Audio-V product audit and capability contract

Last reviewed: 2026-07-30

This document is the product truth source between the approved vision, the original request, competing applications, and what the repository actually implements. A feature is not considered present because a mock-up displays it.

Original request baseline

The request in DRAZY/deemix-remastered discussion 128 asked for a lightweight, standalone analyzer with:

A full-track spectrogram for inspecting real frequency cutoffs.
CBR/VBR, bitrate, encoding, and sample-rate details.
Peak level, RMS, dynamic-range, and loudness measurements.
Channel layout and encoding-mode details such as stereo, joint stereo, and mono.

Those four outcomes are the minimum viable product and are implemented. Quick Inspect preserves the requested lightweight one-file workflow by placing the verdict, essential measurements, MP3 encoded channel mode, and deep-view actions on one screen. Batch folders, evidence-backed verdicts, integrity checks, comparison, and reports extend that workflow rather than making them mandatory.

The request’s phrase “100% accurate, unaltered, raw spectrogram” is treated as a request for an honest, reproducible view—not as a literal product claim. A spectrogram is an STFT-derived representation whose appearance changes with window function, FFT size, overlap, scaling, normalization, and color floor. Audio-V exposes and exports those settings, analyzes decoded samples without applying EQ or enhancement, and never implies that one rendering proves source history. Sonic Visualiser documents the same time-versus-frequency-resolution tradeoff in its spectrogram reference.

Lightweight workflow contract

“Lightweight” describes the shortest user journey, not a promise that Audio-V has a small capability set:

Choose one file.
Run the complete Oracle audit.
Read Quick Inspect on one screen.
Open a deep view only when the summary names something worth inspecting.

Quick Inspect and the engineering views consume the same Oracle record. There is no fast-but-weaker verdict, hidden playback dependency, or alternate quality score. The larger workstation remains available for people who choose folders, comparison, reports, or advanced inspection.

Product boundary

Audio-V is an audio integrity and fidelity workstation, not a general music player, tag editor, downloader, or mastering suite.

No playback transport is in scope. Audio comparison uses decoded measurements, waveform envelopes, spectra, and identity evidence without becoming a player. Lyrics, scrobbling, equalizers, spatial effects, and library playback are explicitly excluded. This keeps every primary action aimed at assessment, verification, comparison, reporting, or safe non-destructive remediation.

AudioAuditor parity contract

The official AudioAuditor website and source repository are the feature benchmark. Audio-V adopts assessor capabilities only when their evidence can be named and reproduced.

AudioAuditor benchmark capability	Audio-V decision	Current state
Automated batch analysis	Adopt	Implemented with recursive discovery, adaptively bounded 1–8 worker decoding, pause/cancellation, truthful finalization progress, checkpoints, and cache reuse
Fake-lossless / spectral-cutoff review	Adopt conservatively	Implemented as an explicitly heuristic origin assessment with inconclusive behavior
AI-generated audio detection	Research-gated	Generic “AI detected” verdict excluded until a validated corpus exists; offline C2PA validation, known generator metadata, and known identifier strings are implemented as separately labeled indicators rather than an AI verdict
MQA detection	Evidence-gated	Implemented only for editable metadata and decoder-profile declarations; Audio-V explicitly does not imply authenticated MQA, proprietary unfolding, or source provenance
Fake stereo detection	Adopt with neutral language	Implemented as mono, sample-identical dual-mono, near-mono, distinct-stereo, or inconclusive using correlation and side-to-mid energy
Clipping analysis	Adopt	Implemented total/per-channel percentages, contiguous events and timeline, near clipping, BS.1770 true peak, and conservative scaled-clipping review; destructive peak reconstruction is never promised
Spectrogram viewer	Adopt and deepen	Implemented 512/2,048/4,096/16,384-point STFT, combined/L/R/L−R views, zoom/pan/region measurement, scientific colormaps, and single/batch PNG export
Waveform and spectrogram compare	Adopt without playback	Implemented independent File A/B loading, bounded waveform envelopes, paired spectra, normalized spectral-difference heatmap, and measurement table
Audio player, equalizer, lyrics	Exclude	No player, EQ, lyrics, queue playback, or transport code
CSV / PDF / XLSX / DOCX export	Adopt and add JSON	Implemented from one evidence model; JSON remains the lossless machine-readable format
Free / actively developed / open source	Product decision	Active development is current; source is AGPL-3.0-only, contributor and trademark policies are published, and development releases disclose their unsigned status

AudioAuditor also advertises MQA and experimental AI checks. Audio-V now covers exact/internal silence, steep-transition, click/pop, and stuck-sample candidates; persistent Chromaprint identity relationships; offline C2PA inspection; known generator/signature inventory; integer lossless bit-utilization review; and explicitly unauthenticated MQA metadata/profile declarations. A statistically calibrated AI classifier remains future work and may not appear as authoritative until a labeled corpus and failure boundaries exist.

Capability matrix

Status meanings:

Implemented — real data flows end to end and is covered by verification.
Foundation — schema or UI exists, but the promised analysis does not.
Planned — specified and prioritized, but not implemented.
Excluded — intentionally outside the core product.

