mtastan commited on
Commit
b5fb06d
·
verified ·
1 Parent(s): 731bfba

Make Hugging Face release checkpoint-only; point code users to GitHub

Browse files
README.md CHANGED
@@ -4,56 +4,22 @@ library_name: pytorch
4
  tags:
5
  - video-anomaly-detection
6
  - weakly-supervised-learning
7
- - transfer-learning
8
  ---
9
- # PACT-VAD: controlled multi-target temporal transfer
10
 
11
- This repository distributes the **180 original validation-selected target checkpoints**
12
- for three datasets, six conditions (B, R, P, PW, PLR, PTSA), and ten seeds.
13
- The checkpoint manifest binds every file to its stored canonical SHA-256 and byte size.
14
- It is a separate release from the historical five-seed `mtastan/PATT-Net` repository.
15
 
16
- Selection uses the earliest epoch attaining maximum validation video AUC. Primary
17
- test metrics are frame ROC-AUC for UCF-Crime and ShanghaiTech and frame AP for
18
- XD-Violence. The stored UCF evaluator covers **1,109,504 frames**, not 1,109,888.
19
- The evidence JSON supplies metrics and selected epochs; no new training was performed
20
- to produce this release. Original PyTorch serialization is preserved to retain hashes.
21
 
22
- ## Interpretation
 
 
23
 
24
- This is a controlled exploratory transfer study, not a SOTA or universal-benefit claim.
25
- Only ShanghaiTech PW minus P survives the declared five-contrast within-dataset Holm
26
- adjustment: +3.435 AUC points, 95% paired-bootstrap interval [2.449,4.477], adjusted
27
- p=0.0098, positive in all ten seeds. Negative and uncertain conditions are retained.
28
- Historical test-epoch maxima are not primary results. Test labels do not select these
29
- checkpoints or enter the training objective.
30
 
31
- ## Use and limits
32
 
33
- Model/evaluation code and split contracts are supplied in `code/pact_vad_code.zip`
34
- with a per-file manifest. Architecture uses frozen CLIP
35
- ViT-L/14 768-dimensional descriptors and a temporal receiver. Loading weights alone
36
- does not reproduce features, frame alignment, or metrics. Historical feature archives
37
- and raw frame-score vectors are not part of this release. Ground-truth arrays must be
38
- retrieved from their original distributors under applicable terms.
39
-
40
- **Compact checkpoints:** 90 P/PW/PTSA files deliberately omit the frozen source.
41
- Retrieve `source/prevad_temporal_e02.pth` from this release and use the strict loader
42
- in `src/released_checkpoint.py`; `code/README.md` provides an example. All 180 files
43
- passed strict CPU state loading and a synthetic forward pass. This is not dataset
44
- metric reevaluation.
45
-
46
- Use `huggingface_hub.hf_hub_download` with the manifest path and pin the repository
47
- revision. Check file SHA-256 against `checkpoint_manifest.json` before loading.
48
- Load only trusted weights on CPU using PyTorch's `weights_only=True` option where
49
- supported; instantiate the condition-specific architecture from the project source.
50
- The release is research material, not a validated operational surveillance system.
51
-
52
- ## Contributions
53
-
54
- Hamza Osman İlhan (Associate Professor, PhD): Conceptualization, Methodology,
55
- Supervision; technical guidance and architectural ideas.
56
- Mehmet Taştan (master's student): Software, Validation; implementation and testing.
57
- These contribution roles were provided by the project author. Manuscript funding,
58
- competing interests, ethics, final author approval and publication readiness remain
59
- separate declarations and are not certified by this checkpoint release.
 
4
  tags:
5
  - video-anomaly-detection
6
  - weakly-supervised-learning
7
+ - checkpoints
8
  ---
 
9
 
10
+ # PACT-VAD checkpoints
 
 
 
11
 
12
+ This repository distributes the validation-selected PACT-VAD model weights used
13
+ in the accompanying controlled transfer study.
 
 
 
14
 
15
+ - Target checkpoints: 3 datasets × 6 conditions × 10 seeds = 180 files
16
+ - Source checkpoint: the fixed PreVAD temporal backbone used by the transfer runs
17
+ - Selection rule: earliest epoch attaining maximum validation video AUC
18
 
19
+ Source code, configurations, the canonical metric ledger, and manuscript sources:
 
 
 
 
 
