# Keep edits or Discard edits: actual MIDI downloads

These four files were newly downloaded on October 19, 2026 (Asia/Shanghai) after an actual browser interaction with an isolated local preview. That preview used the exact deployed build at source commit `eeb011cf0e3dd0394946b50a9020fbc700b696da`. The engine was not changed. These are actual downloads from that run, not manually generated replacement MIDI files or a production-user-audio test.

## Reused input and settings

The input is this site's original CC0 artificial sine recording, reused from the September 26 fixture:

- [reference.wav](https://getmp3tomidi.com/examples/audio-to-midi/reference.wav)
- SHA-256: `14c6538c8e5861e7a6fc99b9036cfb99d4033331e26fbb21db768217c3855c35`
- [Original fixture method and reuse permission](https://getmp3tomidi.com/examples/audio-to-midi/README.txt)

This was a fresh execution using old audio; the recording was not newly generated. It is not a singer, full song or microphone recording. The eight-note original output happens to match the earlier archived output's bytes; its provenance in this collection is the fresh browser download.

## Repeat the UI sequence

1. Use the byte-matched source build above in an isolated browser preview. A later version of the public site may behave differently. Open the converter and choose **Melody before selecting the file**. Choose the linked `reference.wav`. Use quiet-note filter **−40 dB**, minimum note length **40 ms** and **Transpose 0**.
2. After analysis, **Download MIDI**. Save that unedited download as `01-original.mid`: eight notes.
3. In **Fix individual notes**, select the first note: MIDI pitch 60, approximately 0.000–0.182 seconds (ticks 0–350). Choose **Delete note**. Download as `02-deleted.mid`: seven notes. The second pitch-60 note is a different event and remains.
4. Change the global **Transpose** setting from 0 to +1. In **Replace your edited notes?**, choose **Keep edits**. The pending global change is canceled, Transpose remains 0 and the deletion stays. Download as `03-kept.mid`: seven notes, byte-identical to `02-deleted.mid`.
5. Request global Transpose +1 again. This time choose **Discard edits**. The original analysis notes are rebuilt and the pending +1 change is applied. Download as `04-discarded-plus1.mid`: eight notes, with each pitch one semitone above `01-original.mid`. This branch does not simply transpose the remaining seven edited notes.

Keep your own original and edited copies before trying replacement controls. The `Discard edits` branch deliberately removes the individual edit; this example is not a promise of undoing every action in every future build.

## Check the downloaded collection without writing files

Download all seven files into the same directory. Python 3 is sufficient; no package installation is required:

```sh
python3 verify.py
```

The script uses a separate strict Standard MIDI File byte parser. It reads the four MIDI files and `measurements.json`; it does not import the converter, its MIDI writer, a generator or Mido. It performs no network calls, browser actions or file writes. Its JSON result goes to standard output. Expected hashes and note events are fixed in the verifier rather than obtained only from the measurements table.

It checks the 8/7/7/8 note counts, first-note deletion, exact equality of files 02 and 03, and the original eight notes plus one semitone in file 04. Every file is Type 0, one track, 960 PPQ, tempo 500,000 microseconds per quarter note, and End-of-Track at tick 3686. All note channels are raw channel 0 (displayed channel 1), Note On velocities are 89 and Note Off velocities are 0. Starts, ends, velocities, program and clock stay unchanged in the stated comparisons. The 120 BPM event is this exporter's time base, not detected song tempo. Timing in seconds follows this fixed tempo; tiny floating-point differences are compared within 1e-12 seconds.

## What this does and does not show

This is an edit-confirmation and export-handling check for one synthetic input and one deployed build. It does not measure recognition accuracy, physical-key timing, DAW compatibility, subjective listening or behavior on every input/device. It is separate from the October 6 singing-recording example in the guide. No new tool or recognition feature was added. The engine, models and quality thresholds were unchanged.

The original synthetic fixture, these observed synthetic MIDI outputs and this standalone verifier may be reused and redistributed under CC0-1.0. Browser session snapshots, local private paths and models are not part of this portable collection.
