Local-first audio variation and evidence tool for material you own or are authorized to modify. SunoJump applies a seedable transform pipeline, supports preview and comparison, and records local signal-change evidence so you can audit each render.
SunoJump applies an 11-pass processing pipeline with non-uniform segment-based transforms. Each segment can receive different processing parameters, producing controlled spectral, temporal, phase, stereo, noise, dynamics, and codec variation.
- Use SunoJump only with audio you own or are authorized to modify.
- Built-in metrics are experimental, local, adapter/version-scoped measurements of the current input and output.
- A signal-change or landmark-overlap value does not predict or guarantee any platform, recognition, moderation, or detector outcome.
- SunoJump does not upload or resubmit audio, call platform APIs, or optimize against a platform response.
| # | Pass | What It Does |
|---|---|---|
| 1 | Ordinary Metadata Strip | Removes tags recognized by Mutagen after the separate C2PA provenance preflight |
| 2 | Spectral Perturbation | Scans stable narrowband candidates, then perturbs frequency magnitudes with per-band controls and randomized STFT window sweeps |
| 3 | Dynamic EQ | Time-varying multiband EQ with LUFS-like gain matching |
| 4 | Pitch Micro-Shift | Non-uniform pitch warping across random segments |
| 5 | Tempo Micro-Variation | Coupled in-segment timing drift that keeps segment boundaries beat-aligned |
| 6 | Phase Scrambling | Randomizes phase relationships in STFT domain |
| 7 | Stereo Manipulation | Mid-side processing to alter stereo field |
| 8 | Noise Injection | Adds masking-aware shaped pink noise below local spectral thresholds |
| 9 | Dynamics Modification | Per-frame random gain variation to break statistical patterns |
| 10 | Humanization | Wow/flutter, dynamic breathing, micro noise floor |
| 11 | Lossy Re-encode | Controlled MP3 encode/decode round trip (requires ffmpeg) |
Unlike a flat whole-track transform, SunoJump splits audio into variable-length segments and applies different transform parameters to each segment. Preview, comparison, run logs, and sidecars make those changes inspectable.
Download SunoJump.exe from the latest release. No Python install required. Double-click to launch.
git clone https://github.com/SysAdminDoc/SunoJump.git
cd SunoJump
# Install dependencies
python -m pip install -r requirements.txt
# Run
python sunojump.py- Python 3.11+
- ffmpeg (optional, for Lossy Re-encode pass)
- PyQt6 Multimedia (optional, for in-app preview playback; usually bundled)
Install Python dependencies before running from source; the packaged executable already includes them.
# Install local audit tooling, then report lock/native/DSP/security status
python -m pip install -r requirements-dev.txt
python tools/audit_dependencies.py
python tools/audit_licenses.pyThe compatibility report separates direct, transitive, build-tool, and native-runtime drift; runs the fixed-seed generated DSP golden; audits the release lock for known vulnerabilities; and prints the recorded rollback commit. Source environments may report drift as a warning because requirements.txt intentionally uses ranges. Release environments use --require-lock-match as a hard gate.
tools/compatibility_baseline.json binds both lock hashes to the tested Windows x64/CPython 3.12 native versions, DSP signature, supported Python 3.11/3.12 source lanes, and rollback point. Any lock update must refresh that record after the compatibility workflow passes. The workflow runs the DSP golden, offscreen GUI tests, packaging smoke tests, and full suite on both supported Python lanes; the CPython 3.12 release-contract job additionally installs both hashed locks and uploads the machine-readable compatibility report.
On Windows x64 with CPython 3.12, the release command creates a temporary virtual environment, installs only the exact wheels in the two hashed locks, builds an unsigned one-file executable, and proves its CLI, fixture-render, and foreground-GUI paths:
python tools/build_release.pyThe gate fails on a missing, stale, or wrong-version executable; a stale compatibility baseline; native-version drift; undeclared bundled packages; incomplete license/source records; or failed smoke tests. It does not perform code signing. dist/ receives the executable, SHA256SUMS, compatibility baseline, CycloneDX 1.7 SBOM, actual PyInstaller inventory, native-version report, license notices/source routing, build provenance, both locks, and the corresponding source archive. Use sha256sum -c SHA256SUMS from inside the downloaded artifact directory to verify the set.