20
 
21
+ https://github.com/tastangh/PACT-VAD
22
 
23
+ The checkpoint manifest records the dataset, condition, seed, selected epoch,
24
+ file size, and SHA-256 digest for every target checkpoint. Dataset videos and
25
+ historical feature arrays are not distributed here.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
code/README.md DELETED
@@ -1,58 +0,0 @@
1
- # PACT-VAD reproducibility code snapshot
2
-
3
- This is a file-hash-bound snapshot, not a claim of bit-identical historical training.
4
- It contains model/evaluator source, configurations, splits, statistical analysis and
5
- CPU tests. Raw videos, feature arrays and third-party annotation datasets are excluded.
6
- Checkpoint metrics remain the stored observations, not a new evaluation.
7
-
8
- ## CPU checks
9
-
10
- ```powershell
11
- python -m venv .venv
12
- .venv\Scripts\python -m pip install -r requirements-dev.txt
13
- .venv\Scripts\python -m pytest -q
14
- .venv\Scripts\python tools/analyze_pact_10seed.py
15
- .venv\Scripts\python tools/fetch_paa_ground_truth.py
16
- ```
17
-
18
- The legacy checkpoint manifest is included for its contract test; absent legacy
19
- weights are skipped. Tests requiring undistributed PreVAD annotations and the obsolete
20
- five-seed manuscript are excluded from this snapshot. Canonical weights are described
21
- by the Hub's root `checkpoint_manifest.json`.
22
- GT downloader pins author-hosted arrays and verifies historical hashes; no upstream
23
- redistribution license is implied. Feature extraction/training are optional expensive
24
- operations and were not run when preparing this release.
25
-
26
- ## Load canonical checkpoint files
27
-
28
- Ninety P/PW/PTSA files omit the frozen source to save space. They require the exact
29
- `source/prevad_temporal_e02.pth` distributed in the same model repository. Ordinary
30
- strict loading of the compact state alone fails; this is not corruption.
31
-
32
- ```python
33
- import json, sys
34
- from huggingface_hub import hf_hub_download
35
- sys.path.insert(0, 'src')
36
- from released_checkpoint import load_released_checkpoint
37
- repo = 'mtastan/PACT-VAD'
38
- revision = '<PINNED_RELEASE_COMMIT>'
39
- manifest = json.load(open(hf_hub_download(repo, 'checkpoint_manifest.json', revision=revision)))
40
- entry = next(r for r in manifest['checkpoints']
41
- if r['dataset'] == 'ShanghaiTech' and r['condition'] == 'PW' and r['seed'] == 1)
42
- target = hf_hub_download(repo, entry['path_in_repo'], revision=revision)
43
- source = hf_hub_download(repo, 'source/prevad_temporal_e02.pth', revision=revision)
44
- cfg = json.load(open('configs/prevad_adaptation_shanghaitech_10seed_v1.json'))['model']
45
- model, metadata = load_released_checkpoint(target, cfg, entry['condition'],
46
- entry['checkpoint_sha256'], source,
47
- 'f3c3ecd73a046beab6bdf4fcff6c2850c0428afa3361236d20ed21ea91d9473a')
48
- ```
49
-
50
- The loader checks hashes before safe CPU deserialization, reconstructs only the
51
- explicitly omitted source, and enforces exact model keys/shapes. It never reads test
52
- labels. All 180 original checkpoint states passed strict loading and a four-snippet
53
- synthetic CPU forward; that check does not recompute dataset AUC/AP.
54
-
55
- `evaluate.py` and `evaluate_multitarget.py` retain the historical B/R/P run-directory
56
- interfaces. They are not six-condition Hub download runners. Use the strict loader
57
- for PW/PLR/PTSA, preserve the exact dataset manifests and evaluator contracts, and do
58
- not infer metric reproducibility without the historical features.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
code/manifest.json DELETED
@@ -1,292 +0,0 @@
1
- {
2
- "files": [
3
- {
4
- "path": "LICENSE",
5
- "size_bytes": 1091,
6
- "sha256": "6a4c211e5d77ccbdded31921170bd8e0c1eea6f64e19caab19a7fab986be404c"
7
- },
8
- {
9
- "path": "README.md",
10
- "size_bytes": 2966,
11
- "sha256": "b15665d8e69627135f5cdd6b5843670c8917b4fce6e0c9e4430c126a373385a5"
12
- },
13
- {
14
- "path": "checkpoints/manifest.json",
15
- "size_bytes": 8157,
16
- "sha256": "7782117ae19ab1a81465a3265970a9e3b57eecd27e48f2c4f05a3d6b9a339646"
17
- },
18
- {
19
- "path": "configs/b_extension_shanghaitech_seeds6_10_v1.json",
20
- "size_bytes": 995,
21
- "sha256": "3a5abe0f826ee7c742795728dc7ad1185f393454941ebdb6e37411356ce37364"
22
- },
23
- {
24
- "path": "configs/b_extension_ucf_seeds6_10_v1.json",
25
- "size_bytes": 1066,
26
- "sha256": "e558ed7575f436b67d5d115dd4982c42ed36208cee1cbbc8f0c0a4738a36392f"
27
- },
28
- {
29
- "path": "configs/b_extension_xd_violence_seeds6_10_v1.json",
30
- "size_bytes": 994,
31
- "sha256": "4db58c6ce936eddc4f803b0850d539d5f8ca18e5cf56ecca8983e130c96f7bed"
32
- },
33