Capability	User value	v0.4.46 source status	Disposition
File and folder ingest	Analyze one track or a full library	Implemented	Current
Recursive bounded discovery	Avoid freezing on large trees	Implemented	Current
Technical metadata	Codec, container, rate, depth, channels, duration	Implemented	Current
CBR/VBR/ABR mode	Verify encoding details requested by Jota	Implemented from packet evidence where the demuxer exposes it	Current
Frame-level bitrate statistics	Show average, percentile range, deviation, and coverage	Implemented for demuxers that expose packet size and duration	Current
Full decode integrity	Detect truncation and malformed frames	Implemented	Current
FLAC STREAMINFO MD5	Deterministic lossless integrity evidence	Implemented	Current
Sample peak and RMS	Requested objective level measurements	Implemented	Current
Integrated LUFS and LRA	Standards-based loudness and variation	Implemented	Current
True peak	Detect inter-sample peaks	Implemented	Current
Delivery profiles	Evaluate a file against an explicitly chosen destination	Implemented for EBU R 128 programme QC, ATSC A/85, and AES internet-music track references with exact limits preserved in evidence	Current
Clipping and near-clipping	Find damaged or overly limited masters	Implemented with per-channel events and timeline	Current
DC offset, silence, dropout	Diagnose signal defects	Implemented	Current
Quick Inspect	Read one file's verdict and essential evidence without entering the engineering workflow	Implemented as the default selected-file summary with direct deep-view actions	Current
Channel layout and codec mode	Stereo, mono, 5.1, joint stereo	Implemented from probe/decoded evidence, with explicit MP3 MPEG-frame mode parsing	Current
Stereo correlation / duplicate mono	Detect phase and fake stereo issues	Implemented	Current
Full-track spectrogram	Direct visual inspection	Implemented with multichannel inspection and export	Current
Adjustable analysis settings	Reproducible visual evidence	Implemented	Current
Effective-bandwidth estimate	Support upsample/transcode review	Implemented as conservative heuristic evidence	Current
Evidence-backed verdict	Explain facts versus heuristics	Implemented with scoped Clear, Review, Failed, and Analysis error outcomes	Current
Human review disposition	Record how a person handled a result without rewriting machine evidence	Implemented with seven explicit outcomes, optional notes, timestamps, and report export	Current
Batch queue, pause, resume, cancel	Library-scale workflow	Implemented with crash-safe checkpoints, immediate cancel acknowledgement, hard worker recycling, and conservative recovery limits	Current
Saved cache and sessions	Resume audits and avoid repeated work	Implemented with SQLite persistence, startup interruption recovery, and suspect-file isolation	Current
Storage reclamation	Keep large cleared histories from consuming disk indefinitely	Implemented with incremental reclamation for new databases, measured legacy free space, and isolated one-time optimization	Current
PDF/XLSX/DOCX/CSV/JSON reports	Share and automate results	Implemented from a shared evidence model	Current
File/edition comparison	Distinguish masters and encodes	Implemented with independent loading and full-overlap multichannel null testing	Current
Duplicate fingerprints	Find identical audio across containers and sessions	Implemented with persistent SQLite index	Current
External MD5/SHA manifests	Verify file identity against supplied sidecars	Implemented	Current
AccurateRip/CTDB verification	Verify CD extraction provenance when context exists	Eligibility contract implemented for complete CD-frame-aligned single-image cue context; checksum/database match remains planned	Post-1.0 lookup
Acceptance/soak evidence	Document real package and library behavior	Implemented privacy-safe workload, timing, sampled memory, storage growth, cancellation, source, and recovery evidence export	Current
AI-generated audio detection	Experimental and difficult to validate responsibly	Deterministic provenance indicators implemented; statistical classifier deferred	Research only
Full music-player features	Does not improve the central verdict	Excluded	Excluded
Metadata editing	Risks turning inspection into mutation	Excluded	Excluded from 1.x

Competitive findings

No single competitor invalidates Audio-V, but the bar is substantially higher than a spectrogram plus metadata:

Spek is fast, cross-platform, supports FFmpeg formats, exports images, and makes spectrogram inspection immediate. Its published purpose is focused and approachable: show the acoustic spectrum of a file. Audio-V should preserve that clarity while serving the substantially larger batch-verdict and evidence workflow.
MediaInfo has deep, dependable container and stream metadata, multiple exports, and a library interface. It does not answer signal-authenticity questions.
Sonic Visualiser is the benchmark for inspectable, adjustable visualization, layered annotations, plugin-derived features, and responsive work on substantial audio. It is an expert exploration environment rather than a guided library verdict system.
Fakin’ The Funk is centered on batch quality estimation and frame scanning. Its value validates the use case, while its classifications also demonstrate why Audio-V needs inspectable evidence and explicit uncertainty.
AudioAuditor currently has the broadest directly comparable feature list: spectral estimation, clipping, MQA, fake stereo, loudness, true peak, comparison, exports, batch tools, and a Windows GUI alongside a player, EQ, lyrics, metadata tools, and experimental AI detection. Audio-V cannot compete by copying breadth; it must stand out through cross-platform parity, scientific transparency, reproducibility, workflow focus, and validation quality.
Lossless Audio Checker research demonstrates that upscaling, upsampling, and some transcoding detection can be validated against ground truth. That is a better model than inventing an opaque “quality percentage.”

