Figure 16 — Artifact layout

Vector PDF

File: fig16_artifact_layout.pdf

Suggested caption. Layout of the benchmark outputs. Every unit of work writes its own Parquet file. The collect stage merges them into the published tables in tables/.

Contents

Path	Row / content
pools/<ds>/<gen>/seed<k>.parquet	pool windows (N, T, 3); .record holds training and sampling time and diagnostics
generators/<ds>/<gen>/…	generator weights (VAE, Diffusion-TS, SDV, plugins)
series/<ds>/<method>/seed<k>.parquet	generated series: t, value, hour of week, window, piece, seam flag
records/…	time, memory, hardware and diagnostics of each generation run
fidelity/…	one row with all F-metrics
labels/<ds>/<task>.parquet	real test labels (split, row, y)
utility/<ds>/<method>/seed<k>/	metrics, every prediction, weights of every fitted model, cells.json (done cells)
tables/	timing, fidelity, utility, pools, cross_seed, datasets, inventory, hardware, config.json
reports/	Markdown reports of the analysis: REPORT.md (research questions RQ0–RQ9, confirmatory family C1–C4, case study CS1–CS15), OVERALL.md (overall leaderboard), LEADERBOARD.md (ranking of every method in every slice), METRIC_VALUES.md (values of the core fidelity metrics and the primary utility metrics by reference, dataset, task, model and split)
overall/	tables of the overall leaderboard (fidelity on the core metrics and utility, noise-aware ranks) and its robustness checks
leaderboard/	tables of the ranking of every method in every slice, incl. rankings.parquet (the four rankings) and rankings_summary.parquet
metric_values/	fidelity.parquet, utility.parquet: median and IQR over seeds of every metric × reference and task × model × split, per dataset and method
hypotheses/	per-question global and per-dataset tables (RQ, CS), confirmatory family
errors.log	failed units (the run continues)

Notes for the text

• Storage: Parquet with zstd; float32 everywhere except series values (float64). Each profile writes to its own artifacts/<profile>/.
• Every model and generator can be reloaded and used without retraining (marathonts/bench/load.py), so reviewers can re-evaluate saved models on new test data.