- {
34
- "path": "configs/prevad_adaptation_shanghaitech_10seed_v1.json",
35
- "size_bytes": 1043,
36
- "sha256": "430969d485aa139a8f465cf863bdc1575958b9813f569c39407c9470b9d56257"
37
- },
38
- {
39
- "path": "configs/prevad_adaptation_ucf_10seed_v1.json",
40
- "size_bytes": 980,
41
- "sha256": "90ad9e6f37c8082712b7ec03b2d7a26bd0076e5f1e9c5589b6cd57664f0fa281"
42
- },
43
- {
44
- "path": "configs/prevad_adaptation_xd_violence_10seed_v1.json",
45
- "size_bytes": 1043,
46
- "sha256": "0ddbf4fa190f5b6bbd8ed4f6a9fed03b2cacad01e5e2d238578589c967f48922"
47
- },
48
- {
49
- "path": "configs/publication_vitl14.json",
50
- "size_bytes": 1174,
51
- "sha256": "d1ef9b86f2b70935a2db9b19ebb022da60dcbaf16705b05017ff9f843b0dfed0"
52
- },
53
- {
54
- "path": "configs/r_extension_shanghaitech_seeds6_10_v1.json",
55
- "size_bytes": 995,
56
- "sha256": "9d1a933c625acce1dbd81cc70c9a56fd80850b322bf4fd92c84f05af41d4bd40"
57
- },
58
- {
59
- "path": "configs/r_extension_ucf_seeds6_10_v1.json",
60
- "size_bytes": 1066,
61
- "sha256": "c972d8c1f596e2c7ad541775ebd41992b5a5a785747a62c87a76d4a62079d4a6"
62
- },
63
- {
64
- "path": "configs/r_extension_xd_violence_seeds6_10_v1.json",
65
- "size_bytes": 994,
66
- "sha256": "3027a3d57f644d3290368a6401bbaff3e1ec3d739e9f97107fe98653cf9deb69"
67
- },
68
- {
69
- "path": "configs/shanghaitech_vitl14.json",
70
- "size_bytes": 1052,
71
- "sha256": "b16f03ecc55e03ffe716ab3a933642de51c442874d1c9715130acdb1c5531d11"
72
- },
73
- {
74
- "path": "configs/xd_violence_vitl14.json",
75
- "size_bytes": 946,
76
- "sha256": "574cf18860008b79dfa9575e3257ae4e072bea6ab23e42b291afaac8d3ff9b30"
77
- },
78
- {
79
- "path": "data/splits/prevad_train_val_duplicate_exclusions.json",
80
- "size_bytes": 2011,
81
- "sha256": "ca33b736ecdbee168f434eb4c18d760337e41d50b624d61f4d883036d829491c"
82
- },
83
- {
84
- "path": "data/splits/shanghaitech_v1/feature_lengths.json",
85
- "size_bytes": 32947,
86
- "sha256": "8c2f8013702b04eed385169453dacfd5b1c794fa8a2fcfc3fc14bbea7812b8b9"
87
- },
88
- {
89
- "path": "data/splits/shanghaitech_v1/protocol.json",
90
- "size_bytes": 998,
91
- "sha256": "f62cc36f73cf007c24ac8256227a77ca5bf8967acc97d2e55fd8511a00ff36d2"
92
- },
93
- {
94
- "path": "data/splits/shanghaitech_v1/test.json",
95
- "size_bytes": 22180,
96
- "sha256": "b6ee9b19a4eb53758462dfba5dfcfe5a724a02cf5c5bc8d979d7e2bdfab1b582"
97
- },
98
- {
99
- "path": "data/splits/shanghaitech_v1/train.json",
100
- "size_bytes": 21132,
101
- "sha256": "c8b02c3237c5623f693cfd15eb3a1fdc366ea25e6070785f863e144eb71dfd71"
102
- },
103
- {
104
- "path": "data/splits/shanghaitech_v1/val.json",
105
- "size_bytes": 5370,
106
- "sha256": "a4d64274a1a4ab2303e96f0063f3266026c59209e8467bc96a7e23a135a57801"
107
- },
108
- {
109
- "path": "data/splits/ucf_crime_unbiased/test.json",
110
- "size_bytes": 8965,
111
- "sha256": "69b429a2e493296c0978b86c6a5d4c9e840c67688f953d2aaff6a6845d2b7a37"
112
- },
113
- {
114
- "path": "data/splits/ucf_crime_unbiased/train.json",
115
- "size_bytes": 37168,
116
- "sha256": "1abae4be6cdcd22be78aad5893961286bfdaa5b1b3e95d323427d0b1c8eb216f"
117
- },
118
- {
119
- "path": "data/splits/ucf_crime_unbiased/val.json",
120
- "size_bytes": 9713,
121
- "sha256": "2a6397b609ca65e49fa89201822fd9a732cb81e96fad1d416e5a5d66a9cae038"
122
- },
123
- {
124
- "path": "data/splits/xd_violence_v1/feature_lengths.json",
125
- "size_bytes": 520976,
126
- "sha256": "1c5aa251e004f384f24601c89f8b21999c535c86dcddce7bcafacd9ea50e22de"
127
- },
128
- {
129
- "path": "data/splits/xd_violence_v1/protocol.json",
130
- "size_bytes": 1146,
131
- "sha256": "d2744ce5939d6afd9737479339862b43fb422c2dfec2ce3276a52ac9d3f43656"
132
- },
133
- {
134
- "path": "data/splits/xd_violence_v1/test.json",
135
- "size_bytes": 146389,
136
- "sha256": "cc4a3e96b64162aeb9462d0fde4e5a0279f7202a39fdcfe44fe83352cd1a54dd"
137
- },
138
- {
139
- "path": "data/splits/xd_violence_v1/train.json",
140
- "size_bytes": 570106,
141
- "sha256": "ef3c628dd01c55ee7815a2552c7d76a8f2023ed73a1d69de93814ecbb0712964"
142
- },
143
- {
144
- "path": "data/splits/xd_violence_v1/val.json",
145
- "size_bytes": 143465,
146
- "sha256": "35457e59c9cbd872330837f5e736f1d0d2568329ab4a1e605e12e5e754ee62aa"
147
- },
148
- {
149
- "path": "requirements-dev.txt",
150
- "size_bytes": 55,
151
- "sha256": "68274312f8393efd3885af6e426b5e2736b58b218d4f757db9f3af445dc42b4c"
152
- },
153
- {
154
- "path": "requirements-extraction.txt",
155
- "size_bytes": 155,
156
- "sha256": "d1507f3ce7555b7c773355156125f8cc1a68d0f919520839a7f48ce9ae6a9f19"
157
- },
158
- {
159
- "path": "requirements.txt",
160
- "size_bytes": 105,
161
- "sha256": "a0acc602f8811ca49f1875bb3d31073df3cc65ddccaf04ea6e5b0e0179f75482"
162
- },
163
- {
164
