# Reproduce the saved-note evidence

The four browser-*.mid files are actual public Chrome SaveLinkAs downloads, not generated substitutes. Browser-*ui.txt and browser-*.jpg retain the UI evidence. Offline-*.mid are separate NodeVM/WASM results; baseline-*.mid use the unchanged original single-pitch detector. Piano differs by one velocity byte between browser and offline, so do not expect all cross-runtime hashes to agree.

## Counts and plots without a model or external package

Download this folder, then run Python3: `python score.py` and `python compare-runtime.py`. Both use only the standard library plus the included midi_reader.py. score.py reparses the actual SMF files, redoes exact-pitch/onset<=50ms one-to-one matching against ref-*.json, and emits note/matchCSV plus event-derivedSVG piano rolls. It never fits an offset or grants octave credit. Compare the printed values with findings.json/result-table.csv and inspect the retained raw notes.

## Original human inputs

See ATTRIBUTION.md, both source-provenance.json files, guitar-original.jams and piano-original-reference.mid. The full original audio is available from the cited dataset/archives with the exact filenames andSHA values in protocol.json and provenance. To regenerate clips from those originals, place them at the explicit D:/JOBS source paths from protocol.json and run `python generate.py D:/JOBS/polyphonic-fresh-inputs`. This is deterministic standard-library cropping/rounding/summing; it does not synthesize the instruments. Match each WAV/ref SHA to inputs.json and the manifest.

## Engine reproduction, separately from browser testing

Use the exact script.js and self-hosted assets/converter-v2 tree from https://audiotomidinow.com/ in a local site-dist folder. Source/model versions and SHA are pinned in protocol.json/offline-conversion.json; never silently accept changed software. Run Node24 or compatible: `node rerun-worker.mjs D:/path/to/downloaded-evidence D:/path/to/site-dist`. The included NodeVM adapter executes the unmodified worker with local-file model/WASM fetches and the archived original writer; this is offline arithmetic, not a browser test. The author repeated generation and worker execution in an independent directory: all26 recorded deterministic artifacts matched exactly, while the actual browser/VM piano velocity difference is preserved separately. Full browser repetition requires selecting each provided WAV in Local note transcription/Balanced and saving the actual MIDI; do not copy offline files into browser filenames.

The two source performances are unrelated. Key-held reference ends omit acoustic pedal decay; carried-in notes and guitar annotation uncertainty limit conclusions. Source labels attached by nearest-reference matching are not instrument identities predicted by the model. No overall accuracy, live latency, source-separator quality, user conversion rate, or SEO effect was measured.