Market position

Audio-V’s category is evidence-backed audio library assessment. It is not positioned as a replacement for every specialist:

Spek remains the quick single-file spectrum check.
Sonic Visualiser remains the deeper annotation and plugin laboratory.
CUETools and AccurateRip remain the authority for database-backed CD-rip verification when complete disc context is available.
AudioAuditor remains a broader Windows media-and-analysis suite.

Audio-V’s attention-worthy combination is full-stream integrity, conservative origin assessment, signal and delivery diagnostics, scientific file comparison, identity/provenance context, resumable large-library operation, human disposition, and durable reports under one versioned Oracle evidence model. The absence of playback, tag editing, lyrics, and generic AI badges is a product-focus decision rather than an unfinished feature list.

Scientific and standards contract

Audio-V measurements must name their governing method:

Programme loudness and true peak: ITU-R BS.1770-5.
EBU mode, momentary/short-term/integrated loudness, and loudness range: EBU R 128 and Tech 3341/3342.
FLAC integrity: decode and compare the unencoded PCM MD5 stored in STREAMINFO as described by Xiph.
CD-rip provenance: only call a rip verified when it matches a suitable external database such as AccurateRip or CTDB with the required whole-disc context. Audio-V currently reports eligibility—not a match. A matching FLAC MD5 proves internal stream integrity, not provenance.
Fingerprints: use local Chromaprint for similarity and duplicates. Any AcoustID lookup must be opt-in because it is a network call with service terms and rate limits.

Every heuristic classifier must ship with:

A versioned algorithm and settings.
A controlled corpus with ground truth.
Per-class false-positive and false-negative results.
Evidence for and against the verdict.
Explicit inconclusive behavior.
Regression protection for historical recordings, vinyl transfers, analog tape, low-pass mastering, silence, speech, and unusual genres.

Architecture decision

The Oracle Engine must be an independent, versioned worker rather than renderer code.

Gate 1 implementation status

Oracle file analysis now executes in a bounded, crash-replacing worker-thread pool. The Electron main process coordinates authorized jobs but does not perform PCM, FFT, or Oracle classification loops.
Audit sources, incremental results, warnings, and terminal states are persisted to a WAL-mode SQLite session database. Canceled and failed sessions retain completed records.
Reports are generated from authoritative main-process session/file records rather than renderer-supplied measurement objects.
True-peak remediation derives its gain from the current authoritative Oracle record rather than accepting a renderer-supplied measurement.
Renderer navigation, permission requests, and content loading use default-deny policies.
Spectrogram power and waveform envelopes are channel-safe and no longer erase opposite-polarity stereo material.
Audit history can reopen persisted evidence or clear every terminal session after an explicit confirmation. Clearing is blocked during an active audit and does not delete source audio, reusable Oracle cache, or the independent Identity index. A small foreground transaction hides terminal sessions immediately; physical saved-file records are then reclaimed in bounded 50-record background transactions with renderer progress, foreground storage interleaving, and launch-time continuation after interruption. At startup, orphaned running sessions become explicit interrupted checkpoints; the last source label is restored, adaptive resume avoids preloading the large partial result set, completed cache records are reused, and files active during interruption are validated serially before the saved worker profile is restored. A candidate that repeats the interruption is quarantined alone on the next resume. Requested and effective limits remain separately visible, and an explicit prior-limits path discloses its repeat-crash risk. Active queues can pause after current jobs finish. Cancellation rejects active work immediately, interrupts discovery/checksum/finalization stages, gives native decoders a cooperative shutdown window, then recycles any analysis worker blocked in synchronous computation. Completed records remain checkpointed, and failed records can be retried.
New SQLite databases enable incremental page reclamation before schema creation. Settings reports total, live, and reclaimable database bytes. A legacy database with material free space can run one isolated VACUUM while no audit or history cleanup is active; this preserves logical cache, fingerprint, and session records, returns free pages to the operating system, and enables bounded future reclamation.
The legacy 300-entry JSON cache migrates once into the unbounded SQLite Oracle cache, keyed by file size, modification time, and Oracle engine version.
Compare now runs a separate decoded-signal worker that measures offset, gain, polarity, correlation, residual energy, and relationship. Stacked, overlay, wipe, and blink modes share zoom and horizontal navigation; aligned residual waveform/spectrogram panels and a selected-region measurement expose local differences without replacing full-track multichannel null evidence.
Spectrum analysis now stores a selectable 512-point overview and 2,048-point detail tier.

Gate 1 is complete at the repository level. Compare uses bounded alignment estimation and then null-tests the complete overlapping decoded tracks across every matching channel; channel-count mismatches are disclosed instead of silently coerced.