- "path": "results/evaluations/pact_10seed_selected_runs.json",
165
- "size_bytes": 126005,
166
- "sha256": "31c1c5be3c338ea53426d7d5b6852db6759aa2fe6f4936903b3cf6eef8633fb7"
167
- },
168
- {
169
- "path": "src/data.py",
170
- "size_bytes": 13853,
171
- "sha256": "27c7e93481e1b5ca59be3a0de18455d1f410770ac8627c41e9ce17f64c855982"
172
- },
173
- {
174
- "path": "src/metrics.py",
175
- "size_bytes": 6953,
176
- "sha256": "1004c08c0de7dd03c9255fc9d3aefb6c77434b2d322494a6ad6642a9d1127dd8"
177
- },
178
- {
179
- "path": "src/model.py",
180
- "size_bytes": 11347,
181
- "sha256": "ab11dec3af5c4aeb868ac8c0cfa0fbacb45b8171c65a795118694c60c4452dee"
182
- },
183
- {
184
- "path": "src/released_checkpoint.py",
185
- "size_bytes": 1732,
186
- "sha256": "7cc54c1df858e2faa6a778ea50dc5edbd0e6a6579bdb6b80685ada6004c4836a"
187
- },
188
- {
189
- "path": "src/utils.py",
190
- "size_bytes": 866,
191
- "sha256": "b7fa9c3a024012fd050b1bacadcdc94a064e24d90f753188dbe7d7ff796fd864"
192
- },
193
- {
194
- "path": "tests/test_checkpoint_manifest.py",
195
- "size_bytes": 2543,
196
- "sha256": "d80559c3bf6134ed28aa657c5f39c5f68412c18d23dce92d147273f2f3cfb601"
197
- },
198
- {
199
- "path": "tests/test_model.py",
200
- "size_bytes": 5416,
201
- "sha256": "40f8ebba24e0d709b46ff901555b98735193a6d05afe8c37b889f4189b35031b"
202
- },
203
- {
204
- "path": "tests/test_multitarget_extraction.py",
205
- "size_bytes": 2827,
206
- "sha256": "466657837b1414e824bbdecd9341d14d7d25a368ab372491bd6bee9a86bee883"
207
- },
208
- {
209
- "path": "tests/test_multitarget_metrics.py",
210
- "size_bytes": 1968,
211
- "sha256": "06bf754499a076b5b1f4a435f05736d3f29d7b400bdf0fc35ad94d99720c6135"
212
- },
213
- {
214
- "path": "tests/test_multitarget_statistics.py",
215
- "size_bytes": 1171,
216
- "sha256": "1353ef1e489d3809a36ffcc4d1f4d5803652b49daf9bb11f6e8f45c9969dd5cc"
217
- },
218
- {
219
- "path": "tests/test_paper_statistics.py",
220
- "size_bytes": 879,
221
- "sha256": "cdcb9180037e708280945c37bda3e1e6c2dee7d265769ed97f525040f3cb1950"
222
- },
223
- {
224
- "path": "tests/test_source_objective.py",
225
- "size_bytes": 991,
226
- "sha256": "d9552b2537f512005ee8b6add3c2f4f25eec050db124076b13130d5c9bd70c04"
227
- },
228
- {
229
- "path": "tools/analyze_pact_10seed.py",
230
- "size_bytes": 8391,
231
- "sha256": "eef3e58ea5944cbfa7953894d614abde5b3edf169ff30dd5935fc20a665e04c4"
232
- },
233
- {
234
- "path": "tools/audit_multitarget_data.py",
235
- "size_bytes": 6414,
236
- "sha256": "1703b8c905b594088f49cba3ba36801cad15fb4ef5d28fad5f6f6e57bfc6ec0e"
237
- },
238
- {
239
- "path": "tools/evaluate.py",
240
- "size_bytes": 4466,
241
- "sha256": "8c0e864dac43c3b25b4ec25d3c1bd41a751cd772dc3a511e9a51df0592317246"
242
- },
243
- {
244
- "path": "tools/evaluate_multitarget.py",
245
- "size_bytes": 6358,
246
- "sha256": "c6b9414ff1c47f7e9bcb51ea0c092dd7b14a4ada169349610c0c012c5fed2ed2"
247
- },
248
- {
249
- "path": "tools/extract_multitarget_vitl14.py",
250
- "size_bytes": 14372,
251
- "sha256": "1c7f85ee58dc34e674dc3f421c29ba9ac41cb13a74d60e94cdb9494f8dc7eef1"
252
- },
253
- {
254
- "path": "tools/fetch_paa_ground_truth.py",
255
- "size_bytes": 1412,
256
- "sha256": "2e2ff299288c8654a03ae2e06982de79527570f74054215be7f021640131e05f"
257
- },
258
- {
259
- "path": "tools/lock_multitarget_lengths.py",
260
- "size_bytes": 5736,
261
- "sha256": "5f3726e28abc46ba907d27d35623a2372415338f66c167291f87111663595566"
262
- },
263
- {
264
- "path": "tools/paper_statistics.py",
265
- "size_bytes": 5329,
266
- "sha256": "a01304f8159b302194f7fe3e3b0f0f86c5e1b739076e8f66c15773727dea2498"
267
- },
268
- {
269
- "path": "tools/prepare_multitarget_manifests.py",
270
- "size_bytes": 5558,
271
- "sha256": "33f00719a298db90e7e72addc2d70b9977b4a75472b4814722adfdac16454842"
272
- },
273
- {
274
- "path": "tools/pretrain_prevad.py",
275
- "size_bytes": 9522,
276
- "sha256": "afe46102e6ef750bb1e03fd5e858885d90bf70cc94b30b641a5de7765f48ce5b"
277
- },
278
- {
279
- "path": "tools/train_ablation.py",
280
- "size_bytes": 7963,
281
- "sha256": "ab14e6625847864b3a6c72fea99d2df23e79375cd3af05562b9fafc27ce4df73"
282
- },
283
- {
284
- "path": "tools/train_prevad_initialization_adaptation.py",
285
- "size_bytes": 15767,
286
- "sha256": "a555bc54eb31c7eb1e637ee166a09394de4d74392badfdd08a9c6c791fcf76b4"
287
- }
288
- ],
289
- "archive_sha256": "64ae1b063feb97fb275e5234ce088e83e98b6655a5ca286fdb396a6a647b1302",
290
- "archive_size_bytes": 266330,
291
- "license": "MIT for project code; third-party data not relicensed"
292
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
code/pact_vad_code.zip DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:64ae1b063feb97fb275e5234ce088e83e98b6655a5ca286fdb396a6a647b1302
3
- size 266330
 
