{"schemaVersion":"1.0","name":"MR Maker manual sample-evaluation checklist","canonicalUrl":"https://mrmrmaker.com/methodology/protocol.json","humanReadableUrl":"https://mrmrmaker.com/methodology","lastReviewed":"2026-08-22","scope":"A public checklist for future quality research and manual evaluation of authorized samples. It is not an automated control applied to every customer job, a copyright clearance, or a guarantee of perfect separation.","execution":{"mode":"manual_review_of_authorized_samples","appliesToEveryCustomerJob":false,"automaticFullFileDecodeForEveryJob":false,"automaticMatchedPassageListeningForEveryJob":false,"evidenceRecordCreatedForEveryJob":false,"intendedUse":"Use this checklist when MR Maker explicitly conducts a permitted sample review or a future documented benchmark. Routine customer processing does not imply that these review stages were performed."},"sourceEligibility":["The file uses a supported format and stays within the published file-size and duration limits.","The audio can be decoded from beginning to end.","The user confirms they are authorized to process the source audio."],"stages":[{"id":"source-validation","method":"For a selected authorized sample, a reviewer records the source format, duration, and decode result before processing so the output can be compared with the same source.","evidence":["format","duration","decode result"]},{"id":"technical-integrity","method":"For that manually evaluated sample, decode the generated file through its full duration and review for truncation, unexpected silence, clipping, or a broken container.","evidence":["full-file decode result","output duration","silence or clipping review"]},{"id":"matched-passage-listening","method":"A reviewer compares matching passages near the beginning, middle, and end at a similar listening level, noting lead-vocal residue, backing-vocal bleed, drum or bass damage, and stereo-width changes.","evidence":["beginning passage","middle passage","ending passage","review notes"]},{"id":"outcome","method":"Classify the evaluated sample as pass, manual review, or retry. A pass applies only to that recorded sample and means it passed the stated checks; it does not mean that every trace of vocals was removed.","evidence":["outcome","reason","review timestamp"]}],"outcomes":{"pass":"The manually evaluated sample passed the recorded generation and integrity checks.","manualReview":"The evaluated sample is playable, but audible artifacts or source conditions require further listening review.","retry":"The evaluated sample failed generation or an integrity check, so another source or processing attempt is required."},"knownLimitations":["Voices and instruments can share frequency and spatial information, leaving vocal residue or instrument damage.","Live recordings, strong reverb, lossy compression, and choirs are harder to separate.","Listening review is source-dependent and is not a substitute for rights clearance or release approval."],"benchmarkPublication":{"status":"not_published","aggregateScore":null,"explanation":"MR Maker does not currently publish an aggregate benchmark score. A score will not be stated until a licensed or openly permitted test set, fixed model versions, repeatable measurements, and the underlying result records can be published together."},"references":[{"name":"Demucs source-separation implementation","publisher":"Meta Research","url":"https://github.com/facebookresearch/demucs","reviewedAt":"2026-08-22"},{"name":"Hybrid Transformers for Music Source Separation","publisher":"arXiv","url":"https://arxiv.org/abs/2211.08553","reviewedAt":"2026-08-22"},{"name":"FFmpeg audio filter documentation","publisher":"FFmpeg","url":"https://ffmpeg.org/ffmpeg-filters.html","reviewedAt":"2026-08-22"}]}