ "target_mass_absent_source_held": 0.22586206896551725,
155
+ "source_held_records": 338,
156
+ "target_evaluation_records": 580
157
+ },
158
+ {
159
+ "task_index": 14,
160
+ "dataset": "gas",
161
+ "environment": "batch6",
162
+ "task": "gas1_vs_3",
163
+ "empirical_joint_tv": 0.5750790638836178,
164
+ "overlap_current_mass": 0.4597780461158071,
165
+ "empirical_conditional_tv_overlap": 0.21295510392558914,
166
+ "empirical_posterior_tv_overlap": 0.07479508985261864,
167
+ "target_mass_absent_source_held": 0.07624633431085043,
168
+ "source_held_records": 255,
169
+ "target_evaluation_records": 341
170
+ },
171
+ {
172
+ "task_index": 15,
173
+ "dataset": "gas",
174
+ "environment": "batch7",
175
+ "task": "gas1_vs_3",
176
+ "empirical_joint_tv": 0.5147367608460707,
177
+ "overlap_current_mass": 0.5225412999845607,
178
+ "empirical_conditional_tv_overlap": 0.10063793699050405,
179
+ "empirical_posterior_tv_overlap": 0.029726850735940166,
180
+ "target_mass_absent_source_held": 0.12204724409448818,
181
+ "source_held_records": 255,
182
+ "target_evaluation_records": 508
183
+ },
184
+ {
185
+ "task_index": 17,
186
+ "dataset": "gas",
187
+ "environment": "batch9",
188
+ "task": "gas1_vs_3",
189
+ "empirical_joint_tv": 0.7607843137254902,
190
+ "overlap_current_mass": 0.2627450980392157,
191
+ "empirical_conditional_tv_overlap": 0.272636815920398,
192
+ "empirical_posterior_tv_overlap": 0.18308457711442785,
193
+ "target_mass_absent_source_held": 0.0,
194
+ "source_held_records": 255,
195
+ "target_evaluation_records": 90
196
+ },
197
+ {
198
+ "task_index": 18,
199
+ "dataset": "gas",
200
+ "environment": "batch10",
201
+ "task": "gas1_vs_3",
202
+ "empirical_joint_tv": 0.7049428104575164,
203
+ "overlap_current_mass": 0.4061887254901961,
204
+ "empirical_conditional_tv_overlap": 0.4271407873087596,
205
+ "empirical_posterior_tv_overlap": 0.34255774684355955,
206
+ "target_mass_absent_source_held": 0.06770833333333333,
207
+ "source_held_records": 255,
208
+ "target_evaluation_records": 576
209
+ },
210
+ {
211
+ "task_index": 20,
212
+ "dataset": "gas",
213
+ "environment": "batch7",
214
+ "task": "gas1_vs_4",
215
+ "empirical_joint_tv": 0.5835438073574398,
216
+ "overlap_current_mass": 0.41809553690485524,
217
+ "empirical_conditional_tv_overlap": 0.15349406350528697,
218
+ "empirical_posterior_tv_overlap": 0.14940959760816513,
219
+ "target_mass_absent_source_held": 0.03934426229508197,
220
+ "source_held_records": 209,
221
+ "target_evaluation_records": 610
222
+ },
223
+ {
224
+ "task_index": 23,
225
+ "dataset": "gas",
226
+ "environment": "batch10",
227
+ "task": "gas1_vs_4",
228
+ "empirical_joint_tv": 0.6279124193883919,
229
+ "overlap_current_mass": 0.4050689966021773,
230
+ "empirical_conditional_tv_overlap": 0.13612500940242103,
231
+ "empirical_posterior_tv_overlap": 0.07040986774360053,
232
+ "target_mass_absent_source_held": 0.04710144927536232,
233
+ "source_held_records": 209,
234
+ "target_evaluation_records": 552
235
+ },
236
+ {
237
+ "task_index": 24,
238
+ "dataset": "gas",
239
+ "environment": "batch6",
240
+ "task": "gas1_vs_5",
241
+ "empirical_joint_tv": 0.4761886761886762,
242
+ "overlap_current_mass": 0.6386964386964387,
243
+ "empirical_conditional_tv_overlap": 0.21147300294298005,
244
+ "empirical_posterior_tv_overlap": 0.19232550737466725,
245
+ "target_mass_absent_source_held": 0.018711018711018712,
246
+ "source_held_records": 385,
247
+ "target_evaluation_records": 481
248
+ },
249
+ {
250
+ "task_index": 25,
251
+ "dataset": "gas",
252
+ "environment": "batch7",
253
+ "task": "gas1_vs_5",
254
+ "empirical_joint_tv": 0.5707052441229656,
255
+ "overlap_current_mass": 0.5403254972875227,
256
+ "empirical_conditional_tv_overlap": 0.2699232523524929,
257
+ "empirical_posterior_tv_overlap": 0.25119377854739744,
258
+ "target_mass_absent_source_held": 0.09945750452079566,
259
+ "source_held_records": 385,
260
+ "target_evaluation_records": 553
261
+ },
262
+ {
263
+ "task_index": 26,
264
+ "dataset": "gas",
265
+ "environment": "batch8",
266
+ "task": "gas1_vs_5",
267
+ "empirical_joint_tv": 0.590566037735849,
268
+ "overlap_current_mass": 0.40943396226415096,
269
+ "empirical_conditional_tv_overlap": 0.07292505923842325,
270
+ "empirical_posterior_tv_overlap": 0.07292505923842325,
271
+ "target_mass_absent_source_held": 0.0,
272
+ "source_held_records": 385,
273
+ "target_evaluation_records": 106
274
+ },
275
+ {
276
+ "task_index": 28,
277
+ "dataset": "gas",
278
+ "environment": "batch10",
279
+ "task": "gas1_vs_5",
280
+ "empirical_joint_tv": 0.7159182366928847,
281
+ "overlap_current_mass": 0.7303228461679165,
282
+ "empirical_conditional_tv_overlap": 0.6639737282499998,
283
+ "empirical_posterior_tv_overlap": 0.652671781560124,
284
+ "target_mass_absent_source_held": 0.08274647887323942,
285
+ "source_held_records": 385,
286
+ "target_evaluation_records": 568
287
+ },
288
+ {
289
+ "task_index": 29,
290
+ "dataset": "gas",
291
+ "environment": "batch6",
292
+ "task": "gas1_vs_6",
293
+ "empirical_joint_tv": 0.6327683615819208,
294
+ "overlap_current_mass": 0.37160800603295874,
295
+ "empirical_conditional_tv_overlap": 0.40441907700569835,
296
+ "empirical_posterior_tv_overlap": 0.40441907700569835,
297
+ "target_mass_absent_source_held": 0.015317286652078774,
298
+ "source_held_records": 177,
299
+ "target_evaluation_records": 457
300
+ },
301
+ {
302
+ "task_index": 30,
303
+ "dataset": "gas",
304
+ "environment": "batch7",
305
+ "task": "gas1_vs_6",
306
+ "empirical_joint_tv": 0.6958751576984258,
307
+ "overlap_current_mass": 0.41407492732159507,
308
+ "empirical_conditional_tv_overlap": 0.6075821071209954,
309
+ "empirical_posterior_tv_overlap": 0.5880430953313543,
310
+ "target_mass_absent_source_held": 0.1197411003236246,
311
+ "source_held_records": 177,
312
+ "target_evaluation_records": 618
313
+ },
314
+ {
315
+ "task_index": 32,
316
+ "dataset": "gas",
317
+ "environment": "batch9",
318
+ "task": "gas1_vs_6",
319
+ "empirical_joint_tv": 0.6327683615819208,
320
+ "overlap_current_mass": 0.3672316384180791,
321
+ "empirical_conditional_tv_overlap": 0.5384615384615384,
322
+ "empirical_posterior_tv_overlap": 0.5384615384615384,
323
+ "target_mass_absent_source_held": 0.0,
324
+ "source_held_records": 177,
325
+ "target_evaluation_records": 91
326
+ },
327
+ {
328
+ "task_index": 33,
329
+ "dataset": "gas",
330
+ "environment": "batch10",
331
+ "task": "gas1_vs_6",
332
+ "empirical_joint_tv": 0.6091301566997123,
333
+ "overlap_current_mass": 0.43262978360515936,
334
+ "empirical_conditional_tv_overlap": 0.558964700310025,
335
+ "empirical_posterior_tv_overlap": 0.550913222405153,
336
+ "target_mass_absent_source_held": 0.01257861635220126,
337
+ "source_held_records": 177,
338
+ "target_evaluation_records": 636
339
+ },
340
+ {
341
+ "task_index": 34,
342
+ "dataset": "gas",
343
+ "environment": "batch6",
344
+ "task": "gas2_vs_3",
345
+ "empirical_joint_tv": 0.6544434519611991,
346
+ "overlap_current_mass": 0.35477097143215336,
347
+ "empirical_conditional_tv_overlap": 0.24161786483966885,
348
+ "empirical_posterior_tv_overlap": 0.1855972808916146,
349
+ "target_mass_absent_source_held": 0.16104868913857676,
350
+ "source_held_records": 239,
351
+ "target_evaluation_records": 267
352
+ },
353
+ {
354
+ "task_index": 35,
355
+ "dataset": "gas",
356
+ "environment": "batch7",
357
+ "task": "gas2_vs_3",
358
+ "empirical_joint_tv": 0.7481554824491361,
359
+ "overlap_current_mass": 0.39395221821201565,
360
+ "empirical_conditional_tv_overlap": 0.6199017107064613,
361
+ "empirical_posterior_tv_overlap": 0.5149738337243449,
362
+ "target_mass_absent_source_held": 0.25,
363
+ "source_held_records": 239,
364
+ "target_evaluation_records": 524
365
+ },
366
+ {
367
+ "task_index": 37,
368
+ "dataset": "gas",
369
+ "environment": "batch9",
370
+ "task": "gas2_vs_3",
371
+ "empirical_joint_tv": 0.8786610878661087,
372
+ "overlap_current_mass": 0.12133891213389121,
373
+ "empirical_conditional_tv_overlap": 0.3977011494252874,
374
+ "empirical_posterior_tv_overlap": 0.3977011494252874,
375
+ "target_mass_absent_source_held": 0.53125,
376
+ "source_held_records": 239,
377
+ "target_evaluation_records": 64
378
+ },
379
+ {
380
+ "task_index": 38,
381
+ "dataset": "gas",
382
+ "environment": "batch10",
383
+ "task": "gas2_vs_3",
384
+ "empirical_joint_tv": 0.7213929005263868,
385
+ "overlap_current_mass": 0.4157713591577811,
386
+ "empirical_conditional_tv_overlap": 0.49642875215628623,
387
+ "empirical_posterior_tv_overlap": 0.44338837322005364,
388
+ "target_mass_absent_source_held": 0.35322580645161294,
389
+ "source_held_records": 239,
390
+ "target_evaluation_records": 620
391
+ },
392
+ {
393
+ "task_index": 40,
394
+ "dataset": "gas",
395
+ "environment": "batch7",
396
+ "task": "gas2_vs_4",
397
+ "empirical_joint_tv": 0.6515750964260292,
398
+ "overlap_current_mass": 0.41268685129699223,
399
+ "empirical_conditional_tv_overlap": 0.4458951213954642,
400
+ "empirical_posterior_tv_overlap": 0.30977899624504307,
401
+ "target_mass_absent_source_held": 0.26837060702875404,
402
+ "source_held_records": 193,
403
+ "target_evaluation_records": 626
404
+ },
405
+ {
406
+ "task_index": 43,
407
+ "dataset": "gas",
408
+ "environment": "batch10",
409
+ "task": "gas2_vs_4",
410
+ "empirical_joint_tv": 0.7339691205619502,
411