Gate 2 implementation status

All SQLite session and cache access now runs in a dedicated storage worker. Renderer and Electron main-process work no longer performs synchronous database operations.
Analysis uses an adaptive, bounded worker schedule derived from available CPU parallelism instead of fixed two-file waves. Result order remains deterministic even when files finish out of order.
Session writes are buffered into bounded batches and committed transactionally, reducing per-file worker messages and SQLite transaction overhead.
Renderer progress events are coalesced to animation frames, and the audit result table is virtualized with overscan. A 10,000-record audit retains authoritative application state while mounting only the visible rows.
Rich waveform and spectrogram arrays no longer accumulate in the Electron main process or renderer. Live queues carry measurement summaries without visual matrices; durable history retains a bounded 32-region spectrogram overview, 160-point waveform, the complete measurement/verdict summary, and the independent 4,096-point classifier summary. Higher inspection tiers regenerate from the selected source on demand, and only 16 recently used detailed records remain resident.
Resource controls are per-file ceilings rather than independent speed settings. Before workers start, Audio-V resolves one to eight requested file workers against a system-wide budget capped at 20% of physical memory or 8 GB and 75% of logical CPUs; unsafe combinations are visibly reduced.
Current-audit fingerprint matching uses exact-hash and ±3-second duration indexes instead of an all-pairs scan. Exact candidate expansion is capped and near-match expansion is deferred above 2,000 files. Historical enrichment uses no more than four outstanding storage requests, indexed exact-hash lookups above 100 files, and exact plus near-duration matching for smaller selections; it remains deferred above 1,000 files. A supplemental history timeout produces one audit notice and cannot discard completed Oracle verdicts or prevent session completion.
Checksum manifests are indexed once per source, and the Oracle reuses an already-computed SHA-256 result instead of rereading the file.
Mounted and removable libraries use eight-way bounded directory discovery. Full audits stage each active macOS /Volumes, Windows UNC, mapped-drive, or non-system-drive file through one 4 MB sequential transfer, run repeated forensic passes locally, and delete the copy afterward. Staging follows the resolved 1–8 file concurrency and an aggregate temporary-space reservation, rather than adding an independent unbounded “network sessions” control.
Authoritative JSON and CSV session reports stream from paged SQLite records without a fixed payload limit. PDF, DOCX, and XLSX generation also runs outside the Electron main process and consumes compact session evidence instead of hydrating every full spectrogram.
Compiled-worker verification now exercises Oracle analysis, comparison, durable storage, and streamed report export.
Rotating JSONL application logs record lifecycle, scan/session identity, resource limits, per-file starts and outcomes, Oracle attempt/retry/timeout events, storage-worker recovery, checkpoint state, and renderer/child-process termination. Logs remain local, are size bounded, and are separate from the privacy-safe diagnostics export.
Successful scan completion logs now include application version, elapsed wall time, file and error counts, point-in-time main-process RSS, database size, reclaimable bytes, source kind, and run mode. The RSS field is explicitly not represented as whole-application peak memory.
Exhausted Oracle worker timeouts persist as per-file Analysis error records with the exact filename, oracle-engine failure stage, worker code, attempt count, and evidence. The failed worker is replaced and the remaining library continues; infrastructure-error cache entries are retried on later audits.
Per-file Reports use independent bounded list and evidence panes. Selecting a row anywhere in a long report keeps the evidence visible beside it, resets the new evidence report to its header, and preserves independent scrolling in both wide and stacked window layouts.
npm run benchmark:scale enforces repeatable persistence and restoration budgets for dense-evidence 100-, 1,000-, and 10,000-record sessions. The v0.4.34 benchmark persisted 10,000 records in 6.95 seconds, restored compact history in 485 ms, and occupied 458.4 MB (437.1 MiB) across the cumulative 11,100 stored records, below the 600 MiB database budget.

These dense-evidence figures validate storage and compact state retrieval on the development Mac; they do not represent full audio decode throughput.

Gate 2 is complete for the repository-level scale foundation. The maintainer has additionally completed a native-Windows real-library audit of at least 6,000 files with the portable distribution. Exact package identity, timing, memory, database growth, and recovery measurements were not captured, so instrumented acceptance and clean-environment profiling remain open for supported Windows/macOS versions, 1/10 GbE shares, removable/network volumes, long-duration files, constrained memory, cancellation under load, and thermal throttling. Discovery still retains a sorted path list, renderer filters remain linear in session size, and rich document exports remain whole-session operations inside the storage worker. Those are measured optimization candidates rather than hidden claims of unlimited scale.

Gate 3 implementation status

Oracle Engine v5 separates heuristic rule strength from evidence coverage. Fixed rule values are no longer presented as provenance probabilities.
Every successfully decoded Origin Assessment reports 0–100% evidence coverage plus a machine-readable reason code. Inconclusive files explain whether duration, measurable bandwidth, or cutoff stability prevented classification.
Origin coverage, reason, and score type flow through the details UI and every report format.
A versioned golden-corpus manifest covers native wideband controls, known MP3-to-FLAC and upsample transformations, intentional low-pass material, digital silence, insufficient duration, and narrow-band tonal content.
npm run validate:fidelity executes the compiled Oracle against the corpus, blocks regressions, and records a machine-readable scorecard in build/fidelity-validation-latest.json.
Changing the failure semantics advanced the engine/cache identity to 0.4.0-oracle-v6, preventing older catch-all damage interpretations from appearing under the new model.
Deep spectrogram inspection and clipping-event evidence advance the current identity to 0.5.0-oracle-v7, invalidating cached v6 measurements that do not contain the expanded evidence model.

