Bounded Adaptation V2 and V2.1: Calibration, Confirmation, Activation, and Transition-Entry Carryover Evidence
Year: 2026
Type: dataset
Abstract: This dataset provides the frozen evidence and reproducibility materials underlying the study “When Reactive Gates Are Too Late: Transition-Entry Carryover Harm in Stateful Computer-Vision Adaptation.” It contains three chronologically distinct evidence layers for a deterministic multi-stage segmentation-OCR pipeline. V2 calibration evaluated a prespecified 25-value grid across 3,000 runs and selected the largest empirically admissible tested radius, $\delta_W=12.0$, with zero harmful trajectories among 84 informative trajectories. The subsequent V2 confirmation used five disjoint trajectories that were not involved in parameter selection. After V2 confirmation had been completed, a separate V2.1 targeted challenge was designed and frozen before its own execution. V2.1 was therefore frozen after V2 confirmation and is not part of the original V2 preregistration. It examined mechanism activation and transition-entry carryover across 18 independent trajectories and 155 state changes. Transition-entry harm occurred in 9/12 trajectories containing an entry into broken-character drift and in 0/6 without such an entry (two-sided Fisher exact $p=0.0090$). The deposit includes frozen archives, input locks, configuration records, audit materials, manifests, and confirmation evidence. The three evidence layers must remain analytically separate and must not be interpreted as a single preregistered experiment. All experimental image inputs are synthetic and contain no real vehicle identifiers or personal data.
Authors: Szabó, Mihály, Kovari, Attila, Kertész, Gábor
Language: enc
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7208838388