# Audio to MIDI offline: current-build bootstrap inspection

New material observed2026-11-18: a static source dependency map, anonymous public-file response/SHA/cache-header audit, exact loader-failure branches and an offline preparation worksheet. No old guitar/piano note statistics were retabulated, and no recording, audio conversion, model execution, actual disconnected launch, cache clear or browser setting change was performed. External products were documented, not installed or benchmarked.

bootstrap-audit.json contains11 named source/UI/runtime files and their raw local/response bytes, SHA256 and allowlisted public response headers. Runtime files match the local public dist byte-for-byte; HTML may differ only by separately normalized hosting instrumentation. asset-manifest.csv is the human-readable companion. loader-excerpts.txt and source-scan.json tie claims to source hashes and five explicitly named source files; literal-pattern absence is not exhaustive dynamic analysis. load-failure-map.json records code branches, not observed offline screenshots; a rejected fetch may bypass a custom non-OK-status error.

The model manifest and weights have distinct paths. WASM variants are alternatives in this inventory, not a claim that all variants download in one browser run. Worker+model+one-variant sums are raw file-byte arithmetic and exclude UI shell, encoding/transport overhead, cache hits, memory and actual request traces. HTTP cache headers do not prove data are present or persistent on any particular user device. In-worker promise reuse is not a persistent offline-cache contract. Cold/warm/restarted offline behavior is explicitly untested.

## Reproduce without recordings or model execution
Use Node with built-in fetch. Place the published current source files under a local dist directory while preserving paths in source-index.json; compare their SHA256 to the index. Download audit.mjs. Run only the auditor:

    node audit.mjs --root ./dist --out ./local-report

To also check the declared public static-file URLs, with no login, cookies, audio upload or model invocation:

    node audit.mjs --root ./dist --out ./public-report --live

The auditor reads source strings/JSON, hashes files and optionally GETs the named public assets; it never imports/evaluates the transcription worker or instantiates WebAssembly. Live response headers/date can change, so retain both reports instead of assuming later cache headers or website versions will match. There is no instruction to run a recording, inference, MIDI conversion or another accepted12/4-case suite.

## Sources and interpretation
Original demand bodies were read for qualitative offline interest only; the black-MIDI requester has an additional musical goal outside this page, and an old Samplab comment is not a current product assertion. documentation-evidence.json attributes current publisher/API statements and distinguishes install/bundled-model/first-download needs. Several reviewed tools already claim offline use; this is not an exclusive-function claim. Current Google order could not be observed because interactive access failed. research-review.json separates public discovery and primary-body review from a dated core-family Google background record, with unknown sources not treated as absent. Exact US volume/competition/CPC remain unknown.

## Internal editorial audit
The complete main prose contains28 sentences;12 are new useful offline-specific source/file findings,42.8571428571%. Rejecting two gives10/28,35.7142857143%, above the30% minimum. Attributed documentation, necessary advice and limitations remain in the denominator. This internal audit is not a Google score, and does not reuse yesterday's43.33% as a pass. The independent topic-specific increment gate is stored with the task.