Gate 3 now has both a small deterministic regression corpus and a separately governed real-world validation matrix. The current matrix contains 2,295 controlled cases derived from 85 independent references across five registered sources, with group-separated splits and published per-class measurements. It reaches the project's numeric public-reference target, but its present zero-positive-detection result for controlled lossy and upsample transformations means the Origin Assessment is not probability-calibrated and cannot claim universal accuracy. Audio-V therefore continues to use possible, bandwidth limited, or inconclusive language, and promotion remains blocked until the declared recall, false-positive, and confidence-bound requirements are met on eligible held-out evidence.

Gate 4 implementation status

Binary identity is explicit rather than dependent on whichever credentials happen to exist on the build machine. The default macOS build disables certificate discovery, notarization, and Hardened Runtime but applies a complete ad-hoc bundle signature so Gatekeeper sees a valid integrity seal; the default Windows build disables executable signing.
Separate future signed build commands retain macOS Hardened Runtime/notarization and Windows Authenticode integration without requiring credentials for open-source development releases.
Package inspection verifies the two required DMGs, Windows installer, Windows portable executable, embedded application archive, FFmpeg/ffprobe engines, notices, engine manifest, PE signatures, and both Intel/Apple Silicon architectures in the Universal application and engines. Packaged-runtime smoke tests then launch the native application, fully analyze a real FLAC fixture, verify the current Oracle engine identity, and exit through a guarded QA channel.
Every release receives SHA-256 checksums and UNSIGNED_RELEASE_MANIFEST.json, which declares the distribution and signing policy for each exact artifact.
Maintainers run the full suite and package inspection locally before publication. An optional manual-only workflow can produce macOS and Windows Oracle snapshots over identical fixtures and compare deterministic values exactly and measured values within documented tolerances.
Release publication is a local-build, direct-upload process modeled on Deemix Remastered. Tag/package version mismatch, missing project licensing, failed artifact inspection, or a failed release-candidate check blocks maintainer publication.
User documentation explains checksum verification, platform warnings, safe per-application approval, portable-state behavior, and why managed Windows policy may prevent unsigned execution.

Gate 4 is complete at the repository and local-package level. Native Windows portable operation and clean Apple Silicon behavior are maintainer-witnessed. Actual macOS-versus-Windows Oracle parity remains unproven until maintainers collect and compare native snapshots; Windows installer upgrade/uninstall, detailed clean-macOS journeys, and Intel launch remain release-candidate acceptance. The former public-release license blocker is resolved by the Gate 5 AGPL-3.0-only decision.

Gate 5 implementation status

Audio-V is licensed AGPL-3.0-only. Every application package includes the canonical license, copyright notice, corresponding-source notice, third-party notices, engine manifest, and trademark policy.
The Contributor License Agreement preserves contributor ownership while granting DRAZY the explicit sublicensing and relicensing authority needed for continued AGPL distribution, commercial editions, and dual licensing. Pull requests must record affirmative agreement.
The Audio-V name, logo/icon, Oracle Engine identity, and official-release claims are reserved separately from the code license; compliant forks must use distinct branding.
Primary navigation, verdict filters, analysis tabs, virtual result rows, canvases, and live scan state expose keyboard and assistive-technology semantics. A skip link, visible focus, reduced-motion behavior, and Command/Ctrl keyboard workflow are implemented.
An automated accessibility gate checks both the production DOM and Chromium accessibility tree. The current audit passes 41 axe-core rules, exposes 154 semantic nodes, requires a named interactive-control tree, and requires workspace landmark and heading structure.
Settings exports a privacy-safe diagnostic record containing application/runtime versions, aggregate session status, and engine capabilities without accepting filenames, paths, hashes, tags, or audio evidence. A unit test enforces the aggregate-only boundary.
The beginner user journey, plain-language Oracle Engine and glossary, scientific methodology, privacy/security, unsigned-installation, release, contribution, and release-candidate documents cover the complete product workflow and limitations at different levels of depth.
npm run verify:release-candidate produces a machine-readable automated readiness result from licensing, documentation, accessibility, fidelity, scale, and artifact evidence.

Gate 5 is complete for automated repository readiness and local-package validation. The application is release-candidate quality but not yet fully market-validated: the remaining matrix in docs/RELEASE_CANDIDATE_CHECKLIST.md includes the clean Windows installer lifecycle, Windows/macOS Oracle parity, Intel hardware launch, VoiceOver/Narrator workflow review, display scaling, and an exported instrumented real-library record. Signing/notarization is an explicit policy waiver rather than a failure.

Real-world validation milestone status

The versioned five-source registry covers Slakh2100, MUSAN, MAESTRO, EBU SQAM, and MUSDB18-HQ with official URLs, access rules, license boundaries, expected audio properties, and intended validation roles.
Dataset-specific adapters discover eligible recordings, prevent duplicate source identities, record SHA-256 and decoded technical properties, and keep restricted references challenge-only.
The external-reference registry enforces rights records, SHA-256 identity, public-versus-challenge disposition, and source-group partitions.
Deterministic recipes generate at least ten native, lossy-transcode, upsample, and intentional-filter cases per imported reference.
Evaluation records accepted classifications, detector confusion counts, inconclusive behavior, and exact 95% Clopper-Pearson binomial intervals grouped by dataset, source category/subcollection, split, recipe, and truth class.
Corpus evaluation uses true worker threads with a conservative four-worker default, resumable 25-case checkpoints, hash-verified result reuse, and an explicit validation-only override of up to 16 workers on capable machines.
Settings reads the exact status shipped with the package and exposes public-reference, source-group, case, corpus-version, and claim-level information.
Reports state that heuristic rule strength is not a probability-calibrated provenance claim.

