Add ICLR 2027 release documentation

versions/iclr2027-v1.0rc2/README.md

@@ -0,0 +1,49 @@
+# SparseWake ICLR 2027 release: 1.0rc2
+
+This version extends the historical single-source benchmark with controlled source mixtures and common-prior nearest-source tasks. It provides processed datasets, a frozen wake-library input, Python generation and analysis code, and reference training/evaluation tools. The historical May 2026 release remains at the repository root.
+
+## Downloads
+
+| Archive | Bytes |
+|---|---:|
+| [SparseWake_core_1.0rc2.zip](SparseWake_core_1.0rc2.zip) | 152,967,917 |
+| [SparseWake_review_supplement_1.0rc2.zip](SparseWake_review_supplement_1.0rc2.zip) | 6,665,864 |
+| [SparseWake_reproduction_assets_1.0rc2.zip](SparseWake_reproduction_assets_1.0rc2.zip) | 2,994,288,349 |
+
+The core archive contains seven complete HDF5 banks and six reference checkpoints. The small review supplement contains three sample banks for integration checks. The reproduction archive contains generation code, the frozen wake library, auxiliary datasets, historical experiments, and saved evidence. Start with the core archive; download reproduction assets for generation and extended studies.
+
+Verify downloads using [SHA256SUMS.txt](SHA256SUMS.txt); [ARCHIVES.json](ARCHIVES.json) also records exact sizes and SHA-256 values. Each archive contains its own MANIFEST.json for individual files.
+
+## Extract and run
+
+Extract each archive into its corresponding directory so companion paths remain clear:
+
+```text
+release/
+  core/
+  review_supplement/
+  reproduction_assets/
+```
+
+These directory names are extraction destinations; each ZIP contains package contents without an extra outer directory. Read core/README.md and reproduction_assets/GENERATION.md after extraction. Example from the core directory, using an environment with NumPy, h5py, and PyTorch:
+
+```bash
+python benchmark.py verify --data data/common.h5
+python benchmark.py train --data data/common.h5 --out smoke_common --smoke --epochs 2
+```
+
+The core utilities are portable. Historical generation campaign drivers use Windows APIs; their scope and commands are documented inside the reproduction archive. Smoke data and scores are integration checks, not the full benchmark results.
+
+The files use HDF5, NumPy, and other native scientific containers. Download and extract them to use the included readers; automatic Dataset Viewer support is not promised. ZIP files are downloadable packages and are not automatically expanded into repository folders.
+
+## Physical basis and scope
+
+The reduced-order forward model is based on Zhou, Seo, and Mittal (2025), *Effect of hydrodynamic wakes in dynamical models of large-scale fish schools*, Physics of Fluids 37, 011912, https://doi.org/10.1063/5.0250013. Benchmark additions include local probe observations, cyclic histories, explicit targets, grouped splits, multi-source tasks, and reference estimators.
+
+Source poses are fixed within each history while prescribed wake phases advance. This is a synthetic sensing benchmark with prescribed-source flow superposition. The release does not include the complete upstream MATLAB schooling application or DNS solver, and it does not claim complete reproduction of free-swimming schooling dynamics. Third-party citations and license notices are included in every relevant package.
+
+## License and versioning
+
+Data and result tables: CC BY 4.0. New original code: MIT. Historical utilities and third-party components retain their own notices. The licenses inside the archives define the relevant scope.
+
+Cite release iclr2027-v1.0rc2 and the final immutable repository commit used for download. The archive checksums identify the frozen scientific release. Packaging and anonymization preserve the numerical data and model weights; earlier run hashes may identify earlier metadata/code snapshots. Some internal documentation records the local preparation state before hosting; this page provides the version download entry point.