On Windows, run pwsh -NoProfile -File tools/smoke_accessibility.ps1 to launch an isolated-settings GUI and verify its UI Automation tree, keyboard focus, primary control names, and unit-bearing slider announcements—the same accessibility surface consumed by Narrator.
- 11-pass audio processing pipeline — metadata strip, spectral perturbation, dynamic EQ, pitch/tempo micro-shift, phase scrambling, stereo manipulation, noise injection, dynamics, humanization, lossy re-encode
- LUFS-preserving Dynamic EQ — reshapes band energy without silently changing perceived loudness
- Psychoacoustic noise shaping — keeps injected pink noise lower in quiet regions and under louder spectral masks
- Narrowband candidate scan — locates stable spectral candidates per file and targets them during spectral perturbation
- STFT window sweeps — varies 1024/2048/4096-point windows across randomized spectral segments
- Per-band spectral controls — tune sub-bass, low-mids, presence, and air perturbation independently
- Coupled non-uniform pitch/tempo processing — varies pitch and timing together while keeping segment boundaries aligned
- 4 built-in presets — Gentle, Moderate, Aggressive, Extreme + Custom
- Per-pass toggles and strength sliders — fine-grained control
- Render Preview — choose any 0.1-second start offset and hear up to 30 seconds with your current settings before committing to full-file processing
- Compare Presets — one click renders a 20-second sample per preset so you can A/B/C/D audition all four, save a per-file winner, and restore that choice when the source is re-added
- In-app A/B playback — play original and processed side-by-side without leaving the app
- Responsive, keyboard-complete GUI — panels reflow down to 560×360, long labels wrap at enlarged font sizes and in RTL layouts, focus is visible, queue actions have keyboard equivalents, and sliders expose displayed units to screen readers
- Explicit codec export — WAV/FLAC/OGG export directly, with MP3/M4A export enabled when ffmpeg is available
- Contained input decoding — bounded pure-Python container inspection runs before a time/memory/output-capped decoder process; mismatched containers, IRCAM, and WAV IMA ADPCM are rejected before libsndfile
- C2PA provenance guardrail — locates embedded or adjacent Content Credentials before decode, blocks by default, and records manifest hashes plus the explicit output-without-source-credentials decision without claiming cryptographic validation
- Transactional output saves — reserves collision-free names across processes, validates same-folder temporary renders, and atomically publishes without replacing an existing audio or sidecar destination
- Privacy-controlled diagnostics — every GUI and CLI run writes a bounded local log with path redaction on by default; users control retention, preview redaction, confirm deletion, and export atomic support ZIPs
- Scoped local evidence — reports
sunojump.signal_change v1plus typedmeasured,unavailable, orerrorresults from experimentalsunojump.constellation v1 - Truthful terminal states — jobs and batches report
succeeded,partial,failed, orcancelledwith stable error codes; only an all-success batch reaches 100% - Lifecycle-safe cancellation — one Cancel control covers batches, previews, and preset comparisons; closing waits for confirmed worker exit and never removes active temp storage
- Responsive queue discovery — recursive folder scans run off the UI thread with counted progress and cancellation; empty and partial states explain the next action
- Actionable failures — queue rows show typed failure reasons, and Retry Failed immediately reuses the latest saved batch manifest
- Replayable output evidence — every render displays its effective seed; WAV/FLAC same-environment replays are byte-tested, while sidecars name exact muxer/codec dependencies when bytes are build-sensitive
- Parallel identity-safe batch queue — drag/drop multiple files, reorder them, and render 1–8 concurrently; stable job IDs keep independent progress and late results attached to the correct rows
- Per-file queue presets — assign any built-in preset or a snapshot of the current controls to selected files while the rest continue to follow the batch settings
- Resumable batch manifests — every batch atomically persists inputs, per-file preset snapshots, seeds, attempts, terminal states, and artifact hashes; interrupted and failed-only work can be resumed without replacing existing files
- Custom preset save/load — export your tuned settings to JSON, share, or reuse
- Bounded long-file processing — isolated decoding writes fixed-size blocks to a disk-backed memory map; payloads above 256 MiB run every enabled transform in overlap-blended chunks with a bounded working set and atomic cleanup