Infrastructure is complete and all five registered sources are measured: 85 independent references produce 2,295 cases across MUSAN, Slakh2100 Redux, EBU SQAM, MAESTRO, and MUSDB18-HQ. The public population reaches the numeric 50-reference target, and the restricted sources remain local challenge evidence. The scorecard demonstrates zero false origin advisories on its eligible negative populations but zero positive detections on controlled lossy and upsample transformations; that published recall gap prevents probability-calibrated or authoritative provenance claims.

Provenance, automation, and identity milestone status

Oracle Engine v10 validates embedded and locally resolvable C2PA Content Credentials with C2PA Tool 0.27.3. Remote-manifest and OCSP fetching are disabled so ordinary audits remain deterministic and private.
Valid, untrusted-signer, invalid, absent, unsupported, and tool-error states remain distinct. A credential can validate a signed provenance statement; it does not establish truthfulness, human authorship, ownership, or audio quality.
Known generator names in editable metadata and known watermark/signature identifier strings in raw file bytes are inventoried with their exact source and limitation. Audio-V does not claim to decode proprietary watermarks and does not produce a generic AI Yes/No badge.
Local Chromaprint fingerprints identify exact-fingerprint and high-similarity relationships within the current audit and across a dedicated, browsable Identity workspace. The historical index can be rebuilt from sessions, pruned for missing sources, or cleared without deleting audit evidence. Identify current audit recognizes and enriches already measured fingerprints in place with visible progress and per-file persistence. Optional AcoustID recognition sends only fingerprint plus rounded duration and returns linked MusicBrainz recording leads; it is off by default, serialized below the service ceiling, cached by fingerprint, and retried only for bounded transient failures. A release-injected client identity or OS-protected per-user key is never persisted in audit evidence. Direct MusicBrainz enrichment is a separate no-key opt-in that prefers an embedded MBID, otherwise uses the strongest AcoustID candidate, caches unique IDs, obeys the public one-request-per-second ceiling, and preserves bounded artist, ISRC, date, and release-group context as neutral identity evidence.
A parallel identity result classifies usable external evidence as metadata-corroborated, identity-matched, metadata-conflict, or inconclusive. The saved comparison names every declared and identified recording ID, title, or artist value considered. The result is intentionally outside the five Oracle verdict lanes and cannot change Clear, Review, Failed, Analysis error, or Not analyzed.
Metadata Inventory includes bounded native tags, BPM, ISRC, MusicBrainz IDs, declared ReplayGain values, embedded cue sheets, and adjacent cue-sheet references without invoking the decoder. Full Audit adds calculated RG2 track gain, album eligibility/reason evidence, grouped album gain, independent cue INDEX 01 programme analysis, and separate INDEX 00 pregap analysis.
Full audits add DR meter values and integer-lossless bit-utilization/truncation assessment. Possible zero-padded depth is an origin advisory, not deterministic file damage.
The accepted format list now follows a broad FFmpeg-backed set of common audio and audio-container extensions. Actual support remains decoder-runtime evidence rather than a promise based only on a filename suffix.
The packaged headless CLI accepts files and recursive folders, emits the same compact JSON evidence model, supports CI evidence-policy exit codes, and exposes worker, JavaScript heap, FFmpeg-thread, and native-process RSS limits. The desktop exposes the same resource policy for the next audit.
EBU R128 and DR measurements share one FFmpeg filter-graph pass. Full PCM analysis remains streaming and bounded; local fingerprinting performs one additional decode capped at 120 seconds. Opting into AcoustID performs a second capped fpcalc pass to create the service’s encoded fingerprint.

This milestone deliberately stops before a statistical AI classifier or automatic network service. Those claims remain corpus-gated.

Oracle v10 refinement status

A three-case native decoded defect corpus gates the clean, single-impulse click/pop, and non-zero stuck-sample behaviors. It is labeled synthetic regression evidence and does not claim real-world calibration.
Compare accepts explicit one-to-one channel maps for differing decoded layouts. Without a map, mismatched layouts remain a bounded preview rather than receiving a misleading null.
ReplayGain reports why album gain is unavailable, including missing album identity, insufficient matching tracks, incompatible channel counts, measurement failure, and large-library deferral. Audits above 1,000 files retain calculated track gain but skip the otherwise unbounded second full-library decode; selecting an album or a smaller source calculates exact grouped album gain.
Cue programme segments stop at the next INDEX 00/01 boundary; declared INDEX 00 pregaps are decoded and reported separately.

