# Reconstructing the fixed benchmark

`manifest.csv` identifies the 100 selected source scores by content ID (CID) and
MuseScore score ID. `attribution.json` preserves the available source links and
embedded credits. Obtain source scores from the original MetaScore distribution
or a source authorized to provide the work, respecting its terms.

Complete MIDIs in this release were converted from the corresponding MSCZ files
before selecting the fixed windows. The window builder is
`preprocess/build_metascore_test.py` in the DAMN development repository. Its
source and the serializer dependency are archived under `reconstruction/` for
inspection and reproduction. They were archived from the local development
checkout on 2026-10-31; this is not a claim that the inference-only submission
repository contains the preprocessing pipeline.

Install `reconstruction/requirements.txt`. The archived portable `config.py`
replaces deployment-specific paths with relative placeholders; window parsing
and serialization are unchanged. From `reconstruction/`, with the original
directory structure restored, including all 100 source entries, use:

```bash
python -m preprocess.build_metascore_test --source /path/to/MetaScore-SFT-100 --output /path/to/rebuilt_challenge_test
```

Use a fresh output directory (or the builder's explicit `--output` option) to
avoid replacing the released original test artifacts. The protocol is a fixed
four-bar prompt plus optional pickup and a twelve-bar continuation.

For `QmbQPgvc93q2bBKvPaVRWNJDvaxSB8kHxVKXxQD2W5W1Qx`, `rights_conflicts.json`
lists all expected source/derived paths and their SHA256 checksums. Use the
source score/MIDI to the paths in `manifest.csv` and the window manifest;
compare checksums and rebuild the specified window. All 100 entries, window
indices and selection reasons remain fixed. This release intentionally does
not provide a downloader that bypasses the original source's access controls.