Anytime Solver Comparison on the Duration-Minimization Time-Dependent VRPTW: Raw Experiment Data

Raw experiment data of a five-solver anytime comparison on the Time-Dependent Vehicle Routing Problem with Time Windows under duration minimization. The campaign comprises 8,800 solver runs: five arms over 212 instances at a wall-clock budget of 3,600 seconds per run, single-threaded, one run per core, with ten seeds per instance on eight of the ten instance panels and five on the two panels of 1,000 customers. The arms are KAYROS 1.6.0 in its default iterated-local-search mode, jsprit 2.0.0 on JDK 21, Timefold Solver 2.3.0, Hexaly Optimizer 15.0 on a rich piecewise-linear model through C++ external functions, and the same Hexaly solver on a 96-slice time-sliced conversion reported as descriptive only. The instance set is ten panels, two per family, over the Dabia2013, Rifki2020, Vu2020, Lera2026 and Poryos2026 families, with two panels running their full canonical set and the other eight seeded solver-blind samples. The rich-model Hexaly arm is a cell-exact re-run of the original campaign wave: the original wave declared the travel-time functions through the solver's Python binding, whose external-function callbacks are evaluated on a single worker, serializing the search. The re-run declares the same functions through a compiled C++ binding proven to evaluate identical values, over exactly the same 1,760 cells, and is the arm of record here. The superseded Python-binding records are not part of this deposit; the thread-scaling ladders documenting the two bindings are published in the companion dataset of the Hexaly thread-scaling report. Every reported value is produced by the reference checker of mamut-routing-lib on the original travel-time functions, never by a solver's own accounting. The deposit contains the per-run records with their full incumbent trajectories, the solver-native outputs kept for diagnostics, the 8,800 re-scored checker trajectories bound to their records by content hash, the per-run summary table, the complete analysis pass, the confirmatory statistics, the publication figures, the job logs, and a metadata directory holding the frozen instance manifest, the statistical pre-commitment as frozen before any result was inspected, and the reference snapshot every gap is measured against. Quality is reported as a gap against that reference snapshot and anytime behaviour as the normalized primal integral of the gap over the run, aggregated with panels weighted equally so that each family contributes one fifth of the total. The reference snapshot is post-campaign and solver-blind: every improving solution any arm found was folded into the public benchmark store before the reference was frozen and before any result was analysed, which makes gaps non-negative everywhere and makes panels referenced against proven optima poolable with panels referenced against best-known values. The rich-model re-run postdates that freeze and was separately verified to improve none of the stored references. The instances themselves are not duplicated here. They are distributed as checksummed artifacts through the MAMUT-routing benchmark platform, and every record identifies its instance by a path relative to that tree. Experiments were carried out on the Grid'5000 testbed.