Sunara Reference
Most fake-lossless checkers look at one thing: where the spectrum stops. Sunara Reference looks at three, and that difference is the whole design.
Lossy encoders low-pass the audio, so a hard cutoff below Nyquist is genuinely strong evidence of a transcode. But genuine recordings can stop early too — old masters, tape sources, gentle mastering filters. A detector that only measures the cutoff will convict them all. Any tool can catch fakes if it's allowed to accuse innocent files; the hard problem is catching fakes without doing that.
Sunara Reference runs five detectors grouped into three families, chosen because they fail in different ways:
Not just where the spectrum ends, but how. One detector tracks the effective high-frequency edge through the whole track — a codec's low-pass filter sits perfectly still, while natural roll-off wanders with the music. Another looks at the track's median spectrum for a cliff: a drop too steep and too deep to be acoustic.
What do the surviving high frequencies look like statistically? Transcodes tend to leave sparse, patchy HF content — bins missing, energy distributed unevenly — that differs measurably from the texture of a genuine recording.
How does the HF region behave over time? Quantization and block switching in lossy codecs leave a distinctive frame-to-frame jitter, and the content near the cutoff shows unnaturally consistent structure between frames.
Every detector reports two numbers: a verdict and a reliability — how much it should be trusted on this particular file. A quiet track with almost no HF energy gives a cutoff detector nothing to measure; it reports low reliability instead of guessing.
The fusion layer weights each detector by its reliability and applies one asymmetric rule: artifacts prove fake, but the absence of artifacts does not prove real. Calling a file fake requires strong evidence from one family with support from another. Calling it real requires broader agreement. And when the families genuinely conflict, the verdict is pulled toward the middle and the file is labeled Borderline — because a tool that admits uncertainty is more useful than one that always answers.
Every threshold in the engine was tuned against a controlled, paired dataset: 101 genuine hi-res FLACs and the same 101 files transcoded through MP3 192, MP3 128, and AAC 160. Overall accuracy is 96.5% under those conditions — including 93.1% on the genuine files it must not falsely accuse. One entire detector was built, evaluated, found to perform at chance level, and deleted. That's the standard everything else had to meet.
The full engineering story — including the detector that died and the assumption that turned out backwards — is in the article: One cutoff is not enough →