# Demo Selection Record - 2026-09-21

Five songs were reselected from existing frozen inference outputs for generation, ground truth, 32/64-bar continuations, controls, and action sequences. This involved selection, export, and rendering only, with no retraining, new inference, parameter changes, or note edits. These recordings are not metric-tuning variants; no variants from `experiments/metric_tuning/` or its registry were used.

## Before and After

Repetition is the Jaccard overlap of MIDI onset sets across adjacent eight-bar phrases, aligned by role, relative beat, and pitch. Only generated segments are included. Lower overlap indicates less literal repetition, not better musical quality. Empty phrase pairs count as 100%. Measurements follow actual MIDI export semantics, including independent sustained notes exported as new notes. Exported events were checked against their source grids.

| Scope | Previous five-song mean | Current five-song mean |
|---|---:|---:|
| Autonomous 32 bars / arrangement | 95.72% | 88.07% |
| Autonomous 32 bars / melody | 97.97% | 90.78% |
| Autonomous 64 bars / arrangement | 87.45% | 79.25% |
| Autonomous 64 bars / melody | 88.01% | 80.83% |
| All 80 trajectories / first 32 bars | 85.69% | 82.45% |

## Current Examples

| Name | Frozen window | BPM | Repeat 32 | Repeat 64 | Melody repeat 32 | Window control / 100 | Checks |
|---|---|---:|---:|---:|---:|---:|---:|
| Fast Rhythm | Track01527.npz:112 | 160 | 87.48% | 77.95% | 89.88% | 87.41 | 16/16 |
| Melody Progression | Track01631.npz:32 | 102 | 85.64% | 72.31% | 94.38% | 83.76 | 16/16 |
| Steady Beat | Track01657.npz:25 | 120 | 92.85% | 79.81% | 93.45% | 87.55 | 16/16 |
| Bright Groove | Track01646.npz:73 | 120 | 87.43% | 86.03% | 87.17% | 85.35 | 16/16 |
| Light Ensemble | Track01672.npz:16 | 126 | 86.95% | 80.13% | 89.00% | 86.95 | 16/16 |

## Selection Scope and Rules

The initial audit covered autonomous continuations from 1,760 windows across 233 songs in the original full validation run. All 16 control configurations retained their original parameters. A total of 248 windows passed all 11 control groups and 16 individual checks, had a melody lead and nonempty restoration targets, achieved a mean window control score of at least 0.8, contained no silent two-bar block across 64 bars, had at least eight melody onsets per eight-bar phrase, and had a last/first phrase arrangement onset ratio between 0.65 and 1.5.

Additional checks used actual MIDI export semantics: guidance strength 1.0 precision and recall at least 80%; alternating octave precision at least 75% and recall at least 80%; all guidance/restoration background IoU at least 75%; and all four restoration F1 scores at least 80%. A total of 77 windows passed. Lower guidance strengths retain parts of the original melody, so their precision against the complete new target is not treated as a failure at full strength.

Eligible windows were sorted by the maximum of four repetition values: autonomous 32-bar arrangement and melody overlap, and late autonomous 64-bar arrangement and melody overlap. Ties used descending control score and then window ID. Five windows were selected with at most one per song. The current presentation order is listed above; the selected windows and recordings are unchanged.

The three fixed-density settings were also checked for all 248 initially eligible windows in `density_repetition_audit.json`. These versions often retain strong loops, so passing density control does not establish natural phrase development. This retrospective qualitative selection was not preregistered and does not replace the original aggregate paper metrics. Historical song means and rankings retain their original values; the displayed control scores refer to the selected windows.

## Control and Audio Verification

- All 80 trajectory checks and 55 control-group checks passed; the previous five songs passed 76/80 trajectory checks.
- Four MIDI guidance strengths show increasing target following. Restoration targets are nonempty, target recall and background preservation are checked, and muted intervals contain no crossing target-role notes.
- Actual MIDI onsets per eight bars satisfy low < medium < high density. Two-bar budgets match the existing targets.
- Exported MIDI onset events match the normal export of the original grids, with no splicing or inference-output edits.
- Ground truth is the actual next 32 bars, without codec reconstruction or future information supplied to the generator.
- Versions of each song share the same eight-bar prelude. MIDI prefixes match, and lossless audio prefixes match sample by sample except the 0.1 seconds before handoff.
- Recordings use the original SoundFont, 44.1 kHz, common gain 1.0, and no per-clip normalization. The 180 GT/LMWM MP3/MIDI hashes are verifiable.
- The previous five-song archive's 180 MP3/MIDI files remain unchanged. Base checkpoint hashes match the original inference protocol.

## Limitations

Recurring motifs remain in autonomous generation. Fixed-density controls still loop; per-song repetition measurements remain in the manifest. They demonstrate density execution, not natural musical development. Muting and density use output constraints, while restoration uses explicit MIDI guidance. Background preservation does not establish harmonic adaptation. No blind listening or subjective quality scores were collected.

BHC and ablation slots remain labelled previews and are outside the current sample replacement.

## Files and Reproduction

- `manifest.json`: 90 GT/LMWM recordings, trajectory metrics, repetition measurements, and input/output hashes.
- `selection.json`: complete selection thresholds and per-window measurements before and after selection.
- `ground-truth.json`: sources, hashes, and alignment checks for the actual future.
- `historical_song_ranking.csv`: unchanged ranking of all 233 original songs.
- Repository directory `reports/demo_reselection_20260921/`: the 1,760-window audit, 248-window density checks, configuration comparisons, original inference protocol, old page snapshots, and preservation checks.
- Repository directory `demo/five-song-20260920/`: original audio, MIDI, manifests, and rankings at their original paths.

Selection used `scripts.audit_demo_repetition`, `scripts.audit_demo_control_candidates`, and `scripts.reselect_demo_low_repetition`; the original report used `scripts.report_demo_reselection`. Re-exporting requires new `--audit` and `--output` directories.

The final presentation revision translates the demo into English and reorders existing examples. Baseline MP3s are trimmed from archived WAV masters to the last generated note plus up to 1.5 seconds of release tail. No MIDI notes are changed. Updated audio hashes and durations are in `baseline/manifest.json` and `baseline/native_length_verification.json`.