saturn-m

This corpus contains no real data. No benchmark file is opened at any point in the build. It is generated end to end from two synthetic processes and a seeded RNG, so window-level leakage against any evaluation suite is structurally impossible rather than audited away — verified by tracing every import reachable from the builder.

Stated precisely, because the stronger version would be wrong: the corpus has no benchmark data contact, but it does have benchmark statistics contact. Its design constants were calibrated against published measurements of real suites — the quantiser targets SMAP's measured frac_repeat of 0.97, the low-rank stratum targets C-MAPSS FD00x's measured corr_erank_frac of 0.11, and gen_broad's feature marginals were tuned for coverage against 13 benchmarks. That is a real, if weak, information channel from the evaluation suite into the corpus, and anyone using this data for an OOD claim should know about it.

What it is

600.0M distinct patch tokens (4,207,591 windows, 2,266,149 panels), stratified across eleven dimensions: domain surrogate, frequency, channel count, sampling regularity, dynamics, noise, stationarity, scale, spatiality, events and sensor quality. The corpus is defined by the variation it spans, not by the sources it draws from -- every stratum label below is a coordinate a downstream user can select on.

Schema

One row per PANEL. target is list<list> with shape (n_channels, length): the panel is the unit of storage, so the multivariate grouping survives any reordering. channel_index is the explicit 0..C-1 index and group_id names the generating panel. This is deliberately NOT CauKer2M's ['start','target'] layout, in which the grouping is recoverable only as index // F.

column	meaning
target	list<list>, shape (n_channels, length)
channel_index	explicit 0..C-1
group_id	{corpus}:{seed} -- names the generating PANEL, so a panel's channels share one id. Distinct ids are fewer than rows; never use row position to group
n_channels, length	panel shape
backbone	gen_broad (eleven dynamical families) or cauker (causal DAG over GP nodes)
family	one of the eleven dynamical families
obs_model	clean / quantised / heavy_tail / missing / regime
erank_target, corr_erank_frac	target and measured effective-rank fraction
L_gen, rate_hz, irregular, acf1_target	generation length, nominal sampling rate, irregular-sampling flag, target persistence
K, max_parents	latent bases (gen_broad) / DAG density (CauKer); -1 where not applicable

Generators

• gen_broad (experiments/exp-22/gen_broad.py) supplies the eleven dynamical families, driven here at a forced (T, C, family, K) rather than through its public API, which draws all four internally.
• CauKer (ed006705ca, github.com/ShifengXIE/CauKer) supplies a causal DAG over GP-valued nodes, confined to panels of >= 14 channels -- below that it has no cross-channel structure to contribute and its cost per token is up to 84x higher.
• A labelled observation-model post-process supplies what a Gaussian process cannot: repeated values, heavy tails, missingness and regime changes. Note that gen_broad also applies its own random, unlabelled behaviours internally; this layer is applied on top and is the one the obs_model column names.

Build gates

gate	threshold	measured	verdict

Channel balance: effective nan levels spanning C=None..None.

Reproducibility

Seed 3405, CauKer commit ed006705cabbc5f0957da6de945e911cda30a387. Every panel is regenerable from its recorded stratum coordinate and seed.

Caveats

• Coverage is measured marginally (a benchmark median inside the corpus p5-p95, one feature at a time) and is therefore an upper bound, never "85% of the benchmark distribution is reproduced".
• Irregular sampling is realised as non-uniform decimation; the pack format carries no timestamp axis, so a model sees a varying effective dt but not a timestamp.
• Window length is capped at 512 samples by the evaluation context. The seconds-to-years scale axis is carried by rate_hz, not by sample count.