update README to signpost CANONICAL/tune5 as the definitive paper results

README.md

@@ -1,4 +1,45 @@
 ---
 license: mit
 pretty_name: Seismic Ensemble
----
+---
+
+# Seismic Ensemble — Data & Model Store
+
+This repo holds the data and model artefacts for the code at
+[ese-ada-lovelace-2025/irp-jas25](https://github.com/ese-ada-lovelace-2025/irp-jas25), an Imperial College
+MSc IRP project on multi-class seismic event discrimination (earthquake / deep earthquake / explosion /
+nuclear explosion / volcanic eruption / noise) for CTBT-style monitoring, using a six-member ensemble
+(three deep models, three classical) with a learned meta-learner on top. It accompanies the paper
+*"Robust and Explainable Multi-Class Seismic Event Discrimination through Benchmarking and Ensemble
+Learning."*
+
+## Start here: `CANONICAL/tune5/`
+
+`CANONICAL/tune5/` is the definitive run behind the final report — everything you need to reproduce the
+paper's tables and figures lives here, and nothing outside it is required to do so.
+
+- **Model checkpoints**: `CANONICAL/tune5/{barama,kasburg,swaine,kong,li,pignatelli}/best_retrain/` — one
+  directory per ensemble member, each with its trained weights (`model.pt` for the three deep models,
+  `model.joblib` for the three classical ones), the config it was trained with, and its Optuna tuning
+  history under the sibling `optuna/` folder. In the paper these six go by codenames: `barama` = CNN,
+  `kasburg` = ViT, `swaine` = SWP-Net, `kong` = Phys-CNN, `li` = XGBoost, `pignatelli` = SVM.
+- **Main results**: `CANONICAL/tune5/ensemble_hand_plus_ssl/` and its calibrated counterpart
+  `ensemble_hand_plus_ssl_temperature/` are the primary ensemble reported in the paper — `ensemble_report.json`
+  in each has the full per-member and ensemble metrics (trace-level and event-level), and the confusion
+  matrices used in the results figures sit alongside it as PNGs.
+- **Everything else** in this folder (`exp_vol_*`, `exp_spatial_*`, `exp_hparam_*`, `exp_pick_perturbation`,
+  `ensemble_depth_*`, `anomaly_*`, `screening_*`, `pretrain_mae*`, `tune5_fewshot_*`,
+  `north_korea_padding_attribution`, `meta_learner_shap_comparison`) backs one specific robustness check,
+  ablation, or appendix figure in the paper — these are the raw outputs behind those sections, not
+  alternative models to pick between.
+- `CANONICAL/tune5/results/splits/` holds the exact train/val/test split used everywhere above, so any of
+  it can be reproduced end to end.
+
+## Everything else in the repo
+
+`checkpoints/`, `experiments/`, `processed/`, and `UNIFIED/` are kept for completeness and traceability —
+they're the trail of earlier iterations (`tune4_colab`, `ssl1_colab`, `exp1_colab`, and so on) that either
+fed into what became `tune5` or were tried and set aside along the way. `processed/` and `UNIFIED/` are the
+dataset itself at two stages of the pipeline (processed features vs. the full unified archive) rather than
+model outputs. If you're only after the results and checkpoints that the paper actually reports, you want
+`CANONICAL/tune5/` above.