# Audio-to-MIDI velocity: archived-data analysis

The real recordings, native browser conversions, raw MIDI, screenshots, references and source permissions were produced/reviewed2026-11-16. This page and its new descriptive tables were prepared2026-11-17 from those unchanged files. No recording, model, audio conversion, scoring rerun or browser conversion was performed for this article. Existing matches are reused, not recomputed.

## Files
velocity-analysis.json: distributions, archived match-label splits, fixed-reference context joins and reference-velocity arithmetic. note-velocity-ledger.csv: all311 stored notes, with original strict-match flags. guitar-context-velocity.csv:16 shared existing reference attacks. piano-reference-velocity.csv:21 existing piano matches. piano-pair-ordering.csv:204 pairs with unequal reference values. The>=90 cutoff is descriptive and post-hoc, not a trained or recommended deletion rule. The flatten-to64 count is a counterfactual count, no edited MIDI was created. Raw-unit/reference-order differences are not calibrated acoustic loudness errors or overall accuracy.

## Reproduce tables without rerunning recordings
Download analyze.py and source-index.json into an empty directory. Download each entry under source-index.json files from publicPrefix+file into an inputs directory, and verify its SHA256 against the index. These small analysis inputs are the old native MIDI, saved CSV/JSON and attribution; audio and screenshots under optionalHistoricalMedia are optional. With Python3 and no external packages:

    python analyze.py --input-dir inputs --output-dir rebuilt

The script only parses stored files; it does not decode audio, transcribe, invoke a model or open a browser. It asserts the saved MIDI SHA and CSV agreement with the original native receipt. Input order is fixed, medians use ordinary sorted middle values, count stats use every stored event, and existing match indexes partition those rows. Guitar correspondence joins original reference-source labels, pitch and originalStart (nine decimal representation), with no new onset/pitch matching. Piano pair order ignores equal reference values and separately counts equal export values.

## Scope
The original strict ledger uses exact pitch and50ms onset matching, with cut-in notes and annotation/alignment limits retained. Reference source labels are not instrument recognition. Original key-strike MIDI and model amplitude are not assumed calibrated. The existing browser/offline piano velocity difference remains disclosed in the original archive. Source comparison reuses the dated real Google core-family order and already-read bodies; current ranking for this velocity phrase is unknown. No unknown body is counted as absence, no global exclusivity/function superiority claim.

## Editorial audit
Complete main prose13/30=43.3333333333% new useful editorial-analysis sentences; reject2 gives11/30=36.6666666667%, still above the current30% minimum. Old experiment facts and previous article's41.94% ratio are not counted as new. sentence-ledger.json and claims.json retain the full denominator, methods, advice and limits. This is not a Google score.