+ "overlap_current_mass": 0.31912230065723124,
412
+ "empirical_conditional_tv_overlap": 0.39851565155399254,
413
+ "empirical_posterior_tv_overlap": 0.3672921077446364,
414
+ "target_mass_absent_source_held": 0.16610738255033558,
415
+ "source_held_records": 193,
416
+ "target_evaluation_records": 596
417
+ },
418
+ {
419
+ "task_index": 44,
420
+ "dataset": "gas",
421
+ "environment": "batch6",
422
+ "task": "gas2_vs_5",
423
+ "empirical_joint_tv": 0.5464200342249123,
424
+ "overlap_current_mass": 0.45603696823209017,
425
+ "empirical_conditional_tv_overlap": 0.1789962863774888,
426
+ "empirical_posterior_tv_overlap": 0.1596942727730253,
427
+ "target_mass_absent_source_held": 0.11302211302211303,
428
+ "source_held_records": 369,
429
+ "target_evaluation_records": 407
430
+ },
431
+ {
432
+ "task_index": 45,
433
+ "dataset": "gas",
434
+ "environment": "batch7",
435
+ "task": "gas2_vs_5",
436
+ "empirical_joint_tv": 0.6536690147217816,
437
+ "overlap_current_mass": 0.47187811069674845,
438
+ "empirical_conditional_tv_overlap": 0.2851729943473662,
439
+ "empirical_posterior_tv_overlap": 0.28068628505264265,
440
+ "target_mass_absent_source_held": 0.3479789103690685,
441
+ "source_held_records": 369,
442
+ "target_evaluation_records": 569
443
+ },
444
+ {
445
+ "task_index": 46,
446
+ "dataset": "gas",
447
+ "environment": "batch8",
448
+ "task": "gas2_vs_5",
449
+ "empirical_joint_tv": 0.8285780027611596,
450
+ "overlap_current_mass": 0.2196911591757427,
451
+ "empirical_conditional_tv_overlap": 0.3386942860467823,
452
+ "empirical_posterior_tv_overlap": 0.3386942860467823,
453
+ "target_mass_absent_source_held": 0.7547169811320754,
454
+ "source_held_records": 369,
455
+ "target_evaluation_records": 106
456
+ },
457
+ {
458
+ "task_index": 48,
459
+ "dataset": "gas",
460
+ "environment": "batch10",
461
+ "task": "gas2_vs_5",
462
+ "empirical_joint_tv": 0.7809660449545672,
463
+ "overlap_current_mass": 0.2971465008767734,
464
+ "empirical_conditional_tv_overlap": 0.30540410245849353,
465
+ "empirical_posterior_tv_overlap": 0.2818978940013974,
466
+ "target_mass_absent_source_held": 0.05392156862745098,
467
+ "source_held_records": 369,
468
+ "target_evaluation_records": 612
469
+ },
470
+ {
471
+ "task_index": 49,
472
+ "dataset": "gas",
473
+ "environment": "batch6",
474
+ "task": "gas2_vs_6",
475
+ "empirical_joint_tv": 0.629307688565266,
476
+ "overlap_current_mass": 0.3733032774921752,
477
+ "empirical_conditional_tv_overlap": 0.33795152248052684,
478
+ "empirical_posterior_tv_overlap": 0.32629448554002843,
479
+ "target_mass_absent_source_held": 0.057441253263707574,
480
+ "source_held_records": 161,
481
+ "target_evaluation_records": 383
482
+ },
483
+ {
484
+ "task_index": 50,
485
+ "dataset": "gas",
486
+ "environment": "batch7",
487
+ "task": "gas2_vs_6",
488
+ "empirical_joint_tv": 0.6953680663048376,
489
+ "overlap_current_mass": 0.4500558418402335,
490
+ "empirical_conditional_tv_overlap": 0.5535311076862758,
491
+ "empirical_posterior_tv_overlap": 0.4116158625907264,
492
+ "target_mass_absent_source_held": 0.26025236593059936,
493
+ "source_held_records": 161,
494
+ "target_evaluation_records": 634
495
+ },
496
+ {
497
+ "task_index": 52,
498
+ "dataset": "gas",
499
+ "environment": "batch9",
500
+ "task": "gas2_vs_6",
501
+ "empirical_joint_tv": 0.8136645962732918,
502
+ "overlap_current_mass": 0.3756330625895843,
503
+ "empirical_conditional_tv_overlap": 0.5039430170440092,
504
+ "empirical_posterior_tv_overlap": 0.5039430170440092,
505
+ "target_mass_absent_source_held": 0.0,
506
+ "source_held_records": 161,
507
+ "target_evaluation_records": 65
508
+ },
509
+ {
510
+ "task_index": 53,
511
+ "dataset": "gas",
512
+ "environment": "batch10",
513
+ "task": "gas2_vs_6",
514
+ "empirical_joint_tv": 0.8132352941176471,
515
+ "overlap_current_mass": 0.46251370113262696,
516
+ "empirical_conditional_tv_overlap": 0.6759810844684891,
517
+ "empirical_posterior_tv_overlap": 0.532736489055476,
518
+ "target_mass_absent_source_held": 0.34705882352941175,
519
+ "source_held_records": 161,
520
+ "target_evaluation_records": 680
521
+ },
522
+ {
523
+ "task_index": 55,
524
+ "dataset": "gas",
525
+ "environment": "batch7",
526
+ "task": "gas3_vs_4",
527
+ "empirical_joint_tv": 0.6470355731225297,
528
+ "overlap_current_mass": 0.3608695652173913,
529
+ "empirical_conditional_tv_overlap": 0.49843510391772583,
530
+ "empirical_posterior_tv_overlap": 0.43945793650682347,
531
+ "target_mass_absent_source_held": 0.21936758893280633,
532
+ "source_held_records": 110,
533
+ "target_evaluation_records": 506
534
+ },
535
+ {
536
+ "task_index": 58,
537
+ "dataset": "gas",
538
+ "environment": "batch10",
539
+ "task": "gas3_vs_4",
540
+ "empirical_joint_tv": 0.8231879606879606,
541
+ "overlap_current_mass": 0.42767199017199015,
542
+ "empirical_conditional_tv_overlap": 0.634041110907394,
543
+ "empirical_posterior_tv_overlap": 0.5045233454696252,
544
+ "target_mass_absent_source_held": 0.18581081081081083,
545
+ "source_held_records": 110,
546
+ "target_evaluation_records": 592
547
+ },
548
+ {
549
+ "task_index": 59,
550
+ "dataset": "gas",
551
+ "environment": "batch6",
552
+ "task": "gas3_vs_5",
553
+ "empirical_joint_tv": 0.44115958668197475,
554
+ "overlap_current_mass": 0.6581515499425947,
555
+ "empirical_conditional_tv_overlap": 0.17422076547450627,
556
+ "empirical_posterior_tv_overlap": 0.0971143394103828,
557
+ "target_mass_absent_source_held": 0.0037313432835820895,
558
+ "source_held_records": 286,
559
+ "target_evaluation_records": 268
560
+ },
561
+ {
562
+ "task_index": 60,
563
+ "dataset": "gas",
564
+ "environment": "batch7",
565
+ "task": "gas3_vs_5",
566
+ "empirical_joint_tv": 0.6446804865513106,
567
+ "overlap_current_mass": 0.6366439796283894,
568
+ "empirical_conditional_tv_overlap": 0.5124004229911332,
569
+ "empirical_posterior_tv_overlap": 0.4630126614281086,
570
+ "target_mass_absent_source_held": 0.18930957683741648,
571
+ "source_held_records": 286,
572
+ "target_evaluation_records": 449
573
+ },
574
+ {
575
+ "task_index": 61,
576
+ "dataset": "gas",
577
+ "environment": "batch8",
578
+ "task": "gas3_vs_5",
579
+ "empirical_joint_tv": 0.9353146853146854,
580
+ "overlap_current_mass": 0.15034965034965037,
581
+ "empirical_conditional_tv_overlap": 0.6046511627906976,
582
+ "empirical_posterior_tv_overlap": 0.5813953488372093,
583
+ "target_mass_absent_source_held": 0.21153846153846156,
584
+ "source_held_records": 286,
585
+ "target_evaluation_records": 104
586
+ },
587
+ {
588
+ "task_index": 63,
589
+ "dataset": "gas",
590
+ "environment": "batch10",
591
+ "task": "gas3_vs_5",
592
+ "empirical_joint_tv": 0.587861152005889,
593
+ "overlap_current_mass": 0.5445574162679425,
594
+ "empirical_conditional_tv_overlap": 0.25630833252918106,
595
+ "empirical_posterior_tv_overlap": 0.18376488214615264,
596
+ "target_mass_absent_source_held": 0.06414473684210527,
597
+ "source_held_records": 286,
598
+ "target_evaluation_records": 608
599
+ },
600
+ {
601
+ "task_index": 64,
602
+ "dataset": "gas",
603
+ "environment": "batch6",
604
+ "task": "gas3_vs_6",
605
+ "empirical_joint_tv": 0.8314417822614544,
606
+ "overlap_current_mass": 0.36433375367801596,
607
+ "empirical_conditional_tv_overlap": 0.5713873665993654,
608
+ "empirical_posterior_tv_overlap": 0.3464089991346985,
609
+ "target_mass_absent_source_held": 0.06147540983606557,
610
+ "source_held_records": 78,
611
+ "target_evaluation_records": 244
612
+ },
613
+ {
614
+ "task_index": 65,
615
+ "dataset": "gas",
616
+ "environment": "batch7",
617
+ "task": "gas3_vs_6",
618
+ "empirical_joint_tv": 0.9218297914795968,
619
+ "overlap_current_mass": 0.2318667065748778,
620
+ "empirical_conditional_tv_overlap": 0.6910856228708551,
621
+ "empirical_posterior_tv_overlap": 0.6237305475607231,
622
+ "target_mass_absent_source_held": 0.2898832684824903,
623
+ "source_held_records": 78,
624
+ "target_evaluation_records": 514
625
+ },
626
+ {
627
+ "task_index": 67,
628
+ "dataset": "gas",
629
+ "environment": "batch9",
630
+ "task": "gas3_vs_6",
631
+ "empirical_joint_tv": 0.6025641025641026,
632
+ "overlap_current_mass": 0.4230769230769231,
633
+ "empirical_conditional_tv_overlap": 0.06060606060606061,
634
+ "empirical_posterior_tv_overlap": 0.030303030303030304,
635
+ "target_mass_absent_source_held": 0.030927835051546393,
636
+ "source_held_records": 78,
637
+ "target_evaluation_records": 97
638
+ },
639
+ {
640
+ "task_index": 68,
641
+ "dataset": "gas",
642
+ "environment": "batch10",
643
+ "task": "gas3_vs_6",
644
+ "empirical_joint_tv": 0.6158777120315583,
645
+ "overlap_current_mass": 0.5443786982248521,
646
+ "empirical_conditional_tv_overlap": 0.32674974000268386,
647
+ "empirical_posterior_tv_overlap": 0.31002079978529257,
648
+ "target_mass_absent_source_held": 0.2751479289940828,
649
+ "source_held_records": 78,
650
+ "target_evaluation_records": 676
651
+ },
652
+ {
653
+ "task_index": 70,
654
+ "dataset": "gas",
655
+ "environment": "batch7",
656
+ "task": "gas4_vs_5",
657
+ "empirical_joint_tv": 0.7660843920145191,
658
+ "overlap_current_mass": 0.5896324863883847,
659
+ "empirical_conditional_tv_overlap": 0.6704926684133363,
660
+ "empirical_posterior_tv_overlap": 0.6623088991152682,
661
+ "target_mass_absent_source_held": 0.21415607985480944,
662
+ "source_held_records": 240,
