LossyTrace is an experimental Rust library, CLI, and research harness for measuring traces that may be consistent with prior lossy audio compression. It emits evidence, not a provenance verdict.
Warning
The current measurement definition is feature version 0. It has known
content-dependence and invariance failures. LossyTrace does not currently
label files as lossless, lossy-derived, authentic, or fraudulent.
Build from source and inspect an audio file:
cargo run --release -- analyze track.flac --prettyThe JSON contract makes the experimental status explicit:
{
"schema_version": 1,
"state": "experimental_measurements_only",
"public_verdict_enabled": false,
"compression_trace": {
"feature_version": 0
}
}By default the CLI inspects at most 120 seconds. Pass --max-seconds 0 to
inspect the complete file. The report includes a SHA-256 input fingerprint but
does not include the input path.
src/: the verdict-free compression-trace measurements and CLI.benchmarks/audio-integrity-v1/: reproducible benchmark contracts, manifests, fingerprints, and aggregate development reports.scripts/: corpus staging, controlled lossy-to-lossless generation, evaluation, safety gates, and research tooling.research/exact-transform/: source recovered from the final exact-transform research cycle, including the exact MP3 hybrid-filterbank probe.docs/research/: the original Reklawdbox plan and the research disposition.
The 33 research iterations were experiments, not 33 software releases. The
latest candidates remain rejected. See
docs/research/status.md.
No copyrighted or privately supplied audio belongs in this repository.
LossyTrace keeps only generators, provenance manifests, hashes, and aggregate
reports in Git. Real-audio corpora live in an external data directory and are
ignored. Research code with unresolved redistribution boundaries is also kept
outside the public repository. See
docs/data-governance.md and
docs/research/licensing-boundary.md.
cargo fmt --check
cargo clippy --all-targets -- -D warnings
cargo test --all-targets
python3 -m unittest discover -s scripts/tests -p 'test_audio_integrity*.py'Some corpus and benchmark checks are intentionally opt-in because they require licensed or locally retained audio. They must fail closed when their declared inputs or fingerprints are missing.
This work began as an opt-in stratum-dsp prototype inside Reklawdbox. It was
extracted because detector research has a different release cadence and risk
profile. Reklawdbox must not consume a LossyTrace verdict until a fresh,
independent validation gate passes and a stable evidence contract is released.
Licensed under either the Apache License, Version 2.0 or the MIT License, at your option.