 
 
 
evidence/checkpoint_cpu_audit.json DELETED
@@ -1,9 +0,0 @@
1
- {
2
- "verified_target_checkpoints": 180,
3
- "compact_checkpoints_requiring_source": 90,
4
- "source_sha256": "f3c3ecd73a046beab6bdf4fcff6c2850c0428afa3361236d20ed21ea91d9473a",
5
- "strict_state_load": true,
6
- "synthetic_cpu_forward": true,
7
- "test_metrics_recomputed": false,
8
- "training_performed": false
9
- }
 
 
 
 
 
 
 
 
 
 
evidence/selected_runs.json DELETED
The diff for this file is too large to render. See raw diff
 
evidence/statistical_analysis.json DELETED
@@ -1,664 +0,0 @@
1
- {
2
- "schema_version": "1.0",
3
- "paper_id": "PACT-VAD-PAA-2026",
4
- "source_ledger": "results/evaluations/pact_10seed_selected_runs.json",
5
- "source_ledger_sha256": "31c1c5be3c338ea53426d7d5b6852db6759aa2fe6f4936903b3cf6eef8633fb7",
6
- "analysis_class": "exploratory paired-seed analysis",
7
- "primary_metrics": {
8
- "UCF-Crime": "frame_auc",
9
- "ShanghaiTech": "frame_auc",
10
- "XD-Violence": "frame_ap"
11
- },
12
- "frame_contract": {
13
- "UCF-Crime": 1109504,
14
- "ShanghaiTech": 142912,
15
- "XD-Violence": 2330384
16
- },
17
- "inference": {
18
- "test": "exact two-sided paired sign-flip permutation",
19
- "interval": "paired-seed percentile bootstrap with 100,000 resamples",
20
- "multiplicity": "Holm adjustment across five contrasts within each dataset",
21
- "effect_size": "paired Cohen dz"
22
- },
23
- "summaries": [
24
- {
25
- "dataset": "UCF-Crime",
26
- "condition": "B",
27
- "primary_metric": "frame_auc",
28
- "n": 10,
29
- "mean_percent": 86.30887153798557,
30
- "sample_std_percent": 0.615004988415848,
31
- "min_percent": 84.85586336920258,
32
- "max_percent": 87.13535957910307
33
- },
34
- {
35
- "dataset": "UCF-Crime",
36
- "condition": "R",
37
- "primary_metric": "frame_auc",
38
- "n": 10,
39
- "mean_percent": 86.02374854564025,
40
- "sample_std_percent": 0.9082256288418151,
41
- "min_percent": 84.49767356578472,
42
- "max_percent": 87.16557788994217
43
- },
44
- {
45
- "dataset": "UCF-Crime",
46
- "condition": "P",
47
- "primary_metric": "frame_auc",
48
- "n": 10,
49
- "mean_percent": 86.49989674273198,
50
- "sample_std_percent": 0.6329946910522793,
51
- "min_percent": 85.7781330642883,
52
- "max_percent": 87.43318636009143
53
- },
54
- {
55
- "dataset": "UCF-Crime",
56
- "condition": "PW",
57
- "primary_metric": "frame_auc",
58
- "n": 10,
59
- "mean_percent": 85.87398674719044,
60
- "sample_std_percent": 1.2851527931288746,
61
- "min_percent": 83.73684463292778,
62
- "max_percent": 87.65190956852129
63
- },
64
- {
65
- "dataset": "UCF-Crime",
66
- "condition": "PLR",
67
- "primary_metric": "frame_auc",
68
- "n": 10,
69
- "mean_percent": 86.51692090449379,
70
- "sample_std_percent": 0.7306234476660713,
71
- "min_percent": 85.40760320658474,
72
- "max_percent": 87.61063051925949
73
- },
74
- {
75
- "dataset": "UCF-Crime",
76
- "condition": "PTSA",
77
- "primary_metric": "frame_auc",
78
- "n": 10,
79
- "mean_percent": 86.40713860219554,
80
- "sample_std_percent": 0.48804925357909995,
81
- "min_percent": 85.8799794528272,
82
- "max_percent": 87.1386981276887
83
- },
84
- {
85
- "dataset": "ShanghaiTech",
86
- "condition": "B",
87
- "primary_metric": "frame_auc",
88
- "n": 10,
89
- "mean_percent": 92.77011539650368,
90
- "sample_std_percent": 1.5618583016487422,
91
- "min_percent": 90.04447631297636,
92
- "max_percent": 94.60276239945281
93
- },
94
- {
95
- "dataset": "ShanghaiTech",
96
- "condition": "R",
97
- "primary_metric": "frame_auc",
98
- "n": 10,
99
- "mean_percent": 92.61952225069379,
100
- "sample_std_percent": 1.547878689209576,
101
- "min_percent": 90.05242571494418,
102
- "max_percent": 94.6884238547841
103
- },
104
- {
105
- "dataset": "ShanghaiTech",
106
- "condition": "P",
107
- "primary_metric": "frame_auc",
108
- "n": 10,
109
- "mean_percent": 92.94802057452986,
110
- "sample_std_percent": 1.5459361172431185,
111
- "min_percent": 90.08941407543868,
112
- "max_percent": 94.5937676067987
113
- },
114
- {
115
- "dataset": "ShanghaiTech",
116
- "condition": "PW",
117
- "primary_metric": "frame_auc",
118
- "n": 10,
119
- "mean_percent": 96.38290027784046,
120
- "sample_std_percent": 0.7660841788990864,
121
- "min_percent": 94.94117346500852,
122
- "max_percent": 97.72209000061719
123
- },
124
- {
125
- "dataset": "ShanghaiTech",
126
- "condition": "PLR",
127
- "primary_metric": "frame_auc",
128
- "n": 10,
129
- "mean_percent": 91.9058010790412,
130
- "sample_std_percent": 2.3114700607261893,
131
- "min_percent": 88.06032035764862,
132
- "max_percent": 94.21655296879356
133
- },
134
- {
135
- "dataset": "ShanghaiTech",
136
- "condition": "PTSA",
137
- "primary_metric": "frame_auc",
138
- "n": 10,
139
- "mean_percent": 94.34888118151518,
140
- "sample_std_percent": 3.051569739813902,
141
- "min_percent": 85.75492064657244,
142
- "max_percent": 95.99684370324694
143
- },
144
- {
145
- "dataset": "XD-Violence",
146
- "condition": "B",
147
- "primary_metric": "frame_ap",
148
- "n": 10,
149
- "mean_percent": 76.96445150049115,
150
- "sample_std_percent": 3.1571025279420413,
151
- "min_percent": 71.777122297029,
