# Additional benchmark details

## Selection and fixed windows

The existing selection pipeline retained 106 candidate scores satisfying:

- A single piano part (`keyboard.piano*`), at least 16 measures, no lyrics.
- The upstream best-unique-arrangement filter.
- Rating >= 4.5 with at least 3 ratings.
- The upstream 4/4 filter; the final MIDI window is checked explicitly against
  the continuation protocol.

Manual listening rejected six clear quality-control failures; their CIDs are
preserved in `summary.json`. The remaining 100 pieces span 15 genre labels.
These genres are not equally represented; see the summary for exact counts.

The deterministic window builder accounts for score pickup information,
merges piano events and duplicate `(start, pitch)` notes, and chooses the first
complete protocol-compatible 16-bar window. **99 pieces use the opening window;
one uses the first subsequent compatible window** because the opening does not
meet the protocol. `challenge_test/manifest.json` records the source SHA256,
window index, pickup, selection reason, note counts, and MIDI parsing diagnostics.
No new selection or window extraction was performed for this release.


## Format

Times use a sixteenth-note grid. A complete window ends at step 272, with steps
0--15 reserved for an optional pickup. The four regular prompt bars occupy
steps 16--79. Prompt notes have start in 0--79, and reference notes in 80--271.
Pitch is an integer in 0--127; duration is a positive integer and notes do not
extend beyond step 272. Notes are sorted by `(start, pitch)`.

```json
{"prompt": [{"start": 16, "pitch": 60, "duration": 4}]}
```

```json
{"generation": [{"start": 80, "pitch": 64, "duration": 4}]}
```

The fixed-window MIDIs use 480 ticks per quarter note, 120 BPM, program 0 and
velocity 64 for comparable listening. Original complete-piece assets retain
their original contents. Genre labels and source previews are auxiliary;
they are not additional generation conditions in the DAMN benchmark.


## Provenance, permissions and limitations

Metadata records 99 source scores as `cc-zero` and one as `publicdomain`.
These are upstream declarations, not an independent verification of rights to
every underlying composition or arrangement. Existing score credits and copyright
notices are retained, and the explicit conflicting entry is included with its original rights notice unchanged. There is
no blanket license replacing source rights; see [LICENSE.md](LICENSE.md).

The benchmark contains familiar repertoire and a small, curated selection of
genres. Source metadata can be imperfect. References are one possible
continuation rather than the only musically valid answer. Similarity to a
reference is a development proxy, not a substitute for listening evaluation.
The DAMN pipeline excludes benchmark matches from its training data; this does
not establish the absence of overlap with other models' training corpora.