663
+ "target_evaluation_records": 551
664
+ },
665
+ {
666
+ "task_index": 73,
667
+ "dataset": "gas",
668
+ "environment": "batch10",
669
+ "task": "gas4_vs_5",
670
+ "empirical_joint_tv": 0.761986301369863,
671
+ "overlap_current_mass": 0.6600456621004567,
672
+ "empirical_conditional_tv_overlap": 0.6487041835499334,
673
+ "empirical_posterior_tv_overlap": 0.6380813145476109,
674
+ "target_mass_absent_source_held": 0.07705479452054795,
675
+ "source_held_records": 240,
676
+ "target_evaluation_records": 584
677
+ },
678
+ {
679
+ "task_index": 75,
680
+ "dataset": "gas",
681
+ "environment": "batch7",
682
+ "task": "gas4_vs_6",
683
+ "empirical_joint_tv": 0.96875,
684
+ "overlap_current_mass": 0.09780844155844155,
685
+ "empirical_conditional_tv_overlap": 0.8437989857076994,
686
+ "empirical_posterior_tv_overlap": 0.16330106039649608,
687
+ "target_mass_absent_source_held": 0.12987012987012989,
688
+ "source_held_records": 32,
689
+ "target_evaluation_records": 616
690
+ },
691
+ {
692
+ "task_index": 78,
693
+ "dataset": "gas",
694
+ "environment": "batch10",
695
+ "task": "gas4_vs_6",
696
+ "empirical_joint_tv": 0.9672162576687117,
697
+ "overlap_current_mass": 0.04505368098159509,
698
+ "empirical_conditional_tv_overlap": 0.6148246405263438,
699
+ "empirical_posterior_tv_overlap": 0.3424842149944288,
700
+ "target_mass_absent_source_held": 0.3711656441717791,
701
+ "source_held_records": 32,
702
+ "target_evaluation_records": 652
703
+ },
704
+ {
705
+ "task_index": 79,
706
+ "dataset": "gas",
707
+ "environment": "batch6",
708
+ "task": "gas5_vs_6",
709
+ "empirical_joint_tv": 0.578525641025641,
710
+ "overlap_current_mass": 0.4919871794871795,
711
+ "empirical_conditional_tv_overlap": 0.19297279172255224,
712
+ "empirical_posterior_tv_overlap": 0.18808679823721022,
713
+ "target_mass_absent_source_held": 0.11718750000000001,
714
+ "source_held_records": 208,
715
+ "target_evaluation_records": 384
716
+ },
717
+ {
718
+ "task_index": 80,
719
+ "dataset": "gas",
720
+ "environment": "batch7",
721
+ "task": "gas5_vs_6",
722
+ "empirical_joint_tv": 0.7978533094812165,
723
+ "overlap_current_mass": 0.44689177101967803,
724
+ "empirical_conditional_tv_overlap": 0.5812557616410505,
725
+ "empirical_posterior_tv_overlap": 0.574980221652726,
726
+ "target_mass_absent_source_held": 0.44186046511627913,
727
+ "source_held_records": 208,
728
+ "target_evaluation_records": 559
729
+ },
730
+ {
731
+ "task_index": 83,
732
+ "dataset": "gas",
733
+ "environment": "batch10",
734
+ "task": "gas5_vs_6",
735
+ "empirical_joint_tv": 0.6132830492860435,
736
+ "overlap_current_mass": 0.529335559649931,
737
+ "empirical_conditional_tv_overlap": 0.274517045008442,
738
+ "empirical_posterior_tv_overlap": 0.27072698878866275,
739
+ "target_mass_absent_source_held": 0.10628742514970059,
740
+ "source_held_records": 208,
741
+ "target_evaluation_records": 668
742
+ }
743
+ ]
revision_r1/external/shift_overall.json ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "empirical_joint_tv": {
3
+ "overall": 0.7264348414836221,
4
+ "by_dataset": {
5
+ "hydraulic": 0.7459120261182116,
6
+ "gas": 0.7069576568490327
7
+ },
8
+ "by_environment": {
9
+ "hydraulic:context1=73": 0.5891016200294551,
10
+ "hydraulic:context1=80": 0.6455194290245836,
11
+ "hydraulic:context1=90": 0.4949390815370197,
12
+ "hydraulic:context0=3": 0.9999999999999999,
13
+ "hydraulic:context0=20": 1.0,
14
+ "gas:batch6": 0.6017849079274071,
15
+ "gas:batch7": 0.6903547112156817,
16
+ "gas:batch10": 0.7201405974293267,
17
+ "gas:batch9": 0.7376884924021828,
18
+ "gas:batch8": 0.7848195752705647
19
+ },
20
+ "tasks": 57
21
+ },
22
+ "overlap_current_mass": {
23
+ "overall": 0.39033438581068947,
24
+ "by_dataset": {
25
+ "hydraulic": 0.40005300541383015,
26
+ "gas": 0.38061576620754883
27
+ },
28
+ "by_environment": {
29
+ "hydraulic:context1=73": 0.4204221894943544,
30
+ "hydraulic:context1=80": 0.4006344171292625,
31
+ "hydraulic:context1=90": 0.5638238050609184,
32
+ "hydraulic:context0=3": 0.30769230769230765,
33
+ "hydraulic:context0=20": 0.30769230769230765,
34
+ "gas:batch6": 0.45400665855867306,
35
+ "gas:batch7": 0.43603222996748564,
36
+ "gas:batch10": 0.44320989173019887,
37
+ "gas:batch9": 0.3100051268515387,
38
+ "gas:batch8": 0.259824923929848
39
+ },
40
+ "tasks": 57
41
+ },
42
+ "empirical_conditional_tv_overlap": {
43
+ "overall": 0.41570487409678125,
44
+ "by_dataset": {
45
+ "hydraulic": 0.4544437569162058,
46
+ "gas": 0.37696599127735675
47
+ },
48
+ "by_environment": {
49
+ "hydraulic:context1=73": 0.022652965903783285,
50
+ "hydraulic:context1=80": 0.11520190023752969,
51
+ "hydraulic:context1=90": 0.13436391843971632,
52
+ "hydraulic:context0=3": 1.0,
53
+ "hydraulic:context0=20": 1.0,
54
+ "gas:batch6": 0.2622787670653783,
55
+ "gas:batch7": 0.4675657953371991,
56
+ "gas:batch10": 0.4615588416674466,
57
+ "gas:batch9": 0.35466971629145877,
58
+ "gas:batch8": 0.33875683602530104
59
+ },
60
+ "tasks": 57
61
+ },
62
+ "empirical_posterior_tv_overlap": {
63
+ "overall": 0.26299167483291297,
64
+ "by_dataset": {
65
+ "hydraulic": 0.19948737143496967,
66
+ "gas": 0.3264959782308563
67
+ },
68
+ "by_environment": {
69
+ "hydraulic:context1=73": 0.0,
70
+ "hydraulic:context1=80": 0.0,
71
+ "hydraulic:context1=90": 2.86413918586809e-17,
72
+ "hydraulic:context0=3": 0.6618681318681318,
73
+ "hydraulic:context0=20": 0.33556872530671655,
74
+ "gas:batch6": 0.20217457790534224,
75
+ "gas:batch7": 0.3693269205782592,
76
+ "gas:batch10": 0.3992748321602165,
77
+ "gas:batch9": 0.3306986624696586,
78
+ "gas:batch8": 0.33100489804080496
79
+ },
80
+ "tasks": 57
81
+ },
82
+ "target_mass_absent_source_held": {
83
+ "overall": 0.2408059865231789,
84
+ "by_dataset": {
85
+ "hydraulic": 0.3062349162538916,
86
+ "gas": 0.17537705679246618
87
+ },
88
+ "by_environment": {
89
+ "hydraulic:context1=73": 0.41904761904761906,
90
+ "hydraulic:context1=80": 0.4,
91
+ "hydraulic:context1=90": 0.3090909090909091,
92
+ "hydraulic:context0=3": 0.30107526881720426,
93
+ "hydraulic:context0=20": 0.1019607843137255,
94
+ "gas:batch6": 0.06470976148846597,
95
+ "gas:batch7": 0.21526001667687727,
96
+ "gas:batch10": 0.16239479122983275,
97
+ "gas:batch9": 0.11243556701030928,
98
+ "gas:batch8": 0.32208514755684564
99
+ },
100
+ "tasks": 57
101
+ }
102
+ }
revision_r1/external/specification.json ADDED
@@ -0,0 +1,123 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "material_passport": {
3
+ "material_type": "post-review exploratory experiment specification",
4
+ "provenance": "Previously opened external sensor archives and original eligible task manifest",
5
+ "license": "Hydraulic UCI CC BY 4.0; gas legacy research-only restriction retained; local research use only",
6
+ "created_utc": "2026-11-09T14:37:09.140791+00:00",
7
+ "scientific_status": "exploratory; does not replace original confirmation"
8
+ },
9
+ "purpose": "Distinguish small within-fitted-law encoding headroom, finite labeled support, and observed source-target transfer failure without claiming an identifiable population decomposition",
10
+ "device": "cuda:4",
11
+ "dtype": "float64",
12
+ "backbone_executions": 0,
13
+ "trainable_parameters": 0,
14
+ "task_indices": [
15
+ 0,
16
+ 1,
17
+ 2,
18
+ 3,
19
+ 4,
20
+ 5,
21
+ 6,
22
+ 7,
23
+ 8,
24
+ 9,
25
+ 10,
26
+ 13,
27
+ 14,
28
+ 15,
29
+ 17,
30
+ 18,
31
+ 20,
32
+ 23,
33
+ 24,
34
+ 25,
35
+ 26,
36
+ 28,
37
+ 29,
38
+ 30,
39
+ 32,
40
+ 33,
41
+ 34,
42
+ 35,
43
+ 37,
44
+ 38,
45
+ 40,
46
+ 43,
47
+ 44,
48
+ 45,
49
+ 46,
50
+ 48,
51
+ 49,
52
+ 50,
53
+ 52,
54
+ 53,
55
+ 55,
56
+ 58,
57
+ 59,
58
+ 60,
59
+ 61,
60
+ 63,
61
+ 64,
62
+ 65,
63
+ 67,
64
+ 68,
65
+ 70,
66
+ 73,
67
+ 75,
68
+ 78,
69
+ 79,
70
+ 80,
71
+ 83
72
+ ],
73
+ "seeds": [
74
+ 20111,
75
+ 20112,
76
+ 20113
77
+ ],
78
+ "methods": [
79
+ "query_joint",
80
+ "posterior_clusters",
81
+ "identity"
82
+ ],
83
+ "estimators": [
84
+ "support_mixture",
85
+ "joint_jeffreys",
86
+ "naive_bayes"
87
+ ],
88
+ "interventions": [
89
+ "Original target64 exact seed selection, nested target256 when available, and all original target adaptation records",
90
+ "Source group-disjoint train/held split: SHA256('review_r1_source_held:' + original_group) mod 5 == 0 held; source64/256/all train only",
91
+ "Each fixed policy replayed on original target evaluation and, for source-trained policies, source held records",
92
+ "All-target-adaptation unsmoothed empirical reference law: encoders and exact policy fitted without target evaluation access, evaluated on original independent-group target evaluation; not a population oracle"
93
+ ],
94
+ "source_preprocessing": "Original source-only binary thresholds kept; source-held records contributed unlabeled preprocessing before this post-review split, never labeled model fitting",
95
+ "metrics": [
96
+ "paired realized utility and no-query utility",
97
+ "model-predicted versus held-realized query increment",
98
+ "fitted encoding loss",
99
+ "posterior/joint/observation conditional log losses",
100
+ "same-policy source-held versus shifted-target scores",
101
+ "finite-sample empirical conditional TV on overlapping current support",
102
+ "target mass outside source-held current support"
103
+ ],
104