- Open Output — one-click to output folder in your file manager
- Versioned signal-change metric — inspect a sample-domain difference value with explicit scope
- Before/after spectrogram export — opt in to a bounded, full-file, side-by-side PNG with a shared dBFS color scale and hash-linked result/manifest evidence
- BS.1770-5 loudness audit — opt in to a bounded before/after integrated-loudness and oversampled true-peak JSON report with explicit deltas and hash-linked evidence
- Crest and stereo-width audit — opt in to bounded before/after crest factor, mid-side energy ratio, and interchannel-correlation JSON evidence with explicit deltas
python sunojump.py- Drop audio files into the file list (or click Browse)
- Select a preset or customize individual parameters
- (Optional) Select one or more queue rows and use Selected preset to assign a built-in preset or a snapshot of the current controls only to those files; Batch settings keeps them linked to the main controls
- (Optional) Choose a Start offset and click Render Preview to process up to 30 seconds from that point so you can hear the result before committing; Compare uses the same offset
- Optionally enable the spectrogram and/or BS.1770-5 report exports, choose Parallel files (2 by default), then click Process All to render every file in the list to the output directory with
_sjsuffix. If a source contains C2PA Content Credentials, review the warning and explicitly choose whether to continue; cancelling preserves the source and creates no output. - To recover a prior batch, click Resume Batch for pending/interrupted jobs or Retry Failed for failed/partial jobs, then select its
.sunojump-batch.jsonfile
Keyboard shortcuts cover the primary workflow: Ctrl+O browse, Ctrl+P preview, Ctrl+Shift+P compare, Ctrl+R resume, Ctrl+Shift+R retry failed, Ctrl+S save a preset, Ctrl+Enter process, and Esc cancel. In the queue, Delete/Backspace removes the selection and Alt+Up/Alt+Down reorders it.
# Basic usage with preset
python sunojump.py -i song.wav -p aggressive
# Custom parameters
python sunojump.py -i song.wav --pitch 1.5 --phase 0.5 --spectral 0.4
# Compose a built-in preset with reusable sparse JSON overrides
python sunojump.py -i song.wav --profile ./gentle-tuned.json --pitch 0.75
# Keep diagnostics on stderr and write one machine-readable result document
python sunojump.py -i ./my_songs --result-format json > results.json
# Emit one record per file plus a final batch record
python sunojump.py -i ./my_songs --result-format jsonl > results.jsonl
# Export a bounded side-by-side audit image beside every usable output
python sunojump.py -i ./my_songs --spectrogram
# Export before/after integrated loudness and true-peak evidence
python sunojump.py -i ./my_songs --loudness-report
# Export crest-factor and stereo-width evidence
python sunojump.py -i ./my_songs --signal-report
# Batch process a directory
python sunojump.py -i ./my_songs/ -o ./output/ -p moderate -f flac
# Watch a folder continuously; stable drops go to ./incoming/output/
python sunojump.py --watch ./incoming -p moderate
# With lossy re-encode
python sunojump.py -i song.wav -p aggressive --reencode 128
# Resume pending/interrupted work, or retry only failed/partial jobs
python sunojump.py --resume ./output/SunoJump_Batch_....sunojump-batch.json
python sunojump.py --resume ./output/SunoJump_Batch_....sunojump-batch.json --retry failed
# Explicitly allow a transformed output that omits source Content Credentials
python sunojump.py -i signed.wav --c2pa-policy allow-removal
# Select a locale for GUI or CLI text (qps-ploc is the RTL test locale)
python sunojump.py --locale en| Flag | Description | Default |
|---|---|---|
-i, --input | Input file or directory; required unless --resume or --watch is used | none |
--watch | Continuously process stable supported files dropped directly into a directory; conflicts with input/resume/manifest | none |
-o, --output | Output directory | ~/Desktop/SunoJump_Output |
-p, --preset | gentle, moderate, aggressive, extreme | moderate |
-f, --format | wav, flac, ogg, mp3, m4a (mp3/m4a require ffmpeg) | wav |
--preset-file | Path to custom JSON preset (overrides -p) | none |
--profile | Versioned JSON composition of a built-in preset plus sparse overrides; conflicts with -p and --preset-file | none |
--locale | Locale catalog for GUI labels and CLI help; falls back from region to language to English | system locale |
--compute | cpu, auto (CUDA with CPU fallback), or strict cuda for FFT-heavy passes | cpu |
--workers | Queued files to render concurrently, from 1 to 8 | 2 |
--result-format | human for stderr logs, json for one batch document, or jsonl for per-file records plus a final batch record | human |
