2026-10-14T03:28:02Z https://api.figshare.com/v2/oai

oai:figshare.com:article/33511363 2026-09-30T22:22:43Z category_28864 item_type_3 month_year_09_2026

Computational record for Who Is Seen in Time? Timely-Warning Opportunity and a Partial Equity Audit of a V-JEPA-Enhanced Pedestrian Warning System Jingyi Huang (24752965) Artificial intelligence not elsewhere classified pedestrian safety warning systems V-JEPA algorithmic fairness reproducibility <p dir="ltr">Code, configurations, environment records, fixed held-out predictions, non-restricted derived results, and verification evidence accompanying manuscript TDS-26-0007. The study measures timely-warning opportunity and conducts a partial equity audit; it does not measure braking, collision avoidance, injury reduction, or field effectiveness.</p><p dir="ltr">The 142 computational-record files are supplied individually, without a ZIP archive. FILE_MAP.json and restore_file_layout.py are also included. Download the individual files into one folder and run python restore_file_layout.py to verify their SHA-256 fingerprints and restore the canonical directory structure. The analysis package is in pedfair; executable procedures are in scripts; regression tests are in tests. README.md, RESEARCH_PROTOCOL.md, NUMERICAL_TRACEABILITY.md, and R2_VERIFICATION.md describe inputs, analysis order, numerical provenance, and verification limits.</p><p dir="ltr">The original VideoMAE comparator state could not be reconstructed. Original-versus-rerun values, corrected JAAD eligibility and test denominators, downstream refit agreement, small raw-input reference checks, and the retained test record are supplied. The operating-definition sensitivity CSV omitted from the earlier delivery is included; its 12 V-JEPA-versus-GRU rows match Supplementary Table S6.</p><p dir="ltr">Raw videos, restricted raw annotations, and pretrained model weights are not redistributed. Obtain these inputs from the official providers under their applicable terms. Bootstrap intervals condition on fixed held-out predictions and do not repeat fitting, calibration, or threshold selection. Full-array feature regeneration requires the official inputs and suitable compute; the reference checks do not establish reproduction of the unavailable original comparator state.</p> 2026-09-30T22:22:43Z Dataset Dataset 10.6084/m9.figshare.33511363.v1 https://figshare.com/articles/dataset/Computational_record_for_Who_Is_Seen_in_Time_Timely-Warning_Opportunity_and_a_Partial_Equity_Audit_of_a_V-JEPA-Enhanced_Pedestrian_Warning_System/33511363 CC BY 4.0