Learning When to Ask: Communication-Constrained UAV Energy Monitoring with Certified Scheduling

Manuscript and reproducible computational artifact for Learning When to Ask: Communication-Constrained UAV Energy Monitoring with Certified Scheduling.

UAV-specific specialization of classical concave coverage and network duality, independently checked arithmetic certificates, and empirical evaluation on one previously inspected AtlantikSolar flight. The archive includes preregistrations, seed and commit records, input/output hashes, complete scalar results and code history. Established greedy, interval-matrix and flow tools are not claimed as novel.

The registered round-four follow-up compares LP, OR-Tools cost scaling, arbitrary-precision network simplex, greedy and MIP across 57 saved models. It includes 285 valid-design attempts and 399 process attempts counting 114 preserved import failures. Cost scaling is faster than LP in 54 of 57 jointly checked cases at original-model tolerance 1e-7, with one repetition and host-contention limitations. Exhaustive small checks cover 7,086 feasible subsets, 693 rational insertion cases and 494 loss-string cases. Negative findings on transfer, finer-grid benefit and fixed-policy power-shock replay remain explicit. All comparisons on exposed flight windows are exploratory.

One 15-page stock-LNCS PDF includes its internal appendices and references. Figures use Computer Modern and one blue accent. OpenAI assistance is disclosed in the manuscript. This is a research archive, not conference acceptance or confirmation of conference AI-policy eligibility.

Restricted dense flight data and credentials are excluded. Author code is MIT licensed, author manuscript and aggregate results CC BY 4.0, inherited notices and third-party terms retained. No battery-safety, physical-calibration, new-flight generalization or radio-energy saving is established.