--spectrogram | Export a before/after side-by-side PNG beside each usable output | off |
--loudness-report | Export an ITU-R BS.1770-5 before/after integrated-loudness and true-peak JSON report | off |
--signal-report | Export before/after crest-factor, mid-side-width, and stereo-correlation JSON evidence | off |
--manifest | Destination for a new batch manifest; must not already exist | generated in output directory |
--resume | Existing batch manifest to reconcile and resume | none |
--retry | With --resume: pending, unfinished, failed, or cancelled jobs | pending |
--c2pa-policy | block, or allow-removal to acknowledge that transformed outputs omit and do not re-sign source Content Credentials | block |
--diagnostic-retention | Number of persistent run logs to retain, from 1 to 365 | 30 |
--no-redact-diagnostics | Explicitly retain absolute paths in new CLI run logs | redaction enabled |
--no-spectral-scan, --no-watermark-scan | Disable the local narrowband candidate scan (--no-watermark-scan is retained for compatibility) | enabled |
--enable-pass PASS | Enable a named pass at its current/preset amount; repeatable | none |
--disable-pass PASS | Disable a named pass; repeatable | none |
--spectral | Spectral perturbation (0.0-1.0) | preset |
--spectral-sub-bass | Sub-bass spectral perturbation (0.0-1.0) | preset |
--spectral-low-mids | Low-mids spectral perturbation (0.0-1.0) | preset |
--spectral-presence | Presence-band spectral perturbation (0.0-1.0) | preset |
--spectral-air | Air-band spectral perturbation (0.0-1.0) | preset |
--dynamic-eq | Dynamic EQ amount (0.0-1.0) | preset |
--pitch | Pitch micro-shift in semitones (0.0-5.0) | preset |
--tempo | Tempo variation (0.0-0.15) | preset |
--phase | Phase scrambling (0.0-1.0) | preset |
--stereo | Stereo manipulation (0.0-0.5) | preset |
--noise | Noise level in dB (-70 to -30) | preset |
--dynamics | Dynamics amount (0.0-1.0) | preset |
--humanize | Humanization amount (0.0-1.0) | preset |
--reencode | Lossy re-encode bitrate (96-320) | disabled |
--seed | Non-negative integer for deterministic output; the effective per-file seed is always displayed | generated per file |
--native-runtime | Print machine-readable decoder version and containment policy, then exit | n/a |
Every numeric override enables its corresponding pass. To disable one, use --disable-pass with a name shown by --help; combining a pass's numeric value with its disable flag is rejected. Values, preset files, and profiles are validated strictly—unknown keys, wrong types, non-finite numbers, future schemas, and out-of-range values exit before an output directory or file is created.
Profiles make sparse tuning reusable without copying a complete preset. The built-in preset is applied first, the JSON overrides object second, and explicit numeric or pass CLI flags last:
{
"name": "Gentle tuned",
"schema_version": 1,
"preset": "gentle",
"overrides": {
"pitch_range": 0.5,
"stereo_enabled": true
}
}Profile documents reject unknown keys, future schemas, invalid values, and operation-level output_format or c2pa_policy settings. Choose those policies explicitly with their CLI flags.
Watch mode scans only the selected directory's top level. A file must keep the same size and modification time for one second before it is accepted, which avoids reading an in-progress copy. Existing files are eligible, unchanged files run once per watch session, and a later modification is eligible again. The default output is the watched directory's output/ child; using the watched directory itself as output is rejected to prevent generated files from feeding back into the watcher. Each accepted file gets an atomic one-job batch manifest and the normal C2PA, validation, sidecar, cancellation, and no-overwrite safeguards. Press Ctrl+C to stop and cancel active work safely.
GUI labels, tooltips, status/error text, accessibility descriptions, and CLI help share the JSON catalogs in locales/. Locale selection follows Qt-style fallback (language-region → language → English), and plural forms plus right-to-left layout direction are catalog controlled. --locale overrides the system locale for one run; SUNOJUMP_LOCALE provides the same override for launches that cannot pass arguments.
Source runs can move STFT/ISTFT work in spectral scanning/perturbation, Dynamic EQ, pitch phase-vocoder, phase scrambling, masking-aware noise, and constellation evidence to a CUDA-enabled PyTorch runtime. Install the platform-specific stable build using the official PyTorch selector, then run python sunojump.py -i song.wav --compute cuda. auto uses CUDA when the on-device SciPy parity check passes and otherwise records an explicit CPU fallback; strict cuda exits before creating batch state when PyTorch, CUDA, or the parity gate is unavailable.
