Music analysis becomes dangerous when a useful estimate is mistaken for ground truth. A tempo detector can be confident and wrong; a phrase grid can be structurally helpful without containing true downbeats; a mood label can be evocative without being factual. MixMan's analysis architecture is built around a more modest promise: machine observations should be rich enough to improve discovery, but specific enough about their origin and limits that they can be challenged, replaced, or ignored.

The deterministic path begins by decoding a track once. From the same audio buffer, the Queen Mary pipeline composes beat and key analysis, normalized energy, quality checks, similarity features, descriptor axes, and—when requested—cue, phrase, bar, and transition hints. There is deliberately no fictional all-knowing analysis call. MixMan assembles a versioned result from smaller capabilities and records best-effort failures in diagnostics rather than allowing one optional feature to erase the successful core.

Profiles keep that work proportional to the consumer. A summary can support review and scalar metadata. The default recommendation profile adds feature evidence useful for ranking. Timeline and full profiles add bounded structural events for more deliberate analysis, while raw matrix-style output remains an explicit debugging operation. This avoids turning every new track into an unbounded research artifact, and it lets richer analysis be introduced without changing the contract expected by lighter-weight clients.

The full result lives separately from the familiar track fields. Each analysis attempt records its subject, provider, profile, analysis and feature schema versions, analyzer version, source-file identity, status, warnings, diagnostics, scalar publication report, and completion time. Only one successful result is current for a given subject and profile, but recent failed or stale attempts remain available for diagnosis. When a review item is promoted, its current analysis is copied to the track and the change in media identity is recorded, rather than pretending the analysis was freshly produced after publication.

Scalar fields such as BPM, key, duration, energy, colour, and mood still matter because people and older clients use them. Publishing those values is governed by three policies. empty_only fills gaps. analyzer_owned may also refresh values previously written by the same analyzer. overwrite_all is a deliberate administrative rebuild. Per-field provenance records who last claimed a value and which analysis produced it. Just before publication, the worker locks and rereads the track or review row, so a human edit made during a long analysis remains authoritative instead of losing a race to an older machine result.

That ownership boundary continues into the interface. An explicit human set operation claims a field even if its value happens to match the current analyzer estimate. Clearing a field removes that claim and makes later automation possible again. Semantic enrichment is treated as another layer, not as a replacement for deterministic evidence: it can receive analysis hints, but its provenance stays distinct, and relevant semantic results are marked stale when the underlying deterministic analysis is rebuilt. Neural embeddings likewise remain separate from Queen Mary feature vectors because their dimensions, meaning, and lineage are different.

Explainability appears in recommendation evidence as well as storage. Candidate comparisons can report component scores for rhythm, timbre, chroma, energy, onset, section, tonality, tempo, and key. Schema mismatches and missing components are explicit states. Reason codes can identify a tempo jump, key clash, energy cliff, phrase mismatch, or low confidence instead of compressing every musical relationship into one unexplained number. The policy layer can then combine that evidence with session intent without claiming that any single component chose the track.

Quality warnings have product consequences. Severe signals can block automatic approval and remove a track from recommendation; less severe warnings can down-rank it. If the primary analysis fails terminally, legacy scalar jobs remain available as a fallback. Persisted timelines and transport state can also drive projections without rerunning digital signal processing inside a playback request, and their external exposure has its own control. Analysis, recommendation weighting, session planning, and projection do not have to be rolled back as one indivisible feature.

The result is music intelligence with an audit trail. MixMan can use more of what the audio reveals while preserving the difference between measurement, heuristic, semantic interpretation, and human curation. For me, that distinction is the real value of explainability: not a paragraph invented after a recommendation, but a system that knows which evidence it used, who is allowed to revise it, and how to behave when the evidence is incomplete.