Metrics and provenance
Counts, rates, latency, non-identifying device class, exact public versions, corpus identifiers, revisions, licenses, checksums, deviations, and limitations.
Public data · sanitized release
Inspect the test record, not the private content.
The public package contains metrics, provenance, and checksums. It deliberately excludes every transcript, reference, recording, prompt term, filename, path, account identifier, and raw error.
01 · Release files
Start with the index and caveats.
Latency evidence audit, 6 September 2026Timer boundaries, reproduced metrics, external sources and remaining evidence gaps.
LATENCY_AUDIT_2026-09-06.md DownloadVerification tools and checksumsDownloadable Python verifier, dependency links, commands and SHA-256 inventory.README.md DownloadLocalized Markdown reportPortable text edition in the current page language, with pilot results, related evidence, platform coverage, limitations, and next steps./research/private-dictation-2026.md DownloadEnglish archival reportStable English source edition stored beside the versioned public data release.REPORT.md DownloadPublic release indexRelease status, authorship disclosure, file paths, and SHA-256 digests.index.json DownloadMethodology 0.1.0Complete preregistration-style protocol and publication gates.METHODOLOGY.md DownloadPilot results noteCanonical interpretation, limitations, and non-comparability caveats.PILOT_RESULTS.md Download02 · Run packages
Three content-free pilot lanes.
Each run package contains the public manifest, a micro-weighted aggregate, and normalized JSONL rows. The rows identify fixtures with non-semantic public IDs and retain only counts, tags, language, latency, and integrity fields.
Core ML Parakeet, cleanup off
Run manifestProduct, model, software, device, corpus, settings, and protocol provenance.pilot-yaps-macos-m1-max-raw/run-manifest.json DownloadAggregate metricsOverall, language, and tag strata with operation counts and latency distributions.pilot-yaps-macos-m1-max-raw/aggregate.json DownloadNormalized result rowsSanitized per-run counts and timings without speech or transcript content.pilot-yaps-macos-m1-max-raw/normalized-results.jsonl DownloadWhisper Base, cleanup off
Run manifestProduct, model, software, device, corpus, settings, and protocol provenance.pilot-yaps-whisper-base-macos-m1-max-raw/run-manifest.json DownloadAggregate metricsOverall, language, and tag strata with operation counts and latency distributions.pilot-yaps-whisper-base-macos-m1-max-raw/aggregate.json DownloadNormalized result rowsSanitized per-run counts and timings without speech or transcript content.pilot-yaps-whisper-base-macos-m1-max-raw/normalized-results.jsonl DownloadCore ML Parakeet + local cleanup
Run manifestProduct, model, software, device, corpus, settings, and protocol provenance.pilot-yaps-macos-m1-max-delivered/run-manifest.json DownloadAggregate metricsOverall, language, and tag strata with operation counts and latency distributions.pilot-yaps-macos-m1-max-delivered/aggregate.json DownloadNormalized result rowsSanitized per-run counts and timings without speech or transcript content.pilot-yaps-macos-m1-max-delivered/normalized-results.jsonl Download03 · Publication boundary
What remains private by design.
Speech and private context
Audio, reference text, hypotheses, prompts, window context, filenames, paths, emails, account IDs, logs, raw errors, and machine identifiers.