The standard frozen release remains CPU-only: Torch is deliberately excluded from PyInstaller inventory, and release verification rejects executables above 250 MiB. Source GUI workers may opt in with SUNOJUMP_COMPUTE_BACKEND=auto or cuda. Sidecars record the selected library/device/runtime and treat accelerated renders as dependency-sensitive rather than promising byte identity across GPUs.
Use Save... in the GUI to export the current settings, then pass the resulting .json to --preset-file on the CLI to reproduce the same configuration across runs. Legacy partial presets are migrated and completed from the documented Moderate defaults; saved Custom sessions persist the full validated configuration rather than only the word Custom.
CLI exit status is 0 only when every job succeeds, 1 when a batch has at least one usable output but is partial, and 2 when no job produces a usable output. Human-readable progress, per-file summaries, and diagnostics always go to stderr, leaving stdout uncontaminated for --result-format json or jsonl. The com.sunojump.cli-results schema includes stable job IDs, states, error codes, effective seeds, validation evidence, artifact paths/hashes, counts, and the final batch state. Each independently versioned replay sidecar is written atomically and records validated audio shape, hashes, complete stochastic-pass trace, native/dependency versions, and replay constraints. A format-native audio tag contains the canonical sidecar-payload hash; the sidecar contains the final audio hash, providing a verifiable two-way binding.
Spectrogram export uses uniformly distributed Hann-window snapshots across the full file, at most a 2048-point FFT per image column, a logarithmic 20 Hz–20 kHz (or Nyquist) axis, and one fixed -100–0 dBFS color scale for both panels. The PNG is atomically published as <output-stem>.spectrogram.png; its SHA-256 and sampling metadata appear in JSON/JSONL results and batch manifests. A missing or changed audit image invalidates a previously successful manifest during recovery. Report failure leaves validated audio and its replay sidecar usable but returns a typed partial result.
The loudness report reuses the regression suite's ITU-R BS.1770-5 K-weighting, 400 ms absolute/relative gated integrated loudness, and sample-rate-aware 4×/2× true-peak oversampling. Filtering, power aggregation, and oversampling run in fixed-size blocks for long files. The atomically published <output-stem>.loudness.json records before/after LUFS, dBTP, oversampling factors, and after-minus-before deltas; its complete schema metadata and SHA-256 flow through the same structured results and recovery checks as spectrograms.
The signal-statistics report defines whole-signal and per-channel crest factor as 20×log10(absolute peak/RMS). For stereo material it also reports 10×log10(side energy/mid energy) with mid=(L+R)/√2 and side=(L−R)/√2, plus normalized interchannel energy correlation from −1 to +1. Mono, zero-side, and zero-mid cases are explicit rather than substituted with arbitrary values. Fixed-size accumulation keeps memory bounded; <output-stem>.signal.json includes before/after values, after-minus-before deltas, definitions, channel coverage, and hash-linked manifest/result evidence.
Every GUI and CLI batch also writes an atomic, versioned manifest. On recovery, an interrupted running job becomes pending, successful audio and sidecar hashes are revalidated, and missing or changed evidence becomes a failed job eligible for retry. Resume reuses the saved configuration and per-file seeds, rejects conflicting CLI overrides, and reserves a new collision-free output name rather than replacing any prior artifact.
Before any decode or transform, SunoJump performs bounded discovery of C2PA manifest stores in supported RIFF/WAVE, ID3-based MP3/FLAC, Ogg, AIFF, and adjacent .c2pa locations. Discovery reports present_unvalidated, absent, or inspection_failed; it does not validate signatures, trust chains, assertions, or hard bindings. A detected store is blocked by default. If explicitly allowed, the original remains byte-for-byte untouched while the re-encoded output omits and does not re-sign the source credentials. The replay sidecar records the source manifest hash, discovery status, policy, and decision. This behavior follows the C2PA 2.4 embedding specification and is not evidence of any acoustic-detector change.
The Session Log panel controls persistent diagnostics. Redact paths is enabled by default and masks registered input/output roots, the home and temporary directories, and other absolute paths. Privacy & Locations previews the current redaction and shows where every persistent category lives. Retain keeps 1-365 logs; cleanup failures are reported instead of ignored. Clear Logs requires confirmation, refuses to run during rendering, closes the active log lifecycle before deletion, and reports both failures and remaining files.
