All notable changes to this project. The format follows Keep a Changelog and the project adheres to Semantic Versioning.

Honest scientific framing + richer model surface. The headline change is that the README now leads with the metrics where the LSTM dominates Markov (top-3 +11.7 pp, MRR +0.044, ECE 0.011) instead of chasing top-1 — a property of the dataset, not the model. Three feature ablations (--use-resource, --use-temporal) plus a cross-seed variance harness confirmed the gap on top-1 is within seed noise and not feature-engineerable on BPI 2020.

• End-to-end Jupyter tutorial (notebooks/01_bpi2020_tutorial.ipynb) walking the full BPI 2020 workflow: load → analyze → baseline → read pinned metrics → explain → predict-suffix → whatif → diff. The notebook reads bench/published/bpi2020_lstm/metrics/ so a reader gets the headline numbers in 30 seconds without retraining. Source- of-truth in notebooks/build.py.
• Resource embedding in the LSTM (--use-resource). Parallel embedding concatenated with the task embedding before the LSTM. No- op on BPI 2020 (only 2 resources) but lands the infrastructure for richer logs.
• Cyclic temporal features (--use-temporal). Per-event sin/cos of day-of-week and hour-of-day as continuous LSTM input. Also no-op on BPI 2020 top-1; included for the same infrastructure reasons.
• Quantile time head (--time-quantiles "0.1,0.5,0.9"). Replaces the MSE point estimate with a K-output head trained with pinball loss; eval reports interval coverage and width alongside median MAE. "Will finish in 80 h with 80 % confidence" instead of "in 80 h."
• Cross-seed variance harness (bench/seeds.py). Runs N seeds in sequence, aggregates per-seed metrics into mean ± std / min / max per leaf metric. The number that tells you whether two leaderboard rows differ meaningfully or just by training noise.
• Property tests for the model layer (tests/test_properties.py): top-K accuracy is monotone in K; temperature scaling is argmax- invariant; ECE is in [0, 1]; LSTM forward shape is invariant under feature flags.

• README leaderboard reframed: top-3, MRR, ECE, dt MAE as the headline numbers; an honest "what we tried, what didn't help" subsection documents the BPI 2020 feature ablation.
• README CLI section now documents every v0.4+ flag and the non-run subcommands (baseline, explain, predict-suffix, whatif, diff, serve) plus the bench/seeds.py cross-seed harness.

• No backwards-incompatible changes. Every new flag defaults to off / empty / no-op so existing callers (predict-suffix, serve, the canary) are unaffected.
• The cross-seed sweep on BPI 2020 (5 × 30 epochs) lands the canonical variance numbers under bench/published/bpi2020_lstm_seeds/ — pinned for reproducibility once the sweep completes.

Reproducibility, baselines, and ergonomics. Every accuracy number now has a baseline to compare against; every full pipeline run is canary-tested in CI; XES logs work natively; both models can be calibrated and predict times.

• gnn baseline <csv> — null (most-common-next-task) + 1st-order Markov baseline; reports markov_accuracy, markov_coverage, most_common_accuracy. The full pipeline now records lift_over_markov in gat_metrics.json and lstm_metrics.json so model accuracy is compared against the trivial floor.
• gnn explain <csv> --case-id <id> [--model <path>] — runs the trained GAT on a single case with forward_with_attention, dumps per-event predictions + attended-predecessor weights as JSON and a heatmap PNG. Loads --model if given, otherwise trains a small demo GAT on the spot.
• gnn diff <run_a> <run_b> — markdown delta report for two run directories. Numbers, dicts, and clustering changes are surfaced.
• LSTM parity with the GAT: optional time-to-next-event regression head, post-hoc temperature scaling, ECE-before-after reporting. Multi-task path falls back to the Python prefix builder when --predict-time is set (Rust hot path doesn't carry dt yet).
• Per-class precision / recall / F1 / support + macro F1 + weighted F1 in both GAT and LSTM metric files.
• Reference-metrics canary (tests/test_canary.py) — runs gnn smoke --device cpu --seed 42 in a subprocess and compares every metric to tests/golden_smoke_metrics.json with per-key tolerances.
• Native XES ingest (.xes, .xes.gz) via PM4Py's importer — every public BPI-style log now works directly with gnn run log.xes.
• Markov-chain smoke generator: a real process graph with branching (Approve / Reject) and bounded loops (Reject → Resubmit → Review). Markov baseline 96.4% on the new generator, vs the random-skip generator that gave both models ~20% with no signal to learn.
• TOML --config file with two-pass argparse (CLI > TOML > defaults).
• gnn version, gnn --version.

• perform_conformance_checking now returns (replayed, summary_dict) with num_deviant, fitness, precision, f_score. Token-replay deviance count alone is misleading on flower-model logs; this surfaces the precision gap that calls them out (BPI 2020: fitness 1.0, precision 0.247).
• evaluate_gat_model accepts a temperature kwarg.
• stage_train_lstm now takes le_task so per-class metric keys are real task names instead of integer indices.

• LSTM evaluate previously emitted a "torch.tensor of list of numpy arrays is extremely slow" warning; stack via numpy first.

Methodological audit. Three correctness fixes that affected published numbers, each shipped with a regression test.

• Causal forward-only edges in build_graph_data (default). Combined with default GATConv self-loops, node j's representation depends only on {j-K, …, j-1, j} after K layers — strictly past + present. --gat-bidirectional opt-in for v0.2 reproducibility.
• Node-level GAT head (default): predict next-task at every event. --gat-graph-label opt-in for v0.2 reproducibility.
• Optional time-to-next-event regression head on the GAT (--predict-time). Joint loss, val-set MAE in hours.
• Post-hoc temperature scaling on the GAT classifier. ECE before/after recorded; --no-calibrate to skip.
• --split-mode temporal: order cases by start time, take last val_frac as val. Surfaces drift random splits hide.
• pm_fast/ Rust hot-path kernels (PyO3 + maturin). 588× / 505× speedup on build_task_adjacency and the LSTM prefix builder. Auto-detected; falls back to Python when not built.

• data_preprocessing.encode_categoricals adds dt_seconds and dt_log columns alongside next_task.

• ProcessEnv._get_state was sized by len(all_tasks) (post-dropna source-task subset) but indexed with current_task from the full label space. Tasks that only appear as terminal events crashed RL. Sized by len(le_task.classes_) now.
• BPI logs ship both case:id and case:concept:name; the previous rename map collapsed both to case_id, producing a duplicate-named column. Priority-based rename + collision drop.

First public audit. Three correctness issues that affected published numbers; full gnn CLI as the user-facing surface.

• gnn CLI with subcommands: run, analyze, cluster, smoke, version. python main.py <csv> legacy entry point preserved as a thin shim.
• Pytest suite (case-isolated split, scaler fit-on-train-only, both model forward+backward passes, RL contract, spectral on a known bipartite graph). CI on Python 3.10 + 3.11.
• modules.utils.set_seed + pick_device reproducibility helpers.

• Case-level train/val split. The previous LSTM pipeline shuffled prefixes before the 80/20 split, so future events of the same case ended up in both halves. Splits now happen on case_id first.
• Train-only scaler fit. MinMaxScaler / Normalizer are fit on training rows only.
• Normalized Laplacian + eigh for spectral clustering. The unnormalized Laplacian + np.linalg.eig returned complex eigenvectors and produced unstable clusters when task degrees were skewed.

Initial release. Process mining with GNN + LSTM + RL + spectral clustering on event logs. Single-script python main.py <csv> entry.