+ "aggregation": "Original weighting: seeds, then tasks within environment, then environments within dataset, then the two datasets equally. Sample-size contrasts restricted to identical tasks. Costs 0,.01,.05 and budgets 1,2 averaged; no-query average from budget0.",
105
+ "inference": "Descriptive 95% paired bootstrap by target environment within fixed dataset; collapse tasks/seeds first. Few clusters; not population transfer inference; no confirmatory gate rescues.",
106
+ "limitations": [
107
+ "True external law is unknown; histogram reference and TV have finite-sample error",
108
+ "Source-target contrast also includes differences in physical groups and sample composition; no causal attribution percentages",
109
+ "Labels-all support varies by task; 256 contrasts require available support and use matched task subsets",
110
+ "Source group split introduces a new held partition; original target adaptation/evaluation split unchanged",
111
+ "Three seeds are repeated fitting samples, not independent environmental replications",
112
+ "Empirical-reference policies are exploratory benchmarks and their NLL is not reported because zero fitted cells yield uninformative infinite log scores"
113
+ ],
114
+ "input_hashes": {
115
+ "artifacts/external_decision_sufficiency_01/prepared/manifest.json": "c97a59bbf5f04fe8dc3ccfe00986b4df288b8e7ef2f3fbd9eb12f3636dceef5d",
116
+ "articles/make_1/release/huggingface-revised/protocols/external_sufficiency_protocol_01.json": "e0229ef8a50270b7a590c6ed79a850ec8f902d17f2c8213111bf432669ef9d04",
117
+ "articles/make_1/release/huggingface-revised/unionmind/decision_sufficiency/core.py": "fa2f7405a8d0dc6ec96ebae1bb2a2b810d03ef1fd2877f5ebef7df1722992cee",
118
+ "articles/make_1/release/huggingface-revised/unionmind/external_sufficiency/runtime.py": "fa19ce1b7fba9e6953e9e4081d4b6001ed55c33edc48a05960d5a13ae6aad61d"
119
+ },
120
+ "script_sha256": "18a499610df5927f78ae97af5bd4249b3c18beaef98ea39caecc5ad1303b7956",
121
+ "no_recipe_selection": true,
122
+ "original_results_and_gates_unchanged": true
123
+ }
revision_r1/external/table_provenance.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "created_utc": "2026-11-09T14:43:13.718361+00:00",
3
+ "script_sha256": "33e4d7ff42763ccd48db10013dd91831af13aa368a115c39987f65636a06c63f",
4
+ "input_metrics_sha256": "9813ff30348ad76cb8b672e022a5f3fe852b8f0ee13625e7a5799400cfc01120",
5
+ "table_sha256": "09b7110cc185574d01f5aeb4681aa20910048a50429b274a5422969c2654fad2",
6
+ "design_changed": false,
7
+ "scientific_status": "post-review exploratory",
8
+ "weighting": "As specification; each table row lists support and independent environment counts",
9
+ "caution": "The 256-label subset has only one hydraulic environment; empirical TV contains finite-count noise and is not a causal decomposition. Gas record artifacts remain research-only local data."
10
+ }
revision_r1/sensor/analysis_spec.json ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "id": "MAKE-review-r1-sensor-01",
3
+ "timestamp_utc": "2026-11-09T14:35:00.586928+00:00",
4
+ "status": "Prospective specification for new post-review exploratory analyses on previously opened data; no new confirmation or model promotion",
5
+ "authorization": "Reviewer-driven additional experiments requested by author, delegated within academic-research-suite experiment workflow",
6
+ "cohorts": {
7
+ "hydraulic": {
8
+ "sizes": [
9
+ 3,
10
+ 4,
11
+ 3,
12
+ 4
13
+ ],
14
+ "seeds": [
15
+ 20751,
16
+ 20752,
17
+ 20753
18
+ ]
19
+ },
20
+ "temperature": {
21
+ "sizes": [
22
+ 10,
23
+ 3
24
+ ],
25
+ "seeds": [
26
+ 20851,
27
+ 20852,
28
+ 20853
29
+ ]
30
+ }
31
+ },
32
+ "budgets": [
33
+ 64,
34
+ 128
35
+ ],
36
+ "prefixes": "Use exact saved fit_ids for every original seed; assert nested prefix, all source, no unseen binding fitting labels; normalize using exact saved prefix mean/std",
37
+ "optimizer": "400 full-batch AdamW updates, lr=.003, weight_decay=.01, float32 neural training, same original initialization seeds and architectures; float64 probability/statistical operations on cuda:3; TF32 disabled",
38
+ "questions": [
39
+ "At fixed pretrained-free forward mean, how do shared shrunken full, diagonal, and identity covariance affect ranking and probabilities?",
40
+ "At fixed joint forward posterior, what changes when replacing it by the product of its exact factor marginals?",
41
+ "With identical factor-one-hot -> SiLU -> sensor-mean decoder and fixed identity Gaussian likelihood, what changes when the decoder training objective is conditional joint cross-entropy instead of observation MSE?",
42
+ "At fixed inverse classifier architecture, does unweighted versus original class-balanced factor cross-entropy account for the gap?",
43
+ "Can a scalar temperature selected using only out-of-fold fitting-prefix predictions address held probability failure?",
44
+ "In hydraulic data, is reversal reproduced under fixed64 normalization/covariance, longer optimization, duplicate64 input, and equal-count additions of new versus already observed bindings?"
45
+ ],
46
+ "main_grid": [
47
+ "MSE forward: full covariance, diagonal, identity; full-covariance posterior projected to product of marginals",
48
+ "CE forward decoder: fixed identity covariance; train conditional CE over full supplied ontology including unobserved bindings without held labels",
49
+ "Inverse MLP: original class-balanced and unweighted factor CE; compare raw posterior and original source per-factor calibration",
50
+ "For full-cov MSE, identity MSE, identity CE, balanced inverse, unweighted inverse: scalar temperature from fixed grid [1,2,4,8,16,32,64] minimizing prefix OOF joint log loss; apply without target tuning; preserve raw outputs"
51
+ ],
52
+ "crossfit": "Four original deterministic folds i mod4, seed+101+fold; one full fit plus four folds for each of four neural objective/architecture choices, 400 steps each. Forward covariance from MSE source OOF residuals with original shrinkage .5/full + .5/diag + .05 I; calibrator OOF forward scores use this same pooled OOF covariance and are not nested. Disclose possible source calibration optimism. No held/source-outside-prefix tuning.",
53
+ "hydraulic_reversal": {
54
+ "normalization_and_inference_covariance": "fixed64 for all controlled refits; use saved64 source OOF covariance; unchanged uniform prior",
55
+ "scenarios": [
56
+ "saved64/400",
57
+ "duplicate64/400 (128 duplicated rows, no new information)",
58
+ "64/800",
59
+ "actual128/400 fixed64 normalization",
60
+ "actual128/800 fixed64 normalization",
61
+ "64 plus q old-binding extra rows/400",
62
+ "64 plus q new-binding extra rows/400"
63
+ ],
64
+ "matched_additions": "Partition later64 rows by joint-binding membership in first64; q=min(partition sizes); preserve prefix order, use first q each. Same initial seed and number of updates. This controls added count, normalization, covariance and initialization; row values and binding identities still differ. All rows inside original128 prefix."
65
+ },
66
+ "outcomes": "Every seed retained. Joint Top1/Top3 with original tolerance1e-12 uniform tie credit, NLL stable log-softmax, multiclass Brier, confidence, ECE10, coverage and accuracy at .9; source OOF and held seen/unseen, temperature each day9-12 and equal-day mean, hydraulic record mean. Standardized sensor-mean MSE and Gaussian observation scores; original source outside128 diagnostic without fitting. No confirmatory p-values or favorable selection.",
67
+ "replay_and_cost": "Save fit_ids, normalization, all full/fold network states, derived mean/covariance, OOF/held log posterior and per-record scores, scalar temperatures, losses at updates0,100,200,400 (and800), raw/source test diagnostics; record all fit/prediction/calibration/serialization times, trainable counts, forward executions, approximate dense arithmetic counts separately from latency, peak GPU bytes and complete process time. No backbone or CPU scientific execution; no offload or DeepSpeed.",
68
+ "checks": [
69
+ "Gaussian CE decoder identity likelihood score matches frozen posterior scores up to additive normalizer",
70
+ "factor-marginal projection normalizes and preserves each factor marginal",
71
+ "duplicate-row MSE gradients agree within float32 tolerance",
72
+ "all original MSE and balanced inverse refit full predictions match saved checkpoints within1e-5 (report residuals)",
73
+ "all reported scalars finite except empty-stratum values explicitly null",
74
+ "all original seeds/prefix counts and original held targets retained"
75
+ ],
76
+ "limitations": "Post-review, opened-data mechanism probes, no independent systems or new confirmation. Objective can be isolated within fixed decoder only; pure modeling direction remains entangled with input/output parameterization. Product projection removes posterior dependence without retraining and does not isolate all factorization assumptions. Temperature calibration cannot establish operational reliability or extrapolate to new platforms.",
77
+ "command": "PYTHONDONTWRITEBYTECODE=1 python articles/make_1/reproduce/review_r1_sensor.py --mode run --device cuda:3 --output articles/make_1/metadata/review_r1/sensor/run_01",
78
+ "hard_timeout_seconds": 1800
79
+ }
revision_r1/sensor/export_safe.py ADDED
@@ -0,0 +1,39 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Export complete review artifacts to lossless non-executable tensor trees."""
