# Recorded drums-to-MIDI local result, 2026-11-19

One original acoustic bass-drum recording, one unchanged public-worker NodeVM/WASM execution, Balanced mode2. No accepted12/4 input was reopened. No other mode, sample or baseline-detector inference was run. Browser file selection failed before conversion; no actual browser MIDI was downloaded. Google showed a CAPTCHA, not an observed organic order. Fixed-rival current bodies and original qualitative user questions are recorded in research-review.json; do not mistake them for Google top3/top10.

Input attribution and redistribution terms are in ATTRIBUTION.md and source-provenance.json; the originalOGG is CC BY-SA3.0. PCM representation, waveform/annotation derivatives and related plots retain CC BY-SA3.0 attribution; no endorsement. Source recording2005-03-31, original acquisition2026-11-17, analysis/model execution2026-11-19. The untouched actual-local-kick.mid is the real local-model output written by the production writer,161bytes; editorial-channel10-only.mid is a clearly labeled manual negative control, not another inference or download.

Timing-proxy count8, saved events15, nearest-unused reference-order50ms matches4, missed proxies4, unmatched outputs11. No fitted shift. Waveform proxies are not performer/hardware ground truth. All calculations use the same FFmpegPCM axis; native browser decoding is unverified and differs in reported duration from the source page by about20ms. There is no universal accuracy rate or full-kit classification score. GeneralMIDI target35/36 on human channel10 is a declared receiving map, not every custom sampler.

## Recalculate saved-file findings without inference
Place the provided MIDI, reference.json, midi_reader.py and Python analysis scripts in one directory. Run:

    python analyze-local-result.py
    python channel-only-control.py

These commands parse saved files, compute pairs and draw plots; neither decodes another audio file nor starts a model. No externalPython dependencies are needed. Numerical pairs and full note ledger are also in onset-pairs.csv/output-notes.csv. source-inspection-protocol.json and inspect-source.py record the original one-time FFmpeg analysis representation; do not run it as a daily gate. The initial200ms grouping failure remains preserved.

## Local model provenance, not a frontend replay
local-runtime-protocol.json, local-model-attempt.json, local-model-result.json and run-single-local-worker.mjs record the single execution. Current worker/model/WASM public paths plus SHA are in local-model-result.json; source functions are in current-pipeline-excerpt.js. The runner uses only the unchanged published bundle and local-file fetch adapter; it contains a single-attempt guard. It has WindowsD paths matching the actual experiment and is supplied as execution provenance, not falsely described as a portable browser test. Independent readers can reconstruct the same declared local environment from those public pinned assets; the executor must not rerun this closed sample to supply another hash or screenshot.

## Publication/review scope
The article follows the accepted static layout, with unobserved new native desktop/mobile render and unavailable current Google order retained as gaps. Full-prose internal originality 12/26 is bookkeeping, not an accuracy score, a Google metric or coach acceptance. Every complete prose claim has a near-neighbor file/primary reference and a sentence ledger.