Export Support creates a no-replace ZIP atomically. Each included log is capped at 5 MiB, and the bundle records hashes, environment details, redaction state, and read failures. Audio, batch manifests, replay sidecars, and GUI settings contents are excluded by design.
- Run logs: the platform state/log directory shown by Privacy & Locations (
%LOCALAPPDATA%\SunoJump\logson Windows) - GUI settings: the QSettings location shown in the same dialog
- Batch history:
*.sunojump-batch.jsonin the chosen output directory - Replay evidence:
*.sidecar.jsonbeside processed audio in the chosen output directory - Preview audio: OS temporary storage, removed after worker exit and clean close
- Support bundles: the user-selected
.zipdestination
| Preset | Pitch | Spectral | Phase | Noise | Use Case |
|---|---|---|---|---|---|
| Gentle | 0.3 st | 0.10 | 0.10 | -60 dB | Lowest transform intensity |
| Moderate | 0.8 st | 0.30 | 0.30 | -50 dB | Moderate transform intensity |
| Aggressive | 1.5 st | 0.50 | 0.50 | -45 dB | High transform intensity |
| Extreme (default) | 3.0 st | 0.70 | 0.70 | -40 dB | Highest transform intensity; audition before use |
Preset names describe transform intensity, not effectiveness. Use Preview or Compare Presets and choose the lowest intensity that meets your audible and analytical goals.
After processing, SunoJump reports sunojump.signal_change v1, a normalized sample-domain difference derived from signal-to-difference ratio:
- 0-25% — low sample-domain change
- 25-50% — moderate sample-domain change
- 50-75% — high sample-domain change
- 75-100% — very high sample-domain change; audition output quality
This metric is not calibrated to any external service or recognition system. Sidecars record the adapter name, version, unit, and scope.
sunojump.constellation v1 compares up to 30 seconds of local landmark hashes. A measured result includes adapter/version, percentage, sample coverage, duration, alignment offset, and landmark counts. Silence, near-silence, clips shorter than two seconds, and inputs with too few landmarks report unavailable without a numeric value; adapter failures report error. These states appear identically in GUI/CLI logs and sidecar JSON and do not imply a platform outcome.
python tools/run_ab_regression.py --output ab-report.json renders two deterministic, repository-generated CC0 stereo music cues at 44.1 and 48 kHz using fixed seeds. The schema-versioned report contains per-pair and unrelated-cue negative-control detector scores, analyzed coverage, alignment offsets, and before/after ITU-R BS.1770-5 integrated loudness and true peak. Its thresholds are local quality/regression gates only and make no external-platform inference.
The suite supports the Apache-2.0 Google ViSQOL audio-mode CLI as its full-reference perceptual metric. Point --visqol-binary or VISQOL_BINARY at a locally built executable; adding --require-visqol makes absence or adapter failure fail the suite. Inputs are resampled to ViSQOL's required 48 kHz and reports identify the adapter by executable hash. Without the optional binary, the result is explicitly unavailable, never an invented quality score.
Input: WAV, MP3, FLAC, OGG, AIFF, Opus
For native-decoder safety, IRCAM payloads and WAV IMA ADPCM (including the extensible subtype) are not accepted. The file extension must match the detected container.
Output: WAV (24-bit), FLAC, OGG Vorbis, MP3, M4A/AAC
SunoJump's own source code is MIT-licensed. Runtime dependencies include copyleft-licensed packages:
| Package | License | Notes |
|---|---|---|
| PyQt6 | GPL-3.0-only | Bundled GUI toolkit |
| mutagen | GPL-2.0-or-later | Bundled metadata tag library |
| PyQt6-Qt6 | LGPL-3.0-only | Bundled shared Qt runtime |
Source installs: users install dependencies themselves via pip install -r requirements.txt. SunoJump source is MIT.
Binary releases (PyInstaller): the combined executable bundles GPL-licensed components. The binary as a whole must comply with GPL-3.0 terms. Source code is always available in this repository.
Run python tools/audit_licenses.py to verify the reviewed runtime inventory. Release packaging also inventories the frozen executable and fails if build-only, undeclared, or unreviewed components appear.
MIT