nur-dev's picture
Add reviewer-requested controlled and sensor mechanism tests
e7a0400 verified
Raw History Blame Contribute Delete
4.67 kB
{
"material_passport": {
"material_type": "post-review exploratory experiment specification",
"provenance": "Previously opened external sensor archives and original eligible task manifest",
"license": "Hydraulic UCI CC BY 4.0; gas legacy research-only restriction retained; local research use only",
"created_utc": "2026-11-09T14:37:09.140791+00:00",
"scientific_status": "exploratory; does not replace original confirmation"
},
"purpose": "Distinguish small within-fitted-law encoding headroom, finite labeled support, and observed source-target transfer failure without claiming an identifiable population decomposition",
"device": "cuda:4",
"dtype": "float64",
"backbone_executions": 0,
"trainable_parameters": 0,
"task_indices": [
0,
1,
2,
3,
4,
5,
6,
7,
8,
9,
10,
13,
14,
15,
17,
18,
20,
23,
24,
25,
26,
28,
29,
30,
32,
33,
34,
35,
37,
38,
40,
43,
44,
45,
46,
48,
49,
50,
52,
53,
55,
58,
59,
60,
61,
63,
64,
65,
67,
68,
70,
73,
75,
78,
79,
80,
83
],
"seeds": [
20111,
20112,
20113
],
"methods": [
"query_joint",
"posterior_clusters",
"identity"
],
"estimators": [
"support_mixture",
"joint_jeffreys",
"naive_bayes"
],
"interventions": [
"Original target64 exact seed selection, nested target256 when available, and all original target adaptation records",
"Source group-disjoint train/held split: SHA256('review_r1_source_held:' + original_group) mod 5 == 0 held; source64/256/all train only",
"Each fixed policy replayed on original target evaluation and, for source-trained policies, source held records",
"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"
],
"source_preprocessing": "Original source-only binary thresholds kept; source-held records contributed unlabeled preprocessing before this post-review split, never labeled model fitting",
"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"
],
"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.",
"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.",
"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"
],
"input_hashes": {
"artifacts/external_decision_sufficiency_01/prepared/manifest.json": "c97a59bbf5f04fe8dc3ccfe00986b4df288b8e7ef2f3fbd9eb12f3636dceef5d",
"articles/make_1/release/huggingface-revised/protocols/external_sufficiency_protocol_01.json": "e0229ef8a50270b7a590c6ed79a850ec8f902d17f2c8213111bf432669ef9d04",
"articles/make_1/release/huggingface-revised/unionmind/decision_sufficiency/core.py": "fa2f7405a8d0dc6ec96ebae1bb2a2b810d03ef1fd2877f5ebef7df1722992cee",
"articles/make_1/release/huggingface-revised/unionmind/external_sufficiency/runtime.py": "fa19ce1b7fba9e6953e9e4081d4b6001ed55c33edc48a05960d5a13ae6aad61d"
},
"script_sha256": "18a499610df5927f78ae97af5bd4249b3c18beaef98ea39caecc5ad1303b7956",
"no_recipe_selection": true,
"original_results_and_gates_unchanged": true
}