A Reproducible Evaluation Protocol for Synthetic Multi-UAV Coverage Planning

Code, frozen configurations, and rendered output for the manuscript A Reproducible Evaluation Protocol for Synthetic Multi-UAV Coverage Planning Under Bounded Execution Stress (IEEE Access).

The study is deliberately synthetic. It evaluates transparent coverage-planning rules inside a local ENU simulator under matched initial state, action contract, safety projection, and scoring map, and reports the verdict as returned -- including the result that the authors' own rule does not dominate the transparent baselines.

Nothing in this repository is geographic, real-flight, or field evidence.

What is here

Directory	Contents
reproducibility/	Simulator, planners, baseline rules, frozen configurations, validation and rendering scripts, tests, the non-evidence smoke fixture, formal-artifact-manifest.json, and the complete self-contained protocol-calibration study.
supplementary/	Rendered output of the two supplementary synthetic studies: the mechanism ablation and the protocol calibration.

reproducibility/README.md is the authoritative document for the environment, the test targets, the reproduction sequence, and the data-access restrictions. Read it before running anything.

Quick start

The protocol-calibration study is the one part that reruns from this repository alone, with no external input:

cd reproducibility
python -m pip install -r requirements-dev.txt
python -m pip install -e program[dev]
python run_protocol_calibration.py

It writes raw, hash-registered artifacts under results/protocol_calibration/runs/protocol-calibration-v1/ and renders tables, a figure, a CSV summary, and a claim ledger into ../supplementary/protocol-calibration/rendered/.

Use --render-only to regenerate the tables and figure from an existing run's provenance. Every value is read back from the hash-registered provenance files, so that path needs no episodes and no recomputation.

About the compressed rollouts file

reproducibility/results/protocol_calibration/runs/protocol-calibration-v1/episodes/rollouts.jsonl is stored here as rollouts.jsonl.gz -- 2.2 MB instead of 90 MB. It was compressed with gzip -9 -n, so the archive is byte-reproducible and carries no name or timestamp in its header:

gunzip rollouts.jsonl.gz        # restores the original 89,929,126-byte file

The uncompressed file's SHA-256 is 7a05939a35f5c9b5a2ee953c4e0c78105774a9e49da7510d67200cb0e3b1b956. That is exactly the digest recorded for episodes\rollouts.jsonl under source_artifact_hashes in supplementary/protocol-calibration/rendered/render-manifest.json, so the recorded provenance still verifies: decompress, then compare. The other seven entries of source_artifact_hashes match the shipped tree directly.

The renderer does not read this file: --render-only rebuilds every reported number from provenance/cluster_summary.json, provenance/paired_contrasts.json, provenance/headline.json, and metrics.json. It is shipped so that a reader can verify the rendered tables without rerunning the study. render-manifest.json is itself an unedited render artifact from the original run, so it is left exactly as produced rather than rewritten to match a repackaged tree.

What is not included

• Complete formal trace archives. The nominal (3,600-episode) and stress (1,920-episode) traces are multi-gigabyte artifacts. Their identifiers, byte sizes, and SHA-256 values are recorded in reproducibility/formal-artifact-manifest.json; the traces themselves are not redistributed. The paths recorded there are local workspace references, not public download links, so complete external rerun is not offered.
• Raw incident text. The descriptor inputs are derived from external held-out text records that remain subject to upstream licensing and institutional data-responsibility requirements.
• Geographic or scene data. No geographic coordinates, GIS provider records, or raw scene data are used as UAV ground truth.

Citation

reproducibility/CITATION.cff carries the machine-readable record: the manuscript title, the three authors with their ORCIDs, and the release version. If you use these materials, please cite the associated IEEE Access manuscript and retain the protocol identifiers and the evidence-boundary statement. A DOI will be added here on publication.

License

See reproducibility/LICENSE. The release is offered for scholarly verification and non-commercial research, on condition that authorship, the provenance manifests, and the stated evidence boundaries are retained. It does not grant rights to redistribute restricted incident-text inputs, third-party IEEE template files, upstream datasets, geographic-provider records, or model checkpoints. The software and artifacts carry no warranty and provide no flight-safety, operational, localization, geographic-reconstruction, or deployment guarantees.