Changelog¶

1.0.0 — 2026-10-20¶

NSMPP: the per-event NLL includes the compensator term (omitted from nll_matrix in 0.1.0).
THP-GMM, RMTPP-GMM: the temporal log-likelihood term is no longer dropped from the reported NLL.
RMTPP: the compensator uses expm1; float32 1 - exp(-x) underflowed for small decay rates.
AutoSTPP, DeepSTPP: targets the lookback window cannot score are excluded from the next-event NLL instead of entering as 0; training.loss_target_scope="last" scores one common target per context.
Rollout coherence is reported as a distance (lower is better) and deprecated in favour of the energy scores.
Intensity recovery is evaluated on all test sequences, not only the first.
HPO config sanitiser no longer collapses model.support and model.grid vectors.
config/tuning.py imports Optional.

EventCapabilities.nll_path and nll_label distinguish closed-form, numerical-quadrature (NSMPP, THP-GMM) and sample-based likelihoods; DSTPP's benchmark value is labelled a joint sample-KDE estimate, not a variational bound.
Intensity thinning requires declared proposal bounds and emits a DeprecationWarning for predictive metrics; factorized GMM presets sample by inverse temporal hazard.
njsde is documented as the Neural-STPP jump-ODE model with a conditional-GMM decoder (alias neural_stpp_gmm).
Run ids carry a process tag; save() writes the effective config and norm_stats.json.
Prepared-data and campaign roots come from SEAHORSE_* environment variables; no default paths.
Observation metadata (ids, window bounds, time axis) is carried through datasets and CLI evaluation.

Exact compensator-based evaluation (seahorse.evaluation.v1): time-rescaling PIT, threshold-weighted CRPS, HPD-PIT, energy scores, inverse-transform samplers, 0.1.0 checkpoint compatibility loader.
Calibration metrics from the NLL compensator (temporal PIT/CRPS, spatial HPD-PIT): seahorse campaign calibrate.
seahorse campaign: frozen campaign compilation, execution, Slurm dispatch, numerical NLL checks, common target sets, campaign specs under experiments/campaign_2026/.
Statistical Poisson/Hawkes presets (NumPy/SciPy, vendored TheHawkesPackage, MIT) and the prepared: dataset protocol.
Parametric static_hawkes and gaussian_poisson baselines; dnsk and glide presets (available; not benchmarked in the paper).
Author-baseline adapters (ETAS, WSM, TriTPP) with pinned provenance.
Released Hugging Face datasets loadable by short name with pinned revisions and card-hash verification (seahorse.data.hf).
Cross-route NLL consistency test; temporal RMSE metric.
CI (pytest on 3.10–3.12, ruff) and PyPI trusted-publishing workflows.

Unpublished in-house presets and the marked-process presets are not part of the release.

seahorse/models/spatial_models/independent_cnf.py is attributed to neural_stpp (CC BY-NC 4.0); the package licence remains Apache-2.0.

Results produced with 0.1.0 for NSMPP, THP-GMM, RMTPP-GMM, AutoSTPP and DeepSTPP should be recomputed.