Evidence bundle for "Deadline-Aware Hardening of Real-Time Object Detection Against Candidate-Inflation Latency Attacks"

The measurement record behind every number in the paper and its revision, with the mappings that tie them together and scripts that check them. Version 3.2 adds no measurement: revised-main.tex carries the final corrections (Table 1 states its scoring points; one interval endpoint corrected to its artifact value), and revision_claims.yaml traces the 70 numbers it prints. Version 3.1 adds no measurement: revised-main.tex is the final resubmitted manuscript after a readability pass, and revision_claims.yaml traces the 69 numbers it prints, listing the values no longer quoted in the prose under retired_from_main with their artifacts. Version 3 adds, under artifacts/research/results/revision/r11-lossless/, a served configuration that passes a 33.3-ms deadline under the principal attack delivered losslessly: one GPU, a worker with native operators and a strict uncompressed-PNG decoder, loopback delivery with other host workloads suspended, and matched uncapped and clean controls. Three campaigns on one frozen design (45 streams, 174,000 offers, two native builds) are given as per-request traces, with the protocol, worker sources, both binaries and verify_r11.py, which recomputes every verdict from the traces. Version 2 accompanies the journal revision. Under artifacts/research/results/revision/ it adds the analysis plan, frozen and hashed before the data, with its dated amendments and post-data notes; the corpus manifests (137 VisDrone inputs, one frame per video sequence, and 137 BDD100K images); every timed request of the served sessions and streams; the suppression-replay, detection-quality and edge-board measurements; and the analysis scripts as they were run. verify_revision.py traces each new or changed number in the revised manuscript (revised-main.tex) to its artifact. The submitted version’s record is kept: evidence_ledger.json holds each registered claim with its sample count, timing scope, split, model, runtime, hardware and caveat; claims_registry.yaml records where each value lives, never the value; verify.py re-resolves all 129 claims. Both scripts need only Python and PyYAML, with no dataset, weights, GPU or network. MANIFEST.sha256 hashes every file. Not included: crafted images, clean twins and captured candidate arrays, which are derivatives of VisDrone2019-DET and BDD100K and are redistributed subject to those corpora’s terms; available from the corresponding author on request. The reference implementation is a separate record, doi:10.5281/zenodo.22117976.