Supplementary Material for "AdaSkip-Gate: Reliability-Aware Post-Export Activation Correction for 8-Bit Medical Image Segmentation" (derived results, provenance and diagnostics)

SUPPLEMENTARY MATERIAL 1AdaSkip-Gate: Reliability-Aware Post-Export Activation Correction for 8-Bit Medical Image Segmentation Prepared for submission to Biomedical Signal Processing and Control. This package contains the derived, machine-readable results supporting themanuscript's "Data and code availability" section. It does NOT contain theMedical Segmentation Decathlon or Severstal images (public third-partydatasets, available from their original providers) and it does NOT containthe implementation code, which is not publicly available. Each folder below corresponds to one group of results in the manuscript. ------------------------------------------------------------------------01_main_results/ -> "the 100-run tables"------------------------------------------------------------------------ adaskip_all_results.csv 100 experimental runs: task, architecture, protocol, fold, FP32/INT8 Dice, ungated and gated effects, CI bounds, deployment decision, deployed sites. adaskip_excluded_runs.csv Runs outside the canonical grid. A reason is supplied where documented; otherwise the file explicitly states that the original artifacts do not record a more specific rationale. adaskip_factorial_ablation.csv Skip-adjacent vs. non-skip location factorial control. severity_variants_registry_ablation.csv Severity-variant registry: FP32/INT8 checkpoints, hashes, calibration case IDs per protocol. ------------------------------------------------------------------------02_provenance_and_splits/ -> "subject-level split manifests, accepted-site registries and graph hashes"------------------------------------------------------------------------ all_split_manifests.csv Per-case assignment to calibration / candidate-selection / combination- validation / locked-evaluation roles, by run and outer fold. provenance_hashes.csv Deduplicated SHA-256 hashes, file sizes, and portable logical artifact paths (one row per unique hash). ------------------------------------------------------------------------03_gate_diagnostics/ -> "the detailed sequential-Holm inventory, practical-effect and data-size sensitivity tables"------------------------------------------------------------------------ holm_sequential_test_inventory.txt Per-step candidate family size, raw and Holm-adjusted p-values, and the accept/reject decision (including the rejected -24.71 pp candidate). tau_sensitivity_all.txt Deployed-set and recovery stability across the practical-effect threshold tau in {0, 0.05, 0.1, 0.2, 0.5}. gate_size_sensitivity.txt Recovery and site-selection stability as the gate subject pool grows from 8 to 100. affine_calibration_subject_sensitivity.txt Recovery stability as the number of calibration subjects varies from 2 upward. partition_sizes_representative_runs.txt Calibration / selection / validation / evaluation partition sizes for the five archived base runs represented in the supplied split manifests. ------------------------------------------------------------------------04_mechanism_analysis/ -> "Shapley pathway attribution, nonlinear mapping-family ablation"------------------------------------------------------------------------ pathway_decomposition_summary.txt Shapley decomposition of INT8 error between the skip pathway and the remainder of the graph, by task. nonlinear_mapping_ablation_all.txt Gated and ungated recovery for scale-only, bias-only, affine, shrinkage-affine, and two-knot piecewise-linear (PWL2) correction mappings. ------------------------------------------------------------------------05_sota_baselines/ -> "measured AdaRound and post-AdaRound gate tables"------------------------------------------------------------------------ adaround_vs_matched_aimet_nearest.csv Matched-control comparison isolating AdaRound's rounding contribution from the AIMET calibration pipeline. aimet_methods_vs_canonical_int8.csv Full comparison of vanilla INT8, AIMET-nearest, and AdaRound against the canonical calibration protocol. post_adaround_gate.csv AdaSkip-Gate applied on top of an AdaRound-quantized model: additional recovery where headroom remains, abstention where it does not. ------------------------------------------------------------------------06_deployment_cost/------------------------------------------------------------------------ archived_latency_context.txt Archived CPU/workstation timing context across five detailed medical runs, reported separately from the mobile benchmark to avoid pooling. mobile_latency_raw_device_logs.txt Raw output of the Flutter/ONNX Runtime benchmark on the four mobile devices and macOS (hippocampus StackLite25D, input 1x7x256x256, 5 warm-up and 20 timed runs): median, mean, p95, p99, std latency and RSS memory per model tag (SL25-INT8, SL25-ADASKIP). The SL25-FP32 rows are included as logged but are not used in the manuscript. mobile_latency_summary.csv Median latency and batch-1 throughput per device (INT8 vs. AdaSkip-Gate), as reported in the manuscript latency table and taken from the raw logs. ------------------------------------------------------------------------07_cross_domain_and_boundary/------------------------------------------------------------------------ severstal_abstention_summary.csv Severstal out-of-domain abstention test: protocol-level effects on 400 locked images. boundary_metrics_representative.csv Representative class-wise Dice, precision, recall, HD95 and ASD for hippocampus (fold 0), prostate and spleen. ------------------------------------------------------------------------NOTE ON FOLDERS 06 AND 07------------------------------------------------------------------------mobile_latency_summary.csv, severstal_abstention_summary.csv andboundary_metrics_representative.csv are summary tables that reproduce thevalues printed in the manuscript; they are not raw per-run or per-imageoutputs. The last column of mobile_latency_summary.csv is recomputed from thetwo latency columns. The raw device logs for the mobile benchmark areprovided in mobile_latency_raw_device_logs.txt. ------------------------------------------------------------------------PACKAGE NORMALIZATION------------------------------------------------------------------------For submission portability, environment-specific /kaggle paths have beenreplaced with dataset- or experiment-relative logical paths. The splitmanifest still contains the original case IDs, partition assignments, run IDs,and seeds. provenance_hashes.csv is deduplicated by SHA-256. The canonical100-run table removes the incomplete hippocampus/UNet2D/minmax_pt_32 cell andincludes the regenerated spleen/StackLite25D/minmax_pt_128 fold 4.PACKAGE_VERIFICATION.txt recomputes the headline numbers from thecanonical 100-run table. ------------------------------------------------------------------------NOTE ON SCOPE------------------------------------------------------------------------BRECQ readiness/validation artifacts (architecture preflight, checkpointmatching, forward-equivalence and smoke-test logs) are intentionallyexcluded from this package. Those runs are flagged "reportable: False" intheir own logs and are not cited as a result anywhere in the manuscript;BRECQ is discussed only as related work. Likewise, exploratorytransformer-architecture files are excluded, as transformer generalizationis named only as future work in the manuscript and is not a reported resultin this study.