152
- "max_percent": 82.01578096149589
153
- },
154
- {
155
- "dataset": "XD-Violence",
156
- "condition": "R",
157
- "primary_metric": "frame_ap",
158
- "n": 10,
159
- "mean_percent": 77.66587084699974,
160
- "sample_std_percent": 2.6160485716669837,
161
- "min_percent": 73.15066356001026,
162
- "max_percent": 80.5807673083777
163
- },
164
- {
165
- "dataset": "XD-Violence",
166
- "condition": "P",
167
- "primary_metric": "frame_ap",
168
- "n": 10,
169
- "mean_percent": 75.16302344531402,
170
- "sample_std_percent": 3.7270156539456374,
171
- "min_percent": 69.0688821649162,
172
- "max_percent": 79.581843117671
173
- },
174
- {
175
- "dataset": "XD-Violence",
176
- "condition": "PW",
177
- "primary_metric": "frame_ap",
178
- "n": 10,
179
- "mean_percent": 78.76864417266334,
180
- "sample_std_percent": 2.4185326398506257,
181
- "min_percent": 73.70322438545143,
182
- "max_percent": 81.47473393762372
183
- },
184
- {
185
- "dataset": "XD-Violence",
186
- "condition": "PLR",
187
- "primary_metric": "frame_ap",
188
- "n": 10,
189
- "mean_percent": 77.06950427797389,
190
- "sample_std_percent": 3.4618134908105174,
191
- "min_percent": 71.53131622636135,
192
- "max_percent": 81.11560047327306
193
- },
194
- {
195
- "dataset": "XD-Violence",
196
- "condition": "PTSA",
197
- "primary_metric": "frame_ap",
198
- "n": 10,
199
- "mean_percent": 76.0807839778277,
200
- "sample_std_percent": 3.785239942146865,
201
- "min_percent": 69.47811336732286,
202
- "max_percent": 80.65091765789946
203
- }
204
- ],
205
- "contrasts": [
206
- {
207
- "dataset": "UCF-Crime",
208
- "primary_metric": "frame_auc",
209
- "contrast": "P-R",
210
- "n": 10,
211
- "mean_difference_pp": 0.47614819709170925,
212
- "sample_std_difference_pp": 0.6738799858957101,
213
- "paired_cohen_dz": 0.7065771458679263,
214
- "bootstrap_percentile_95_ci_pp": [
215
- 0.08866852962618688,
216
- 0.874144922510286
217
- ],
218
- "exact_two_sided_sign_flip_p": 0.05859375,
219
- "positive_seeds": 7,
220
- "zero_seeds": 0,
221
- "paired_differences_pp": [
222
- -0.501350064060957,
223
- -0.1625708889240185,
224
- 0.8606907712169765,
225
- -0.1379006225970314,
226
- 1.3731621919270132,
227
- 0.07826817617967441,
228
- 0.41151835242717594,
229
- 1.498934354606296,
230
- 0.9048673846615896,
231
- 0.43586231548037313
232
- ],
233
- "holm_adjusted_p_within_dataset": 0.29296875,
234
- "exploratory_reject_0_05_after_holm": false
235
- },
236
- {
237
- "dataset": "UCF-Crime",
238
- "primary_metric": "frame_auc",
239
- "contrast": "P-B",
240
- "n": 10,
241
- "mean_difference_pp": 0.19102520474639229,
242
- "sample_std_difference_pp": 0.6148254894061234,
243
- "paired_cohen_dz": 0.3106982518420125,
244
- "bootstrap_percentile_95_ci_pp": [
245
- -0.1710222928306364,
246
- 0.5521164314698079
247
- ],
248
- "exact_two_sided_sign_flip_p": 0.34765625,
249
- "positive_seeds": 6,
250
- "zero_seeds": 0,
251
- "paired_differences_pp": [
252
- -0.49900722099898154,
253
- 0.1571261092780274,
254
- 0.9222696950857201,
255
- -0.08419634613602289,
256
- 0.34429841511230963,
257
- 0.10848648701877739,
258
- -0.11028462712412423,
259
- -0.8219818254977085,
260
- 1.0580260509306605,
261
- 0.835515309795265
262
- ],
263
- "holm_adjusted_p_within_dataset": 1.0,
264
- "exploratory_reject_0_05_after_holm": false
265
- },
266
- {
267
- "dataset": "UCF-Crime",
268
- "primary_metric": "frame_auc",
269
- "contrast": "PW-P",
270
- "n": 10,
271
- "mean_difference_pp": -0.6259099955415337,
272
- "sample_std_difference_pp": 1.1833212462666256,
273
- "paired_cohen_dz": -0.5289434272529776,
274
- "bootstrap_percentile_95_ci_pp": [
275
- -1.3552855155492303,
276
- 0.02540626899903531
277
- ],
278
- "exact_two_sided_sign_flip_p": 0.13671875,
279
- "positive_seeds": 5,
280
- "zero_seeds": 0,
281
- "paired_differences_pp": [
282
- -1.2236151016239316,
283
- 0.523216448076913,
284
- 0.48323042715608944,
285
- 0.1596235672418067,
286
- 0.09844634212835546,
287
- -0.19458409924232312,
288
- -1.3151153746127742,
289
- -2.2597632874632345,
290
- -2.74926208550611,
291
- 0.2187232084298718
292
- ],
293
- "holm_adjusted_p_within_dataset": 0.546875,
294
- "exploratory_reject_0_05_after_holm": false
295
- },
296
- {
297
- "dataset": "UCF-Crime",
298
- "primary_metric": "frame_auc",
299
- "contrast": "PLR-P",
300
- "n": 10,
301
- "mean_difference_pp": 0.017024161761827195,
302
- "sample_std_difference_pp": 0.7395930869805823,
303
- "paired_cohen_dz": 0.023018281351613223,
304
- "bootstrap_percentile_95_ci_pp": [
305
- -0.3748270770347925,
306
- 0.48567602306528457
307
- ],
308
- "exact_two_sided_sign_flip_p": 0.94921875,
309
- "positive_seeds": 6,
310
- "zero_seeds": 0,
311
- "paired_differences_pp": [
312
- 1.7634403418945732,
313
- 0.2705592639308807,
314
- 0.06794761528136073,
315
- -0.6083552701377948,
316
- 0.010503711827858364,
317
- -0.49671370026659334,
318
- 0.1247801977640628,
319
- -0.32874716884756516,
320
- 0.31480406940804784,
321
- -0.9479774432365584
322
- ],
323
- "holm_adjusted_p_within_dataset": 1.0,
324
- "exploratory_reject_0_05_after_holm": false
325
- },
326
- {
327
- "dataset": "UCF-Crime",
