# Darker: Roland music with Roland GM-bank percussion

The native-timing exports below are 48 kHz, 24-bit stereo FLAC, using the original LAPC-I sequencer and custom instruments. Missing percussion is supplied by a second Munt instance loaded with Roland MTGM.MID. No SoundFont, fades, normalisation, filtering or added release tails.

## Hand-corrected opening mixes

These team-supplied edits are retained unchanged as research references for a
proposed musical alignment, separately from the native-timing exports:

| Recording | Duration | Format |
| --- | ---: | --- |
| [Corrected intro](group-1-corrected-intro.flac) | 139.359 s | 48 kHz, stereo, 16-bit FLAC |
| [Corrected loop](group-1-corrected-loop.flac) | 139.359 s | 48 kHz, stereo, 16-bit FLAC |

The exact edit operations have not been independently reconstructed. These files
are not covered by the native render's no-processing or PCM-verification claims.
Their hashes and metadata are recorded under `corrected_research_mixes` in the
[manifest](manifest.json). See the [full Roland investigation](../../../docs/music-hardware-variants.md#roland-gm-bank-experiments)
for percussion-bank experiments, loop evidence and the interpretation of intent.

## Opening piece: two parallel stems

Start **both intro files at time zero**. Each includes its original leading silence. Follow each intro immediately with its matching loop file, repeating that loop independently. The two intros have different lengths; do not wait for both to finish before starting either loop. No manual offset is needed.

Track2 is originally silent until roughly 16.56 seconds; that silence is present in its intro file. Its later notes are retained.

| Stem | Intro | Loop |
| --- | ---: | ---: |
| track1 | [22.376625 s](group-1-track1-intro.flac) | [58.491188 s](group-1-track1-loop.flac) |
| track2 | [24.816271 s](group-1-track2-intro.flac) | [56.051542 s](group-1-track2-loop.flac) |

## Remaining pieces

Play each intro once, then repeat its loop. Group 6 starts directly with its loop.

| Group | Intro | Loop |
| --- | ---: | ---: |
| 2 | [0.679917 s](group-2-intro.flac) | [32.015167 s](group-2-loop.flac) |
| 3 | [0.659917 s](group-3-intro.flac) | [31.315271 s](group-3-loop.flac) |
| 4 | [0.679917 s](group-4-intro.flac) | [32.015167 s](group-4-loop.flac) |
| 5 | [0.679917 s](group-5-intro.flac) | [32.015167 s](group-5-loop.flac) |
| 6 | None | [38.434208 s](group-6-loop.flac) |

## Timing and provenance

Sections were cut at native sequencer event boundaries, not arbitrary durations. They are continuous first-pass renders with no synthesiser reset at the split. Repeating captured PCM preserves that captured pass rather than recreating live oscillator/reverb history or the original timer’s changing interrupt phase. The opening stems were rendered in isolation, so their sum need not be bit-identical to shared-device playback.

See [manifest.json](manifest.json) for source hashes, original part numbers, exact sample boundaries and verification statistics. Every native-timing FLAC was decoded and checked against the source float PCM to within one 24-bit quantisation step.

## Original driver verification

[Native execution trace](native-loop-trace.json) records the original LAPC-I
driver's loop-marker and jump handlers, independently of the reconstruction.
Both track1 and track2 first jump at scheduler call 4043 (80.867799 seconds).
The next jumps are call 6968 for track1 and 6846 for track2. The native loop
handler loads the saved stream pointer without inserting a compensating delay.

The matching intro melody is present three times in the source: notes 65, 72,
68 start at ticks 2208/2304/2400 in original part 8, 2400/2496/2592 in part 9,
and 2592/2688/2784 in part 10 (track2). These staggered entrances explain why
moving track2 earlier makes it coincide with the melody in track1. This establishes
the encoded timing, not whether the unequal repeat lengths were intentional.