Oracle v12 evidence-model status

Results now expose separate integrity, signal-defect, spectral-origin, provenance, and delivery lanes. Advisory provenance strings, stereo relationships, bit padding, positive true peak without a selected delivery profile, and steep transitions no longer force the overall verdict to Review.
Optional delivery conformance now evaluates exact, persisted EBU R 128 programme-QC, ATSC A/85, or AES internet-music track limits. A miss produces Review and never changes deterministic integrity to Failed.
Complete CD-frame-aligned single-image cue sources receive explicit CTDB/AccurateRip eligibility evidence without being mislabeled as database-verified. MQA text/profile declarations remain neutral format markers.
Every audit checkpoints a privacy-safe acceptance record with package, platform, workload, throughput, sampled process working set, storage growth, cancellation response, source class, and recovery behavior. The packaged reference-audit gate validates this evidence contract.
Strict full-decode failures receive a complete tolerant confirmation analysis. Only failure of both paths can establish deterministic stream damage; tolerant success becomes recoverable structural nonconformance for Review.
The origin classifier uses a dedicated 4,096-point, 48-region tier, retains bounded summary evidence rather than another persisted matrix, and requires a stable band edge plus corroborating evidence before issuing a strong possible-upsample or possible-lossy-transcode pattern.
Top-level and origin probability-like percentages are retired. The engine publishes ordinal rule strength, evidence coverage, regional stability, and independent-indicator count as distinct quantities.
Click/pop detection uses a locally adaptive robust residual threshold. Finding promotion considers magnitude, repetition, or duration; steep transitions alone remain advisory.

Recommended layers:

Probe — container, codec, tags, frame structure, and declared properties.
Decode — bounded streaming conversion to normalized PCM without loading entire albums into memory.
Measure — deterministic integrity, levels, loudness, stereo, and defect metrics.
Spectrum — reusable STFT tiles, spectrogram pyramids, and spectral descriptors.
Classify — versioned evidence rules and calibrated probabilistic models.
Persist — SQLite session/cache keyed by file identity, engine version, and settings.
Present — renderer consumes typed results and never invents measurements.

FFmpeg is a viable decoder and measurement backend and exposes filters such as astats, ebur128, silencedetect, aphasemeter, and showspectrumpic in its official filter documentation. Packaging must not begin until a redistributable binary policy is selected: FFmpeg is normally LGPL, but optional components can make a build GPL, as described in its legal guidance. Audio-V must record the exact binary version/configuration in every report and ship the required notices and source offer.

Release gates

No build may be called a beta until:

The original four-feature request works on real files on both macOS and Windows.
The packaged application contains every required engine binary; it cannot depend on Homebrew, PATH, or a developer machine.
Decorative plots and demo verdicts are absent from production mode.
Results from macOS and Windows agree within documented tolerances.
DMG and EXE installation is tested on clean virtual machines.
The analysis corpus includes known-good and intentionally transformed fixtures.
Every release publishes checksums, third-party notices, and an engine manifest.
Signing/notarization status is explicit.

Immediate build sequence

Remove fictional production data and repair metadata semantics.
Add single-file and multi-file ingest beside folder ingest.
Establish scanner and contract tests with generated fixtures.
Select and license the packaged decode backend.
Implement full decode, technical probe, peak/RMS, LUFS/LRA/true-peak, and channel mode.
Implement real spectrogram tiles and settings.
Add the golden corpus before adding transcode or upsample verdicts.
Add queueing, cache, reports, and comparison after the measurements are trusted.