2
+ from pathlib import Path
3
+ import argparse,hashlib,json,shutil,time
4
+ import torch
5
+ from safe_tree import save_tree,load_tree
6
+
7
+
8
+ def sha(path):return hashlib.sha256(path.read_bytes()).hexdigest()
9
+
10
+
11
+ def compare(a,b):
12
+ if torch.is_tensor(a):assert torch.is_tensor(b) and torch.equal(a,b);return 1
13
+ if isinstance(a,dict):assert list(a)==list(b);return sum(compare(a[k],b[k]) for k in a)
14
+ if isinstance(a,(tuple,list)):assert type(a)==type(b) and len(a)==len(b);return sum(compare(x,y) for x,y in zip(a,b))
15
+ assert a==b;return 0
16
+
17
+
18
+ def main():
19
+ parser=argparse.ArgumentParser();parser.add_argument('--source',default=str(Path(__file__).resolve().parent));parser.add_argument('--output',required=True);parser.add_argument('--device',default='cuda:3');args=parser.parse_args()
20
+ source=Path(args.source).resolve();out=Path(args.output).resolve();assert not out.exists();out.mkdir(parents=True);torch.cuda.set_device(args.device);start=time.perf_counter();trees=[];tensor_count=0
21
+ for name in ['run_01','objective_extension_01']:
22
+ directory=out/name;directory.mkdir()
23
+ for file in sorted((source/name).iterdir()):
24
+ if file.suffix=='.pt':
25
+ original=torch.load(file,map_location=args.device,weights_only=True);base=directory/file.stem;save_tree(original,base)
26
+ restored=load_tree(base,args.device);count=compare(original,restored);tensor_count+=count
27
+ trees.append(dict(file=str(base.relative_to(out)),source_sha256=sha(file),tensors=count,json_sha256=sha(base.with_suffix('.json')),safetensors_sha256=sha(base.with_suffix('.safetensors'))))
28
+ elif file.suffix in ('.json','.csv','.md','.py'):shutil.copyfile(file,directory/file.name)
29
+ for name in ['analysis_spec.json','objective_extension_spec.json','findings.json','findings.md','primary_table.csv','full_replay.json','verify_all.py','safe_tree.py','objective_extension.py','export_safe.py']:
30
+ shutil.copyfile(source/name,out/name)
31
+ trainer=source.parents[2]/'reproduce/review_r1_sensor.py';shutil.copyfile(trainer,out/'review_r1_sensor.py')
32
+ assert not list(out.rglob('*.pt'))
33
+ report=dict(format='Lossless Safetensors + JSON tensor trees; no executable pickle checkpoints',trees=trees,roundtrip_tensor_count=tensor_count,roundtrip_maximum_difference=0,seconds=time.perf_counter()-start,
34
+ source_artifacts_modified=False,device=args.device,export_uses_cpu_only_for_serialization=True,
35
+ replay_command='python verify_all.py --release /path/to/original-release --base run_01 --extension objective_extension_01 --output replay.json --device cuda:0')
36
+ (out/'EXPORT_MANIFEST.json').write_text(json.dumps(report,indent=2)+'\n');print(json.dumps({k:v for k,v in report.items() if k!='trees'},indent=2))
37
+
38
+
39
+ if __name__=='__main__':main()
revision_r1/sensor/findings.json ADDED
@@ -0,0 +1,123 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "Post-review exploratory, all data previously opened",
3
+ "aggregation": {
4
+ "hydraulic": "Mean across all304held unseen-binding rows, then3seeds",
5
+ "temperature_primary_top3": "Equal average across days9,10,11,12, then3seeds;65uniqueunseenepisodes",
6
+ "probability_metrics": "Pool unseen rows and seeds;304x3=912 hydraulic and65x3=195 gas prediction instances, not independent observations or systems"
7
+ },
8
+ "primary_table": [
9
+ {
10
+ "method": "forward_mse_full",
11
+ "description": "MSE decoder; full covariance",
12
+ "hydraulic_64": 0.4682017543859649,
13
+ "hydraulic_128": 0.3267543859649123,
14
+ "temperature_64": 0.6700367647058822,
15
+ "temperature_128": 0.7267156862745098
16
+ },
17
+ {
18
+ "method": "forward_mse_diag",
19
+ "description": "MSE decoder; diagonal covariance",
20
+ "hydraulic_64": 0.15901428538103854,
21
+ "hydraulic_128": 0.12720831799779167,
22
+ "temperature_64": 0.4818206377378537,
23
+ "temperature_128": 0.4679810501786921
24
+ },
25
+ {
26
+ "method": "forward_mse_identity",
27
+ "description": "MSE decoder; identity covariance",
28
+ "hydraulic_64": 0.10416666666666667,
29
+ "hydraulic_128": 0.06579719545342229,
30
+ "temperature_64": 0.3230540060851927,
31
+ "temperature_128": 0.3443985891770501
32
+ },
33
+ {
34
+ "method": "forward_mse_full_product",
35
+ "description": "MSE/full; product of factor marginals",
36
+ "hydraulic_64": 0.46381578947368424,
37
+ "hydraulic_128": 0.42214912280701755,
38
+ "temperature_64": 0.6804534313725489,
39
+ "temperature_128": 0.7058823529411765
40
+ },
41
+ {
42
+ "method": "forward_ce_identity",
43
+ "description": "Joint CE decoder; identity covariance",
44
+ "hydraulic_64": 0.0668859649122807,
45
+ "hydraulic_128": 0.041666666666666664,
46
+ "temperature_64": 0.015318627450980393,
47
+ "temperature_128": 0.05606617647058824
48
+ },
49
+ {
50
+ "method": "forward_factor_ce_identity",
51
+ "description": "Factor marginal CE decoder; identity covariance",
52
+ "hydraulic_64": 0.11622807017543858,
53
+ "hydraulic_128": 0.06578947368421052,
54
+ "temperature_64": 0.07659313725490197,
55
+ "temperature_128": 0.0974264705882353
56
+ },
57
+ {
58
+ "method": "inverse_balanced",
59
+ "description": "Direct balanced CE; raw",
60
+ "hydraulic_64": 0.24122807017543857,
61
+ "hydraulic_128": 0.5285087719298246,
62
+ "temperature_64": 0.3633578431372549,
63
+ "temperature_128": 0.45067401960784315
64
+ },
65
+ {
66
+ "method": "inverse_unweighted",
67
+ "description": "Direct unweighted CE; raw",
68
+ "hydraulic_64": 0.2434210526315789,
69
+ "hydraulic_128": 0.5131578947368421,
70
+ "temperature_64": 0.3026960784313726,
71
+ "temperature_128": 0.42524509803921573
72
+ },
73
+ {
74
+ "method": "inverse_balanced_factor_temperature",
75
+ "description": "Direct balanced CE; original source factor temperatures",
76
+ "hydraulic_64": 0.25548245614035087,
77
+ "hydraulic_128": 0.5219298245614036,
78
+ "temperature_64": 0.3633578431372549,
79
+ "temperature_128": 0.45067401960784315
80
+ },
81
+ {
82
+ "method": "inverse_unweighted_factor_temperature",
83
+ "description": "Direct unweighted CE; source factor temperatures",
84
+ "hydraulic_64": 0.24890350877192982,
85
+ "hydraulic_128": 0.5043859649122807,
86
+ "temperature_64": 0.3026960784313726,
87
+ "temperature_128": 0.42524509803921573
88
+ }
89
+ ],
90
+ "compute": {
91
+ "new_fits": 318,
92
+ "optimizer_updates": 129600,
93
+ "measured_fit_inference_analysis_seconds": 246.56308321398683,
94
+ "training_seconds": 222.80916958197486,
95
+ "estimated_training_dense_matmul_flops": 1256574873600,
96
+ "peak_gpu_allocated_bytes": 288381440,
97
+ "main_process_seconds_including_startup": 171.79019222594798,
98
+ "startup_limitation": "Extension timer starts after imports/CUDA context initialization; extension framework startup is not measured. All fitting/inference/calibration/serialization/replay costs are inside measured analysis time. Summary/replay postprocessing recorded separately. FLOPs are partial dense matmul estimates, not total inference compute.",
99
+ "cpu_scientific_execution": false,
100
+ "backbone_executions": 0,
101
+ "gpu": "NVIDIA H200 cuda:3",
102
+ "frozen_release_changed": false
103
+ },
104
+ "files": {
105
+ "main_summary": "run_01/summary.csv",
106
+ "main_all_metrics": "run_01/metrics.csv",
107
+ "extension_summary": "objective_extension_01/summary.csv",
108
+ "extension_per_seed": "objective_extension_01/summary_per_seed.csv",
109
+ "hydraulic_reversal": "run_01/reversal_summary.csv",
110
+ "hydraulic_mean_diagnostics": "run_01/reversal_diagnostics.json",
111
+ "frozen_source_diagnostics": "run_01/mean_diagnostics.csv",
112
+ "source_calibration_choices": "run_01/calibration_choices.csv"
113
+ },
114
+ "safe_public_export": "public_safe/",
115
+ "verification": {
116
+ "new_saved_metric_score_error": 0,
117
+ "original_archived_top1_top3_vector_error": 0,
118
+ "original_archived_maximum_confidence_error": 1.6825558705191312e-06,
119
+ "original_archived_maximum_nll_error": 2.488220451368761e-05,
120
+ "evaluated_method_arrays": 192,
121
+ "reversal_scenarios": 21
122
+ }
123
+ }
revision_r1/sensor/full_replay.json ADDED
@@ -0,0 +1,39 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "passed",
3
+ "refitted_models": 0,
4
+ "cells": 24,
5
+ "held_methods": 192,
6
+ "reversal_scenarios": 21,
7
+ "maximum_saved_score_error": 0.0,
8
+ "maximum_metric_error": 0.0,
9
+ "maximum_reversal_score_error": 0.0,
10
+ "original_archived_metric_vector_max_errors": {
11
+ "hydraulic_factorized_neural_top1": 0.0,
12
+ "hydraulic_factorized_neural_top3": 0.0,
13
+ "hydraulic_factorized_neural_nll": 2.488220451368761e-05,
14
+ "hydraulic_factorized_neural_confidence": 1.6825558705191312e-06,
15
+ "hydraulic_factorized_classifier_top1": 0.0,
16
+ "hydraulic_factorized_classifier_top3": 0.0,
17
+ "hydraulic_factorized_classifier_nll": 1.6876770096985183e-06,
18
+ "hydraulic_factorized_classifier_confidence": 4.1392002436690944e-08,
19
+ "temperature_factorized_neural_top1": 0.0,
20
+ "temperature_factorized_neural_top3": 0.0,
21
+ "temperature_factorized_neural_nll": 1.0818260330225371e-05,
22
+ "temperature_factorized_neural_confidence": 1.6244467709869426e-06,
23
+ "temperature_factorized_classifier_top1": 0.0,
24
+ "temperature_factorized_classifier_top3": 0.0,
25
+ "temperature_factorized_classifier_nll": 9.632170758777647e-07,
26
+ "temperature_factorized_classifier_confidence": 2.6531173480393022e-08
27
+ },
28
+ "archive_tolerances": {
29
+ "top1": 1e-05,
30
+ "top3": 1e-05,
31
+ "confidence": 1e-05,
32
+ "nll": 0.001
33
+ },
34
+ "notes": "New refits use float32 training and explicit dense evaluation; original archived scores use original Linear kernels. NLL tolerance accommodates small probability changes in log space; all ranking vectors agree exactly when reported error0.",
35
+ "seconds": 1.074040636071004,
36
+ "device": "cuda:3",
37
+ "gpu": "NVIDIA H200",
38
+ "cpu_scientific_execution": false
39
+ }
revision_r1/sensor/objective_extension.py ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Prospectively amended factor-marginal CE decoder, all original cells."""