328
- "primary_metric": "frame_auc",
329
- "contrast": "PTSA-P",
330
- "n": 10,
331
- "mean_difference_pp": -0.09275814053641418,
332
- "sample_std_difference_pp": 0.7503461280082827,
333
- "paired_cohen_dz": -0.12362046937275628,
334
- "bootstrap_percentile_95_ci_pp": [
335
- -0.526177422632681,
336
- 0.35848158284327203
337
- ],
338
- "exact_two_sided_sign_flip_p": 0.671875,
339
- "positive_seeds": 3,
340
- "zero_seeds": 0,
341
- "paired_differences_pp": [
342
- 1.3603963378865491,
343
- -0.004807978094911736,
344
- 0.5112478158395439,
345
- -0.09444633545633296,
346
- -0.5156392623704487,
347
- -1.3542746840327857,
348
- 0.4133830735109645,
349
- -0.017285240490694243,
350
- -0.6814970905943629,
351
- -0.5446580415616631
352
- ],
353
- "holm_adjusted_p_within_dataset": 1.0,
354
- "exploratory_reject_0_05_after_holm": false
355
- },
356
- {
357
- "dataset": "ShanghaiTech",
358
- "primary_metric": "frame_auc",
359
- "contrast": "P-R",
360
- "n": 10,
361
- "mean_difference_pp": 0.3284983238360861,
362
- "sample_std_difference_pp": 1.0746526966344507,
363
- "paired_cohen_dz": 0.30567859259541474,
364
- "bootstrap_percentile_95_ci_pp": [
365
- -0.030493048949836732,
366
- 1.0161596418407282
367
- ],
368
- "exact_two_sided_sign_flip_p": 0.755859375,
369
- "positive_seeds": 7,
370
- "zero_seeds": 0,
371
- "paired_differences_pp": [
372
- 0.03698836049449428,
373
- 0.004697239564255273,
374
- 3.384886387161856,
375
- 0.004423886522608722,
376
- -0.06276333594557038,
377
- 0.00746475441291139,
378
- 0.021865288163536167,
379
- 0.007744017790578539,
380
- -0.09465624798540295,
381
- -0.025667111818405974
382
- ],
383
- "holm_adjusted_p_within_dataset": 0.755859375,
384
- "exploratory_reject_0_05_after_holm": false
385
- },
386
- {
387
- "dataset": "ShanghaiTech",
388
- "primary_metric": "frame_auc",
389
- "contrast": "P-B",
390
- "n": 10,
391
- "mean_difference_pp": 0.1779051780261931,
392
- "sample_std_difference_pp": 0.4726917251775792,
393
- "paired_cohen_dz": 0.37636617810341044,
394
- "bootstrap_percentile_95_ci_pp": [
395
- 0.004431126684261155,
396
- 0.48667714029927417
397
- ],
398
- "exact_two_sided_sign_flip_p": 0.1015625,
399
- "positive_seeds": 7,
400
- "zero_seeds": 0,
401
- "paired_differences_pp": [
402
- 0.04493776246232706,
403
- 0.003009838626122363,
404
- 0.09581984539240862,
405
- 1.5132617523670566,
406
- 0.15843837810507466,
407
- 0.0072756236597637525,
408
- -0.007851881423237561,
409
- 0.004044147432336054,
410
- -0.008994792654104522,
411
- -0.030888893705816045
412
- ],
413
- "holm_adjusted_p_within_dataset": 0.3046875,
414
- "exploratory_reject_0_05_after_holm": false
415
- },
416
- {
417
- "dataset": "ShanghaiTech",
418
- "primary_metric": "frame_auc",
419
- "contrast": "PW-P",
420
- "n": 10,
421
- "mean_difference_pp": 3.4348797033105845,
422
- "sample_std_difference_pp": 1.741038468369999,
423
- "paired_cohen_dz": 1.9728913322210504,
424
- "bootstrap_percentile_95_ci_pp": [
425
- 2.4492103972060497,
426
- 4.47660805910962
427
- ],
428
- "exact_two_sided_sign_flip_p": 0.001953125,
429
- "positive_seeds": 10,
430
- "zero_seeds": 0,
431
- "paired_differences_pp": [
432
- 6.412744150255989,
433
- 4.221113236320184,
434
- 2.071108819082801,
435
- 1.9378351742330824,
436
- 5.610093920115644,
437
- 4.042718608591045,
438
- 3.793642272212716,
439
- 3.4654627371265567,
440
- 1.6779495720377402,
441
- 1.1161285431300905
442
- ],
443
- "holm_adjusted_p_within_dataset": 0.009765625,
444
- "exploratory_reject_0_05_after_holm": true
445
- },
446
- {
447
- "dataset": "ShanghaiTech",
448
- "primary_metric": "frame_auc",
449
- "contrast": "PLR-P",
450
- "n": 10,
451
- "mean_difference_pp": -1.0422194954886899,
452
- "sample_std_difference_pp": 1.3587267189636838,
453
- "paired_cohen_dz": -0.7670560098233753,
454
- "bootstrap_percentile_95_ci_pp": [
455
- -1.8314914749977884,
456
- -0.24081327287709836
457
- ],
458
- "exact_two_sided_sign_flip_p": 0.046875,
459
- "positive_seeds": 3,
460
- "zero_seeds": 0,
461
- "paired_differences_pp": [
462
- -1.876652853177041,
463
- 0.7155762044965597,
464
- -1.9459626250739714,
465
- -3.029022099293144,
466
- -2.5150518897213314,
467
- -1.0562642270099842,
468
- 0.5626654681687326,
469
- 0.7705231210350938,
470
- -0.8951698916489059,
471
- -1.152836162662907
472
- ],
473
- "holm_adjusted_p_within_dataset": 0.1875,
474
- "exploratory_reject_0_05_after_holm": false
475
- },
476
- {
477
- "dataset": "ShanghaiTech",
478
- "primary_metric": "frame_auc",
479
- "contrast": "PTSA-P",
480
- "n": 10,
481
- "mean_difference_pp": 1.4008606069853036,
482
- "sample_std_difference_pp": 3.6099230318993056,
483
- "paired_cohen_dz": 0.38805830335065683,
484
- "bootstrap_percentile_95_ci_pp": [
485
- -1.0019320629923074,
486
- 3.1242835656885903
487
- ],
488
- "exact_two_sided_sign_flip_p": 0.25390625,
489
- "positive_seeds": 9,
490
- "zero_seeds": 0,
491
- "paired_differences_pp": [
492
- 4.707918063888572,
493
- 2.380188364478064,
494
- 0.7254139588277364,
495
- -7.977462026940585,
496
- 4.280166358819426,
497
- 3.4794118689624676,
498
- 1.9621813259475651,
499
