UnWeave: reproduction package for "Unlearning in Incrementally Fine-Tuned Object Detectors: A Fisher-Ratio Selection Rule and Its Measurement Protocol"

Reproduction package for a study of machine unlearning in incrementally fine-tuned single-stage object detectors. The archive contains every experiment script, in the order in which it is run, together with the derived result tables in CSV form from which each table and figure in the paper is generated. Model checkpoints are omitted because of their size and the image data is not redistributed; the weapon chain is built from a publicly available dataset distributed through Roboflow Universe and the second chain from PASCAL VOC, both obtainable from their original sources. Contents: scripts/ 73 Python scripts, 14 queue scripts and 2 dataset configuration files (the full pipeline, the published baselines, the measurement studies, and the GPU queues as they were actually run) results/ 59 CSV files (2,348 rows) covering the weapon and VOC chains, the five published baselines, the probe-strength and probe-repeatability sweeps, the schedule and random-selection controls, the degenerate-state diagnosis, the cross-architecture selection geometry, and the method run on RT-DETR itself README.txt environment, data provenance, pipeline order, protocol notes, and the verifier scripts Five verifier scripts are included and are the quickest way into the package. Each recomputes the paper's numbers frothe CSV files and then checks that the same in the rendered PDF, on the principle that a compile without warnings is no evidence that a figure survived into the text. A single mismatch fails the script. Five other files are superseded pilots or smoke tests, named as such rather than removed, so that every number in the paper can be traced to the run that produced it. None of them feeds a table or figure. The queue scripts are included deliberately, scope bug: two GS-LoRA seeds were launchedwith the adapter scope left at its default and so adapted every convolution rather than the neck and head. The paper reports both the corrected runs and the accidental ones, and the scripts let a reader see how the discrepancy was found. Version 1.1 adds the arm in which the method is run on RT-DETR, the architecture on which its selection criterion is not head-local: unweave_rtdetr_seeds.py, res and verify_rtdetr_claims.py, together withthe README section describing the verifiers.