2
+ from pathlib import Path
3
+ import argparse, hashlib, importlib.util, json, sys, time
4
+ import torch
5
+
6
+ ROOT=Path(__file__).resolve().parents[3]
7
+ local_trainer=Path(__file__).with_name('review_r1_sensor.py')
8
+ spec=importlib.util.spec_from_file_location('review_sensor',local_trainer if local_trainer.exists() else ROOT/'reproduce/review_r1_sensor.py')
9
+ M=importlib.util.module_from_spec(spec);spec.loader.exec_module(M)
10
+
11
+
12
+ @torch.no_grad()
13
+ def main():
14
+ parser=argparse.ArgumentParser();parser.add_argument('--output',required=True);parser.add_argument('--device',default='cuda:3');parser.add_argument('--release');args=parser.parse_args()
15
+ torch.cuda.set_device(args.device);torch.backends.cuda.matmul.allow_tf32=False;torch.backends.cudnn.allow_tf32=False
16
+ out=Path(args.output);assert not out.exists();out.mkdir();start=time.perf_counter();rows=[];timings=[];summary=[]
17
+ specpath=Path(__file__).with_name('objective_extension_spec.json')
18
+ M.write(out/'manifest.json',dict(timestamp_utc=M.stamp(),command=sys.argv,status='Post-review exploratory objective extension',
19
+ spec_sha256=hashlib.sha256(specpath.read_bytes()).hexdigest(),helper_sha256=hashlib.sha256(Path(__file__).read_bytes()).hexdigest(),
20
+ trainer_sha256=hashlib.sha256((ROOT/'reproduce/review_r1_sensor.py').read_bytes()).hexdigest(),torch_version=torch.__version__,gpu=torch.cuda.get_device_name()))
21
+ for cohort,sizes,seeds in [('hydraulic',[3,4,3,4],[20751,20752,20753]),('temperature',[10,3],[20851,20852,20853])]:
22
+ p=M.load(f'data/{cohort}/prepared',args.device)
23
+ if cohort=='hydraulic':p.update(p['masks'])
24
+ original=M.load(f'results/{cohort}/predictions',args.device)['results'];grid=M.F.hypotheses(sizes,args.device);zg=M.F.design(grid,sizes,True)
25
+ for n in [64,128]:
26
+ for seed in seeds:
27
+ cell_start=time.perf_counter();c=M.load(f'models/{cohort}/n{n}_s{seed}/factorized_neural',args.device);ids=c['fit_ids'];h=p['h'][ids]
28
+ record=next(r for r in original if r['seed']==seed and r['samples']==n and r['model']=='factorized_neural');held=record['held_rows'];true=M.F.joint_index(p['h'][held],sizes)
29
+ x=(p['x'][ids]-c['mean'])/c['scale'];target=(p['x'][held]-c['mean'])/c['scale'];st=M.F.joint_index(h,sizes);cov=torch.eye(x.shape[1],device=args.device,dtype=x.dtype)
30
+ fits=M.fit_cross(x,h,sizes,seed,'forward_factor_ce');fold=torch.arange(n,device=args.device)%4;oof=x.new_empty((n,len(grid)))
31
+ for f in range(4):oof[fold==f]=M.gaussian_scores(x[fold==f],M.dense(fits[f+1]['state'],zg),cov).log_softmax(-1)
32
+ logp=M.gaussian_scores(target,M.dense(fits[0]['state'],zg),cov).log_softmax(-1);calibrated,cal=M.calibrate(oof,st,logp)
33
+ scores={'forward_factor_ce_identity':logp,'forward_factor_ce_identity_scalar_temperature':calibrated};vectors={}
34
+ for name,v in scores.items():
35
+ rr,vectors[name]=M.score_rows(cohort,n,seed,name,'held',v,true,record['unseen'],record.get('days'));rows+=rr
36
+ source=oof if name=='forward_factor_ce_identity' else (oof/cal['temperature']).log_softmax(-1)
37
+ rr,_=M.score_rows(cohort,n,seed,name,'source_oof',source,st,torch.zeros(n,device=args.device,dtype=torch.bool));rows+=rr
38
+ mask=record['unseen'];top=vectors[name]['top3']
39
+ primary=torch.stack([top[mask&(record['days']==d)].mean() for d in (9,10,11,12)]).mean() if cohort=='temperature' else top[mask].mean()
40
+ summary.append(dict(cohort=cohort,labels=n,seed=seed,method=name,top3_primary=float(primary),**M.aggregate(vectors[name],mask),temperature=cal['temperature']))
41
+ artifact=dict(cohort=cohort,labels=n,seed=seed,fit_ids=ids,mean=c['mean'],scale=c['scale'],held_rows=held,true=true,unseen=record['unseen'],units=record.get('days'),fits=fits,held_logpost=scores,source_oof_logpost=oof,calibration=cal,metrics=vectors)
42
+ torch.save(artifact,out/f'{cohort}_n{n}_s{seed}.pt');M.sync()
43
+ # Replay full decoder scores from saved tensors and all per-row metrics.
44
+ saved=torch.load(out/f'{cohort}_n{n}_s{seed}.pt',map_location=args.device,weights_only=True)
45
+ fresh=M.gaussian_scores(target,M.dense(saved['fits'][0]['state'],zg),cov).log_softmax(-1);error=float((fresh-saved['held_logpost']['forward_factor_ce_identity']).abs().max());assert error<1e-8
46
+ timings.append(dict(cohort=cohort,labels=n,seed=seed,training_seconds=sum(f['training_seconds'] for f in fits),cell_seconds=time.perf_counter()-cell_start,
47
+ new_fits=len(fits),trainable_parameters=fits[0]['trainable_parameters'],network_training_executions=sum(f['steps'] for f in fits),estimated_training_dense_matmul_flops=sum(f['estimated_training_dense_matmul_flops'] for f in fits),replay_error=error))
48
+ M.dump_csv(out/'metrics.csv',rows);M.dump_csv(out/'summary_per_seed.csv',summary);M.write(out/'cell_compute.json',timings)
49
+ print(json.dumps(dict(cohort=cohort,labels=n,seed=seed,top3=float(primary),temperature=cal['temperature'],seconds=timings[-1]['cell_seconds'])),flush=True)
50
+ aggregate=[]
51
+ for cohort in ['hydraulic','temperature']:
52
+ for n in [64,128]:
53
+ for name in scores:
54
+ rr=[r for r in summary if r['cohort']==cohort and r['labels']==n and r['method']==name]
55
+ aggregate.append(dict(cohort=cohort,labels=n,method=name,**{k:sum(r[k] for r in rr)/len(rr) if all(r[k] is not None for r in rr) else None for k in ['top3_primary','top1','top3','nll','brier','confidence','ece10','coverage90','accuracy90']},seed_top3=[r['top3_primary'] for r in rr],temperatures=[r['temperature'] for r in rr]))
56
+ M.write(out/'summary.json',aggregate);M.dump_csv(out/'summary.csv',aggregate)
57
+ M.write(out/'compute.json',dict(status='completed',timestamp_utc=M.stamp(),new_fits=60,optimizer_updates=24000,complete_seconds=time.perf_counter()-start,
58
+ training_seconds=sum(t['training_seconds'] for t in timings),estimated_training_dense_matmul_flops=sum(t['estimated_training_dense_matmul_flops'] for t in timings),
59
+ flop_scope='Partial dense matmuls, forward/backward; complete time includes all inference/calibration/serialization/replay',
60
+ peak_cuda_allocated_bytes=torch.cuda.max_memory_allocated(),peak_cuda_reserved_bytes=torch.cuda.max_memory_reserved(),cpu_scientific_execution=False,cpu_offload=False,backbone_executions=0,device=args.device,gpu=torch.cuda.get_device_name()))
61
+
62
+
63
+ if __name__=='__main__':main()
revision_r1/sensor/objective_extension_01/cell_compute.json ADDED
@@ -0,0 +1,146 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "cohort": "hydraulic",
4
+ "labels": 64,
5
+ "seed": 20751,
6
+ "training_seconds": 9.247948644100688,
7
+ "cell_seconds": 11.992396004963666,
8
+ "new_fits": 5,
9
+ "trainable_parameters": 19936,
10
+ "network_training_executions": 2000,
11
+ "estimated_training_dense_matmul_flops": 33868800000,
12
+ "replay_error": 0.0
13
+ },
14
+ {
15
+ "cohort": "hydraulic",
16
+ "labels": 64,
17
+ "seed": 20752,
18
+ "training_seconds": 7.901687476900406,
19
+ "cell_seconds": 7.933070569997653,
20
+ "new_fits": 5,
21
+ "trainable_parameters": 19936,
22
+ "network_training_executions": 2000,
23
+ "estimated_training_dense_matmul_flops": 33868800000,
24
+ "replay_error": 0.0
25
+ },
26
+ {
27
+ "cohort": "hydraulic",
28
+ "labels": 64,
29
+ "seed": 20753,
30
+ "training_seconds": 7.900175572023727,
31
+ "cell_seconds": 7.930946532986127,
32
+ "new_fits": 5,
33
+ "trainable_parameters": 19936,
34
+ "network_training_executions": 2000,
35
+ "estimated_training_dense_matmul_flops": 33868800000,
36
+ "replay_error": 0.0
37
+ },
38
+ {
39
+ "cohort": "hydraulic",
40
+ "labels": 128,
41
+ "seed": 20751,
42
+ "training_seconds": 8.145796927856281,
43
+ "cell_seconds": 8.176515604951419,
44
+ "new_fits": 5,
45
+ "trainable_parameters": 19936,
46
+ "network_training_executions": 2000,
47
+ "estimated_training_dense_matmul_flops": 33868800000,
48
+ "replay_error": 0.0
49
+ },
50
+ {
51
+ "cohort": "hydraulic",
52
+ "labels": 128,
53
+ "seed": 20752,
54
+ "training_seconds": 7.722086073132232,
55
+ "cell_seconds": 10.04799306706991,
56
+ "new_fits": 5,
57
+ "trainable_parameters": 19936,
58
+ "network_training_executions": 2000,
59
+ "estimated_training_dense_matmul_flops": 33868800000,
60
+ "replay_error": 0.0
61
+ },
62
+ {
63
+ "cohort": "hydraulic",
64
+ "labels": 128,
65
+ "seed": 20753,
66
+ "training_seconds": 8.524675879976712,
67
+ "cell_seconds": 8.556502330000512,
68
+ "new_fits": 5,
69
+ "trainable_parameters": 19936,
70
+ "network_training_executions": 2000,
71
+ "estimated_training_dense_matmul_flops": 33868800000,
72
+ "replay_error": 0.0
73
+ },
74
+ {
75
+ "cohort": "temperature",
76
+ "labels": 64,
77
+ "seed": 20851,
78
+ "training_seconds": 7.763742531067692,
79
+ "cell_seconds": 7.835612436989322,
80
+ "new_fits": 5,
81
+ "trainable_parameters": 19936,
82
+ "network_training_executions": 2000,
83
+ "estimated_training_dense_matmul_flops": 7054560000,
84
+ "replay_error": 0.0
85
+ },
86
+ {
87
+ "cohort": "temperature",
88
+ "labels": 64,
89
+ "seed": 20852,
90
+ "training_seconds": 10.738027861109003,
91
+ "cell_seconds": 10.862897809944116,
92
+ "new_fits": 5,
93
+ "trainable_parameters": 19936,
94
+ "network_training_executions": 2000,
95
+ "estimated_training_dense_matmul_flops": 7054560000,
96
+ "replay_error": 0.0
97
+ },
98
+ {
99
+ "cohort": "temperature",
100
+ "labels": 64,
101
+ "seed": 20853,
102
+ "training_seconds": 8.26926949096378,
103
+ "cell_seconds": 8.328956548939459,
104
+ "new_fits": 5,
105
+ "trainable_parameters": 19936,
106
+ "network_training_executions": 2000,
107
+ "estimated_training_dense_matmul_flops": 7054560000,
108
+ "replay_error": 0.0
109
+ },
110
+ {
111
+ "cohort": "temperature",
112
+ "labels": 128,
113
+ "seed": 20851,
114
+ "training_seconds": 7.83897131413687,
115
+ "cell_seconds": 7.901723493938334,
116
+ "new_fits": 5,
117
+ "trainable_parameters": 19936,
118
+ "network_training_executions": 2000,
119
+ "estimated_training_dense_matmul_flops": 7054560000,
120
+ "replay_error": 0.0
121
+ },
122
+ {
123
+ "cohort": "temperature",
124
+ "labels": 128,
125
+ "seed": 20852,
126
+ "training_seconds": 7.855879723909311,
127
+ "cell_seconds": 7.91638747905381,
128
+ "new_fits": 5,
129
+ "trainable_parameters": 19936,
130
+ "network_training_executions": 2000,
131
+ "estimated_training_dense_matmul_flops": 7054560000,
132
+ "replay_error": 0.0
133
+ },
134
+ {
135
+ "cohort": "temperature",
136
+ "labels": 128,
137
+ "seed": 20853,
138
+ "training_seconds": 7.856508697965182,
139
+ "cell_seconds": 7.917562803952023,
140
+ "new_fits": 5,
141
+ "trainable_parameters": 19936,
142
+ "network_training_executions": 2000,
143
+ "estimated_training_dense_matmul_flops": 7054560000,
144
+ "replay_error": 0.0
145
+ }
146
+ ]
revision_r1/sensor/objective_extension_01/compute.json ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "completed",
3