- 2.9818177229611886,
500
- 0.15758507333993865,
501
- 1.311385359568662
502
- ],
503
- "holm_adjusted_p_within_dataset": 0.5078125,
504
- "exploratory_reject_0_05_after_holm": false
505
- },
506
- {
507
- "dataset": "XD-Violence",
508
- "primary_metric": "frame_ap",
509
- "contrast": "P-R",
510
- "n": 10,
511
- "mean_difference_pp": -2.5028474016856936,
512
- "sample_std_difference_pp": 3.7970316694064854,
513
- "paired_cohen_dz": -0.6591589482520471,
514
- "bootstrap_percentile_95_ci_pp": [
515
- -4.83103231511396,
516
- -0.41566813133122943
517
- ],
518
- "exact_two_sided_sign_flip_p": 0.060546875,
519
- "positive_seeds": 3,
520
- "zero_seeds": 0,
521
- "paired_differences_pp": [
522
- -1.782718050947063,
523
- -9.605781674238411,
524
- -4.081781395094053,
525
- 1.593739777610137,
526
- 1.679566900462559,
527
- -0.9989241907066959,
528
- -3.301268935568169,
529
- 0.11533996775947575,
530
- -7.903595827666631,
531
- -0.743050588468086
532
- ],
533
- "holm_adjusted_p_within_dataset": 0.2421875,
534
- "exploratory_reject_0_05_after_holm": false
535
- },
536
- {
537
- "dataset": "XD-Violence",
538
- "primary_metric": "frame_ap",
539
- "contrast": "P-B",
540
- "n": 10,
541
- "mean_difference_pp": -1.801428055177113,
542
- "sample_std_difference_pp": 4.250677489559765,
543
- "paired_cohen_dz": -0.423797867422702,
544
- "bootstrap_percentile_95_ci_pp": [
545
- -4.4065599088380125,
546
- 0.5847677754112156
547
- ],
548
- "exact_two_sided_sign_flip_p": 0.220703125,
549
- "positive_seeds": 5,
550
- "zero_seeds": 0,
551
- "paired_differences_pp": [
552
- 1.2900142493247557,
553
- -8.664476863984982,
554
- -2.7082401321127936,
555
- 0.9040846572590966,
556
- 2.1517060376898822,
557
- 2.532855901140829,
558
- -5.697944513036479,
559
- 1.9378528975128484,
560
- -7.644066508631764,
561
- -2.1160662769325267
562
- ],
563
- "holm_adjusted_p_within_dataset": 0.662109375,
564
- "exploratory_reject_0_05_after_holm": false
565
- },
566
- {
567
- "dataset": "XD-Violence",
568
- "primary_metric": "frame_ap",
569
- "contrast": "PW-P",
570
- "n": 10,
571
- "mean_difference_pp": 3.6056207273493195,
572
- "sample_std_difference_pp": 4.2465214601129615,
573
- "paired_cohen_dz": 0.8490762995586997,
574
- "bootstrap_percentile_95_ci_pp": [
575
- 1.136738386257494,
576
- 6.146456140493614
577
- ],
578
- "exact_two_sided_sign_flip_p": 0.03515625,
579
- "positive_seeds": 8,
580
- "zero_seeds": 0,
581
- "paired_differences_pp": [
582
- 1.9790758469699687,
583
- 8.16098965444495,
584
- 10.227860604746809,
585
- 6.2254074973216,
586
- -0.28093745827090233,
587
- 1.6220381891238755,
588
- -2.6146120630079817,
589
- 2.572776828793888,
590
- 7.885133200267836,
591
- 0.27847497310314884
592
- ],
593
- "holm_adjusted_p_within_dataset": 0.17578125,
594
- "exploratory_reject_0_05_after_holm": false
595
- },
596
- {
597
- "dataset": "XD-Violence",
598
- "primary_metric": "frame_ap",
599
- "contrast": "PLR-P",
600
- "n": 10,
601
- "mean_difference_pp": 1.906480832659863,
602
- "sample_std_difference_pp": 6.0098513126238995,
603
- "paired_cohen_dz": 0.31722595676455995,
604
- "bootstrap_percentile_95_ci_pp": [
605
- -1.5559367128000012,
606
- 5.4961911197720035
607
- ],
608
- "exact_two_sided_sign_flip_p": 0.34375,
609
- "positive_seeds": 6,
610
- "zero_seeds": 0,
611
- "paired_differences_pp": [
612
- 1.394389607797586,
613
- 8.699277658612658,
614
- 11.981760304260437,
615
- -1.5021117572136977,
616
- -7.039058682927424,
617
- -2.6783364022229916,
618
- -3.574025147876947,
619
- 2.2475161336638183,
620
- 7.764540072090831,
621
- 1.7708565404143561
622
- ],
623
- "holm_adjusted_p_within_dataset": 0.6875,
624
- "exploratory_reject_0_05_after_holm": false
625
- },
626
- {
627
- "dataset": "XD-Violence",
628
- "primary_metric": "frame_ap",
629
- "contrast": "PTSA-P",
630
- "n": 10,
631
- "mean_difference_pp": 0.9177605325136806,
632
- "sample_std_difference_pp": 5.785539205278465,
633
- "paired_cohen_dz": 0.15863007750018482,
634
- "bootstrap_percentile_95_ci_pp": [
635
- -2.3597143500990683,
636
- 4.400735321882423
637
- ],
638
- "exact_two_sided_sign_flip_p": 0.6171875,
639
- "positive_seeds": 5,
640
- "zero_seeds": 0,
641
- "paired_differences_pp": [
642
- 1.5893649854901604,
643
- 2.973767489094925,
644
- 9.894136036583179,
645
- -5.771213072979265,
646
- -1.5655987664038085,
647
- -0.5537027174259967,
648
- 4.0724211073075,
649
- -5.973926632876814,
650
- 9.393896394596934,
651
- -4.881539498250009
652
- ],
653
- "holm_adjusted_p_within_dataset": 0.6875,
654
- "exploratory_reject_0_05_after_holm": false
655
- }
656
- ],
657
- "limitations": [
658
- "Contrasts and multiplicity families were specified after the experiments and are exploratory.",
659
- "Ten paired seeds quantify optimizer/initialization variability, not dataset sampling uncertainty.",
660
- "Seed pairing does not make results independent replications.",
661
- "Stored metrics were audited but not recomputed from raw frame-score vectors in this package.",
662
- "UCF-Crime uses a 1,109,504-frame evaluator variant; comparisons require the same ground-truth contract."
663
- ]
664
- }