Corrections completed during this audit

Production now starts empty; fictional tracks and verdicts were removed.
Decorative spectrogram and loudness charts no longer masquerade as measurements.
Individual and multiple file selection now sit beside recursive folder ingest.
Codec profile is no longer mislabeled as CBR/VBR mode.
Metadata-probe failure is no longer called proven stream damage.
Source warnings no longer abort all valid files in a mixed selection.
Packaged FFmpeg/ffprobe 8.1.2 LGPL engines perform complete multi-codec stream decoding and technical probing on macOS and Windows.
Real sample peak, RMS, DC offset, clipping, near-clipping, stereo correlation, duplicate-mono, EBU R128 integrated loudness/LRA, and BS.1770 true-peak metrics run automatically during batch ingest.
Measured 512- and 2,048-point Hann-window STFT tiers are created for successfully decoded formats during active inspection. Durable large-library history retains a bounded 32-region overview and the 4,096-point classifier summary; 4,096- and 16,384-point combined/L/R/L−R presentation views decode on demand and expose zoom, pan, region measurement, scientific colormaps, and batch PNG export.
Streaming packet analysis reports p05/p95 bitrate, deviation, duration coverage, and observed CBR/VBR for every demuxer that exposes packet size and duration.
Native FLAC files receive an independent canonical decoded-PCM comparison against the MD5 stored in STREAMINFO.
Exact digital-silence runs, internal dropout candidates, steep transition candidates, crest factor, and peak-to-loudness ratio are measured and disclosed.
Conservative spectral-origin review is regression-tested against native-wideband, MP3-to-FLAC, and 44.1-to-96 kHz controls; it never claims that bandwidth alone proves provenance.
The Oracle produces scoped Clear, Review, and Failed outcomes for complete-decode, signal, loudness, and spectral checks, with deterministic evidence and an explicit provenance disclaimer.
Compare performs real file-to-file property and measurement comparison; Repair provides non-destructive triage; Reports summarizes the session and exports complete JSON evidence.
Deterministic decoder corruption and decoded-audio checksum mismatches finish as Failed with a stage and exact evidence. Tool, timeout, resource, probe, measurement, and internal failures finish as Analysis error without making a damage claim.
Metadata Inventory is an explicit non-decoding workflow whose records remain Not analyzed until a Full Oracle Audit runs; it is no longer represented in verdict distribution as “Metadata only.”
Batch audits can terminate active decoders, preserve completed rows, and restore unchanged v3 results from a persistent size/mtime-invalidated cache.
Spectrogram inspection supports linear/log frequency display, −120/−100/−80 dBFS floors, cursor time/frequency/level readout, zoom/pan, exact dragged regions, combined/left/right/L−R views, scientific colormaps, and single/batch PNG export.
Clipping evidence includes total and per-channel percentages, contiguous event grouping, a location timeline, and a conservative repeated-plateau indicator for possible scaled clipping. These remain Review findings unless independent deterministic integrity evidence establishes failure.
Review now explains the exact trigger, distinguishes integrity failure from measured or heuristic warnings, and offers acknowledgement, evidence export, source reveal, re-analysis, and Repair routing.
Repair Lab provides a visible diagnose/choose/verify process. Positive true-peak findings can create a separate −1 dBTP FLAC working copy that is automatically decoded and audited; sources are never overwritten, and the UI explicitly states that gain reduction does not reconstruct clipped peaks.
Origin Assessment separates container-declared profile data from measured spectral evidence and exposes the classification, heuristic rule strength, evidence coverage, measured bandwidth, cutoff data, supporting basis, and provenance limitation.
Repair is modeled as attention and remediation rather than an assertion that every listed file needs repair. Each item declares whether mitigation is optional, automatic repair is unavailable, manual verification is required, source comparison is recommended, or replacement is recommended.
True-peak remediation preserves a 16-bit or 24-bit source word length by default and offers explicit 16-bit compatibility or 24-bit processing copies. Sixteen-bit output is triangular-dithered after gain processing; sample rate, channels, metadata, and embedded artwork are preserved and the verified output depth must match the selection.
Compare has independent File A and File B disk selectors in addition to current-audit choices; directly loaded comparison files are decoded without replacing the audit session.
Reports is explicitly the current audit report. Every file row opens identity, engine, decoded measurements, origin assessment, expandable evidence, and direct audit/compare/remediation/export actions.
Folder and adjacent-file audits discover GNU, BSD, and filename-specific MD5, SHA-1, SHA-256, and SHA-512 sidecar manifests. Matching entries become deterministic support; mismatches become Review because they disprove manifest identity without pretending to know whether the file or manifest is authoritative.
Stereo authenticity now distinguishes exact dual mono, near mono, distinct stereo content, mono, and inconclusive results using both Pearson correlation and side-to-mid energy. Dual/near mono is a review signal, not an accusation of deceptive intent.
Compare renders bounded full-track waveform envelopes, paired measured spectrograms, and a normalized B-minus-A spectral heatmap alongside identity, format, loudness, clipping, bandwidth, stereo, and checksum values; no playback surface was introduced.
Compare now hydrates full visual evidence on demand when the active audit uses compact in-memory records. File A and File B expose independent loading/error states, every canvas clears stale pixels when evidence changes, and the B-minus-A spectrum follows the shared zoom and horizontal position.
Full-track comparison no longer labels a short or uncorrelated overlap aligned-equivalent; equivalence requires at least 80% duration coverage plus corroborating preview correlation unless the user supplied an explicit channel map.
AcoustID enablement, the validated personal application key, and the independent MusicBrainz preference now survive restarts. The key is encrypted by macOS Keychain or Windows DPAPI, never persisted as plaintext or audit evidence, and can be explicitly cleared.
Audit and per-file evidence export to PDF, XLSX, DOCX, CSV, or JSON from a shared 40+ column evidence model. Production dependencies remain free of known audit advisories.
The desktop typography floor is 11px for utility labels and 12–14px for working text, with larger rows, controls, and analysis panels.
The 168-test regression suite covers discovery, mounted-source staging and identity preservation, malformed metadata, PCM math, internal WAVE decoding, spectral-bin detection, multi-codec full decoding, FLAC audio-MD5 mismatch, loudness/true peak, delivery profiles, continuity, packet-rate behavior, fidelity controls, persistent fingerprint management, bounded 640-file historical finalization, indexed exact-history search, nonfatal enrichment timeout, AcoustID application-key normalization, service preflight, transient retry, cached-record identity enforcement, direct MusicBrainz enrichment, transient MusicBrainz throttling recovery, unique-ID caching, parallel identity corroboration/conflict classification, CD-database eligibility, neutral MQA marker inventory, privacy-safe acceptance evidence, decoded residual and selected-region comparison, desktop IPC error normalization, cue programme/pregap segments, calculated and large-library ReplayGain eligibility, explicit comparison channel mapping, resource enforcement, adaptive crash recovery with cache bypass, structured log rotation, cache invalidation, prompt cancellation of blocked workers, virtualized large-session viewport recovery, process cancellation, source-preserving repair output, attached-artwork retention, source preservation, truncation, and verdict truthfulness.
Production dependencies pass npm audit --omit=dev; remaining advisories are confined to the upstream packaging toolchain.
Release packages contain third-party notices and a machine-readable engine capability manifest.

FilesExpand file tree

PRODUCT_AUDIT.md

Latest commit

History