+ "timestamp_utc": "2026-11-09T14:45:55.244783+00:00",
4
+ "new_fits": 60,
5
+ "optimizer_updates": 24000,
6
+ "complete_seconds": 105.47859474900179,
7
+ "training_seconds": 99.76477019314189,
8
+ "estimated_training_dense_matmul_flops": 245540160000,
9
+ "flop_scope": "Partial dense matmuls, forward/backward; complete time includes all inference/calibration/serialization/replay",
10
+ "peak_cuda_allocated_bytes": 229255680,
11
+ "peak_cuda_reserved_bytes": 255852544,
12
+ "cpu_scientific_execution": false,
13
+ "cpu_offload": false,
14
+ "backbone_executions": 0,
15
+ "device": "cuda:3",
16
+ "gpu": "NVIDIA H200"
17
+ }
revision_r1/sensor/objective_extension_01/hydraulic_n128_s20751.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"dict":[["cohort","hydraulic"],["labels",128],["seed",20751],["fit_ids",{"tensor":"t000000"}],["mean",{"tensor":"t000001"}],["scale",{"tensor":"t000002"}],["held_rows",{"tensor":"t000003"}],["true",{"tensor":"t000004"}],["unseen",{"tensor":"t000005"}],["units",null],["fits",{"list":[{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000006"}],["net.0.bias",{"tensor":"t000007"}],["net.2.weight",{"tensor":"t000008"}],["net.2.bias",{"tensor":"t000009"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20751],["training_seconds",1.7130040070042014],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.3153672218322754]]},{"dict":[["update",100],["preupdate_loss",0.11518652737140656]]},{"dict":[["update",200],["preupdate_loss",0.07788994908332825]]},{"dict":[["update",400],["preupdate_loss",0.1391909271478653]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000010"}],["net.0.bias",{"tensor":"t000011"}],["net.2.weight",{"tensor":"t000012"}],["net.2.bias",{"tensor":"t000013"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20852],["training_seconds",1.7215261249803007],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.26554274559021]]},{"dict":[["update",100],["preupdate_loss",0.10422307252883911]]},{"dict":[["update",200],["preupdate_loss",0.060535937547683716]]},{"dict":[["update",400],["preupdate_loss",0.21105968952178955]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000014"}],["net.0.bias",{"tensor":"t000015"}],["net.2.weight",{"tensor":"t000016"}],["net.2.bias",{"tensor":"t000017"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20853],["training_seconds",1.632717469939962],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.3454340696334839]]},{"dict":[["update",100],["preupdate_loss",0.08547832071781158]]},{"dict":[["update",200],["preupdate_loss",0.04600851982831955]]},{"dict":[["update",400],["preupdate_loss",0.03054380789399147]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000018"}],["net.0.bias",{"tensor":"t000019"}],["net.2.weight",{"tensor":"t000020"}],["net.2.bias",{"tensor":"t000021"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20854],["training_seconds",1.564134583924897],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.3180103302001953]]},{"dict":[["update",100],["preupdate_loss",0.10781454294919968]]},{"dict":[["update",200],["preupdate_loss",0.06580552458763123]]},{"dict":[["update",400],["preupdate_loss",0.04327245429158211]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000022"}],["net.0.bias",{"tensor":"t000023"}],["net.2.weight",{"tensor":"t000024"}],["net.2.bias",{"tensor":"t000025"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20855],["training_seconds",1.5144147420069203],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.234365701675415]]},{"dict":[["update",100],["preupdate_loss",0.10966001451015472]]},{"dict":[["update",200],["preupdate_loss",0.058905404061079025]]},{"dict":[["update",400],["preupdate_loss",0.2836851179599762]]}]}]]}]}],["held_logpost",{"dict":[["forward_factor_ce_identity",{"tensor":"t000026"}],["forward_factor_ce_identity_scalar_temperature",{"tensor":"t000027"}]]}],["source_oof_logpost",{"tensor":"t000028"}],["calibration",{"dict":[["temperature",8.0],["grid",{"list":[1.0,2.0,4.0,8.0,16.0,32.0,64.0]}],["oof_nll",{"list":[11.849315593533733,6.273474467157997,3.9106504839236744,3.373548745521792,3.6546327409681982,4.099476012113334,4.459736913706311]}]]}],["metrics",{"dict":[["forward_factor_ce_identity",{"dict":[["top1",{"tensor":"t000029"}],["top3",{"tensor":"t000030"}],["nll",{"tensor":"t000031"}],["confidence",{"tensor":"t000032"}],["brier",{"tensor":"t000033"}]]}],["forward_factor_ce_identity_scalar_temperature",{"dict":[["top1",{"tensor":"t000034"}],["top3",{"tensor":"t000035"}],["nll",{"tensor":"t000036"}],["confidence",{"tensor":"t000037"}],["brier",{"tensor":"t000038"}]]}]]}]]}
revision_r1/sensor/objective_extension_01/hydraulic_n128_s20751.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d7a21511d36785300dd873f79aadcec87bf8de59a73b621328e26cc45a137bda
3
+ size 2258151
revision_r1/sensor/objective_extension_01/hydraulic_n128_s20752.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"dict":[["cohort","hydraulic"],["labels",128],["seed",20752],["fit_ids",{"tensor":"t000000"}],["mean",{"tensor":"t000001"}],["scale",{"tensor":"t000002"}],["held_rows",{"tensor":"t000003"}],["true",{"tensor":"t000004"}],["unseen",{"tensor":"t000005"}],["units",null],["fits",{"list":[{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000006"}],["net.0.bias",{"tensor":"t000007"}],["net.2.weight",{"tensor":"t000008"}],["net.2.bias",{"tensor":"t000009"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20752],["training_seconds",1.521117574069649],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.2658435106277466]]},{"dict":[["update",100],["preupdate_loss",0.10903939604759216]]},{"dict":[["update",200],["preupdate_loss",0.16448232531547546]]},{"dict":[["update",400],["preupdate_loss",0.03899599239230156]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000010"}],["net.0.bias",{"tensor":"t000011"}],["net.2.weight",{"tensor":"t000012"}],["net.2.bias",{"tensor":"t000013"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20853],["training_seconds",1.534898433019407],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.3638601303100586]]},{"dict":[["update",100],["preupdate_loss",0.09481599181890488]]},{"dict":[["update",200],["preupdate_loss",0.04286932945251465]]},{"dict":[["update",400],["preupdate_loss",0.023131797090172768]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000014"}],["net.0.bias",{"tensor":"t000015"}],["net.2.weight",{"tensor":"t000016"}],["net.2.bias",{"tensor":"t000017"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20854],["training_seconds",1.5903978729620576],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.3067092895507812]]},{"dict":[["update",100],["preupdate_loss",0.1004253476858139]]},{"dict":[["update",200],["preupdate_loss",0.27468007802963257]]},{"dict":[["update",400],["preupdate_loss",0.04639362171292305]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000018"}],["net.0.bias",{"tensor":"t000019"}],["net.2.weight",{"tensor":"t000020"}],["net.2.bias",{"tensor":"t000021"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20855],["training_seconds",1.5241462660487741],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.2226519584655762]]},{"dict":[["update",100],["preupdate_loss",0.061646513640880585]]},{"dict":[["update",200],["preupdate_loss",0.028353851288557053]]},{"dict":[["update",400],["preupdate_loss",0.02910551242530346]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000022"}],["net.0.bias",{"tensor":"t000023"}],["net.2.weight",{"tensor":"t000024"}],["net.2.bias",{"tensor":"t000025"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20856],["training_seconds",1.551525927032344],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.2441983222961426]]},{"dict":[["update",100],["preupdate_loss",0.09223012626171112]]},{"dict":[["update",200],["preupdate_loss",0.05816187337040901]]},{"dict":[["update",400],["preupdate_loss",0.04458361864089966]]}]}]]}]}],["held_logpost",{"dict":[["forward_factor_ce_identity",{"tensor":"t000026"}],["forward_factor_ce_identity_scalar_temperature",{"tensor":"t000027"}]]}],["source_oof_logpost",{"tensor":"t000028"}],["calibration",{"dict":[["temperature",4.0],["grid",{"list":[1.0,2.0,4.0,8.0,16.0,32.0,64.0]}],["oof_nll",{"list":[6.900842189252238,3.717516208510413,2.527189305081876,2.615379022527092,3.2545442193886887,3.8928117248978285,4.353807781944336]}]]}],["metrics",{"dict":[["forward_factor_ce_identity",{"dict":[["top1",{"tensor":"t000029"}],["top3",{"tensor":"t000030"}],["nll",{"tensor":"t000031"}],["confidence",{"tensor":"t000032"}],["brier",{"tensor":"t000033"}]]}],["forward_factor_ce_identity_scalar_temperature",{"dict":[["top1",{"tensor":"t000034"}],["top3",{"tensor":"t000035"}],["nll",{"tensor":"t000036"}],["confidence",{"tensor":"t000037"}],["brier",{"tensor":"t000038"}]]}]]}]]}
revision_r1/sensor/objective_extension_01/hydraulic_n128_s20752.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:84e37c6d7a7ec02217682fedad56211ef67be4757a0376682fc6428f70f92f61
3
+ size 2258151
revision_r1/sensor/objective_extension_01/hydraulic_n128_s20753.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"dict":[["cohort","hydraulic"],["labels",128],["seed",20753],["fit_ids",{"tensor":"t000000"}],["mean",{"tensor":"t000001"}],["scale",{"tensor":"t000002"}],["held_rows",{"tensor":"t000003"}],["true",{"tensor":"t000004"}],["unseen",{"tensor":"t000005"}],["units",null],["fits",{"list":[{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000006"}],["net.0.bias",{"tensor":"t000007"}],["net.2.weight",{"tensor":"t000008"}],["net.2.bias",{"tensor":"t000009"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20753],["training_seconds",1.7182763000018895],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.3494446277618408]]},{"dict":[["update",100],["preupdate_loss",0.09609153866767883]]},{"dict":[["update",200],["preupdate_loss",0.04170093685388565]]},{"dict":[["update",400],["preupdate_loss",0.11369231343269348]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000010"}],["net.0.bias",{"tensor":"t000011"}],["net.2.weight",{"tensor":"t000012"}],["net.2.bias",{"tensor":"t000013"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20854],["training_seconds",1.6955450871028006],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.3265430927276611]]},{"dict":[["update",100],["preupdate_loss",0.08156754076480865]]},{"dict":[["update",200],["preupdate_loss",0.025151949375867844]]},{"dict":[["update",400],["preupdate_loss",0.03886566683650017]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000014"}],["net.0.bias",{"tensor":"t000015"}],["net.2.weight",{"tensor":"t000016"}],["net.2.bias",{"tensor":"t000017"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20855],["training_seconds",1.7018361849477515],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.2241082191467285]]},{"dict":[["update",100],["preupdate_loss",0.0709477961063385]]},{"dict":[["update",200],["preupdate_loss",0.03540191799402237]]},{"dict":[["update",400],["preupdate_loss",0.02366386540234089]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000018"}],["net.0.bias",{"tensor":"t000019"}],["net.2.weight",{"tensor":"t000020"}],["net.2.bias",{"tensor":"t000021"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20856],["training_seconds",1.7114315149374306],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.2223014831542969]]},{"dict":[["update",100],["preupdate_loss",0.09891912341117859]]},{"dict":[["update",200],["preupdate_loss",0.05410601198673248]]},{"dict":[["update",400],["preupdate_loss",0.1466662436723709]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000022"}],["net.0.bias",{"tensor":"t000023"}],["net.2.weight",{"tensor":"t000024"}],["net.2.bias",{"tensor":"t000025"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20857],["training_seconds",1.69758679298684],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.3295478820800781]]},{"dict":[["update",100],["preupdate_loss",0.08611084520816803]]},{"dict":[["update",200],["preupdate_loss",0.03314342349767685]]},{"dict":[["update",400],["preupdate_loss",0.027886219322681427]]}]}]]}]}],["held_logpost",{"dict":[["forward_factor_ce_identity",{"tensor":"t000026"}],["forward_factor_ce_identity_scalar_temperature",{"tensor":"t000027"}]]}],["source_oof_logpost",{"tensor":"t000028"}],["calibration",{"dict":[["temperature",4.0],["grid",{"list":[1.0,2.0,4.0,8.0,16.0,32.0,64.0]}],["oof_nll",{"list":[5.086095402329374,2.831339798198033,2.1274483841182823,2.397373893919174,3.089160951570917,3.771096803890446,4.2730538150441015]}]]}],["metrics",{"dict":[["forward_factor_ce_identity",{"dict":[["top1",{"tensor":"t000029"}],["top3",{"tensor":"t000030"}],["nll",{"tensor":"t000031"}],["confidence",{"tensor":"t000032"}],["brier",{"tensor":"t000033"}]]}],["forward_factor_ce_identity_scalar_temperature",{"dict":[["top1",{"tensor":"t000034"}],["top3",{"tensor":"t000035"}],["nll",{"tensor":"t000036"}],["confidence",{"tensor":"t000037"}],["brier",{"tensor":"t000038"}]]}]]}]]}
revision_r1/sensor/objective_extension_01/hydraulic_n128_s20753.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fb1eb8085bbd37b334f4a7033d7b0ca0863981f0263528ccced9ada34167dace
3
+ size 2258151
revision_r1/sensor/objective_extension_01/hydraulic_n64_s20751.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"dict":[["cohort","hydraulic"],["labels",64],["seed",20751],["fit_ids",{"tensor":"t000000"}],["mean",{"tensor":"t000001"}],["scale",{"tensor":"t000002"}],["held_rows",{"tensor":"t000003"}],["true",{"tensor":"t000004"}],["unseen",{"tensor":"t000005"}],["units",null],["fits",{"list":[{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000006"}],["net.0.bias",{"tensor":"t000007"}],["net.2.weight",{"tensor":"t000008"}],["net.2.bias",{"tensor":"t000009"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20751],["training_seconds",2.5850065340055153],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.3451077938079834]]},{"dict":[["update",100],["preupdate_loss",0.055335015058517456]]},{"dict":[["update",200],["preupdate_loss",0.009100543335080147]]},{"dict":[["update",400],["preupdate_loss",0.0018230908317491412]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000010"}],["net.0.bias",{"tensor":"t000011"}],["net.2.weight",{"tensor":"t000012"}],["net.2.bias",{"tensor":"t000013"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20852],["training_seconds",1.6536050019785762],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.2514609098434448]]},{"dict":[["update",100],["preupdate_loss",0.049963369965553284]]},{"dict":[["update",200],["preupdate_loss",0.007042030803859234]]},{"dict":[["update",400],["preupdate_loss",0.0014284837525337934]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000014"}],["net.0.bias",{"tensor":"t000015"}],["net.2.weight",{"tensor":"t000016"}],["net.2.bias",{"tensor":"t000017"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20853],["training_seconds",1.6466316950973123],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.3599019050598145]]},{"dict":[["update",100],["preupdate_loss",0.03786250203847885]]},{"dict":[["update",200],["preupdate_loss",0.0057496195659041405]]},{"dict":[["update",400],["preupdate_loss",0.0012487568892538548]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000018"}],["net.0.bias",{"tensor":"t000019"}],["net.2.weight",{"tensor":"t000020"}],["net.2.bias",{"tensor":"t000021"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20854],["training_seconds",1.6881663490785286],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.3245327472686768]]},{"dict":[["update",100],["preupdate_loss",0.04571075364947319]]},{"dict":[["update",200],["preupdate_loss",0.006450130604207516]]},{"dict":[["update",400],["preupdate_loss",0.0015182875795289874]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000022"}],["net.0.bias",{"tensor":"t000023"}],["net.2.weight",{"tensor":"t000024"}],["net.2.bias",{"tensor":"t000025"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20855],["training_seconds",1.674539063940756],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.220789909362793]]},{"dict":[["update",100],["preupdate_loss",0.07876121997833252]]},{"dict":[["update",200],["preupdate_loss",0.04276982694864273]]},{"dict":[["update",400],["preupdate_loss",0.03329639509320259]]}]}]]}]}],["held_logpost",{"dict":[["forward_factor_ce_identity",{"tensor":"t000026"}],["forward_factor_ce_identity_scalar_temperature",{"tensor":"t000027"}]]}],["source_oof_logpost",{"tensor":"t000028"}],["calibration",{"dict":[["temperature",4.0],["grid",{"list":[1.0,2.0,4.0,8.0,16.0,32.0,64.0]}],["oof_nll",{"list":[6.792243070171102,3.5698963043271705,2.327733437548014,2.3977731484902653,3.0476395521039077,3.7180662668380786,4.225876077364923]}]]}],["metrics",{"dict":[["forward_factor_ce_identity",{"dict":[["top1",{"tensor":"t000029"}],["top3",{"tensor":"t000030"}],["nll",{"tensor":"t000031"}],["confidence",{"tensor":"t000032"}],["brier",{"tensor":"t000033"}]]}],["forward_factor_ce_identity_scalar_temperature",{"dict":[["top1",{"tensor":"t000034"}],["top3",{"tensor":"t000035"}],["nll",{"tensor":"t000036"}],["confidence",{"tensor":"t000037"}],["brier",{"tensor":"t000038"}]]}]]}]]}
revision_r1/sensor/objective_extension_01/hydraulic_n64_s20751.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3e5a56962a870d83d2cd4037f8ec1c0aa39fe0b20111579bda7d435a98683f28
3
+ size 2183911
revision_r1/sensor/objective_extension_01/hydraulic_n64_s20752.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"dict":[["cohort","hydraulic"],["labels",64],["seed",20752],["fit_ids",{"tensor":"t000000"}],["mean",{"tensor":"t000001"}],["scale",{"tensor":"t000002"}],["held_rows",{"tensor":"t000003"}],["true",{"tensor":"t000004"}],["unseen",{"tensor":"t000005"}],["units",null],["fits",{"list":[{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000006"}],["net.0.bias",{"tensor":"t000007"}],["net.2.weight",{"tensor":"t000008"}],["net.2.bias",{"tensor":"t000009"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20752],["training_seconds",1.7232382470974699],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.2678219079971313]]},{"dict":[["update",100],["preupdate_loss",0.0787072479724884]]},{"dict":[["update",200],["preupdate_loss",0.056699153035879135]]},{"dict":[["update",400],["preupdate_loss",0.7450239658355713]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000010"}],["net.0.bias",{"tensor":"t000011"}],["net.2.weight",{"tensor":"t000012"}],["net.2.bias",{"tensor":"t000013"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20853],["training_seconds",1.6219279859215021],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.383439302444458]]},{"dict":[["update",100],["preupdate_loss",0.05290248245000839]]},{"dict":[["update",200],["preupdate_loss",0.022894427180290222]]},{"dict":[["update",400],["preupdate_loss",0.017069250345230103]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000014"}],["net.0.bias",{"tensor":"t000015"}],["net.2.weight",{"tensor":"t000016"}],["net.2.bias",{"tensor":"t000017"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20854],["training_seconds",1.518231512978673],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.2927300930023193]]},{"dict":[["update",100],["preupdate_loss",0.07159949839115143]]},{"dict":[["update",200],["preupdate_loss",0.026884913444519043]]},{"dict":[["update",400],["preupdate_loss",0.018253501504659653]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000018"}],["net.0.bias",{"tensor":"t000019"}],["net.2.weight",{"tensor":"t000020"}],["net.2.bias",{"tensor":"t000021"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20855],["training_seconds",1.5201711799018085],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.2043473720550537]]},{"dict":[["update",100],["preupdate_loss",0.040777452290058136]]},{"dict":[["update",200],["preupdate_loss",0.02084321342408657]]},{"dict":[["update",400],["preupdate_loss",0.017022639513015747]]}]}]]},{"dict":[["state",{"dict":[["net.0.weight",{"tensor":"t000022"}],["net.0.bias",{"tensor":"t000023"}],["net.2.weight",{"tensor":"t000024"}],["net.2.bias",{"tensor":"t000025"}]]}],["trainable_parameters",19936],["steps",400],["kind","forward_factor_ce"],["seed",20856],["training_seconds",1.5181185510009527],["training_network_executions",400],["estimated_training_dense_matmul_flops",6773760000],["loss_trace",{"list":[{"dict":[["update",1],["preupdate_loss",1.2638598680496216]]},{"dict":[["update",100],["preupdate_loss",0.0664515271782875]]},{"dict":[["update",200],["preupdate_loss",0.038003109395504]]},{"dict":[["update",400],["preupdate_loss",0.03006787970662117]]}]}]]}]}],["held_logpost",{"dict":[["forward_factor_ce_identity",{"tensor":"t000026"}],["forward_factor_ce_identity_scalar_temperature",{"tensor":"t000027"}]]}],["source_oof_logpost",{"tensor":"t000028"}],["calibration",{"dict":[["temperature",8.0],["grid",{"list":[1.0,2.0,4.0,8.0,16.0,32.0,64.0]}],["oof_nll",{"list":[9.869375650516783,5.1392550287919665,3.160878618792568,2.883185872356636,3.3682674820768237,3.9486493509371563,4.385519147434494]}]]}],["metrics",{"dict":[["forward_factor_ce_identity",{"dict":[["top1",{"tensor":"t000029"}],["top3",{"tensor":"t000030"}],["nll",{"tensor":"t000031"}],["confidence",{"tensor":"t000032"}],["brier",{"tensor":"t000033"}]]}],["forward_factor_ce_identity_scalar_temperature",{"dict":[["top1",{"tensor":"t000034"}],["top3",{"tensor":"t000035"}],["nll",{"tensor":"t000036"}],["confidence",{"tensor":"t000037"}],["brier",{"tensor":"t000038"}]]}]]}]]}
revision_r1/sensor/objective_extension_01/hydraulic_n64_s20752.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e4ff49f92f19eeb70adf1a6c200eb238839153310e435c938951a6658adfbc7c
3
+ size 2183911