01 / CUE
Let the evidence speak.
No prompt required. Cue starts from music you loved, accepted, and actually entered, then chooses a useful direction for now.
Personal evidence → relevant cluster → five-track probe
Private beta · Apple Music
Music that understands the moment.
Start with a phrase, a Moment, one song, or simply press Cue. CUED turns your evidence + musical intent into five full tracks, then remembers what actually worked.
Five-track packsFull Apple Music playbackBuilt one listener at a time
Personal sessionPACK 01 / 05
What fits now
Your music, right now
Starts fromloves, accepts, and entered listening
Retrievesa personal cluster that fits this session
Learnswhich direction earns another pack
The bet
Shared music knowledge should widen what CUED can find. Your behavior should decide what fits.
CUED uses catalog identity and aggregated music relationships to propose candidates. Other listeners do not become your taste profile. Your connected library, explicit verdicts, and entered listening are the personal evidence that shapes your model.
One personal model · four paths
Ask for music in the way the thought arrives.
The lanes begin differently because “play something,” “2000s indie rock for a walk,” and “more like this exact Drake song” are different problems.
02 / MOMENT
A sound with an authored point of view.
CUED defines the musical center of a Moment. Your taste bends its artists, eras, and edges without erasing what the Moment means.
Authored intent → personal bend → stable identity
03 / CUSTOM
Say what you mean, even badly.
An AI interpreter turns natural language into explicit musical constraints and preferences. Retrieval and ranking choose the songs. The language model never does.
Language → structured intent → qualified candidates
04 / RADIO
Anchor to this recording.
Radio begins with the exact song, not a random sample from its artist. Song-level relationships expand outward; your taste orders the result.
Exact recording → song neighbors → personal order
AI interprets · ML retrieves and ranks
The language model never gets to make up the playlist.
CUED separates understanding a request from deciding what can play. Every candidate must come from a provider-traceable, playable music corpus and pass the same song-level eligibility path.
Turn human language into a testable request.
The LLM identifies what the listener named and what remains flexible.
- 01Recognizes exact artists, songs, eras, languages, activities, and energy.
- 02Separates named constraints from flexible preferences and unresolved meaning.
- 03Produces a versioned intent envelope the rest of the system can inspect.
Guardrail: the LLM may propose meaning. It cannot invent a recording, confer eligibility, or decide that a song belongs to a culture.
Find supply, score fit, and learn from exposure.
Retrieval and ranking combine multiple signals instead of trusting one model.
- 01Catalog identity, music graphs, audio embeddings, and personal exemplars propose candidates.
- 02Hard filters remove identity, playability, request, and rights failures before ranking.
- 03Lane-specific component scores and finite repeat penalties compose a five-track pack.
Learning contract under validation: entered songs lock; feedback may reorder only the unentered tail; thin supply triggers a fresh retrieval pass.
Illustrative Custom request: “2000s indie rock for a walk”
System contract under validation · not a live website query
01 · INTERPRETEra 2000s · indie rock · walking context · flexible energy
02 · PROPOSEQualified catalog + personal exemplars + source-labelled graph neighbors
03 · FILTERCanonical recording · playable · request-compatible · rights-clean
04 · RANKIntent fit + personal fit + seed fidelity + quality components
05 · COMPOSEStepwise choice · canonical dedupe · soft artist/album spread
06 · UPDATEEntered prefix locks · tail reranks · retrieval expands only when needed
Ranking objective
Discovery should widen taste without erasing identity.
CUED moves from music you already know, out through the scenes and artists around it, toward recordings you have not heard. The objective is new-to-you × good, never another route back to the same familiar hits.
score(candidate)
=
taste fit
×
moment fit
×
unfamiliarity
Deep catalog
How close a candidate sits to evidence you actually produced: your connected library, your explicit verdicts, and the songs you entered rather than skipped.
Artist + scene
Whether the recording belongs to what you asked for right now. Artist similarity stays constrained by scene and culture instead of drifting toward whatever merely sounds adjacent.
New to you
How far a candidate sits outside what you have already heard. Drop this term and the other two collapse into a familiar-hits engine that never moves.
Read honestly: 73% of served tracks fell outside the connected library. That means unowned, not proven unheard. Unfamiliarity is scored against the evidence CUED actually holds, which is your library, your verdicts, and your entered listening, never a claim about everything you have ever heard.
Evidence, not adjectives
Two ledgers. No blended claims.
Listening numbers describe the live beta path at its stated read date. Catalog-foundation numbers describe verified production data that is not yet the listener-facing D1.4 engine.
Live listening evidenceProduction snapshot · 2026-08-04
What four private-beta listeners actually received and judged.
8,012full-track servesAcross 726 sessions
2,940distinct artists servedObserved delivery
73%served outside the connected libraryUnowned, not claimed unheard
1,630explicit accept / reject verdictsBanked evidence
Catalog foundationVerified · 2026-08-28
What the next candidate engine can draw from once serving integration passes.
151,980tracks in the audio-embedding indexProvider identity mapped
1.38Msource-labelled music relationships ingestedNot yet a serving claim
117 / 139representative Radio seeds with song-level supplySparse seeds fail closed
50median candidates for supplied test seedsFive-song pack target
A real listening instrument
- Four listening lanes and five full Apple Music tracks
- Personal library, verdict, and entered-listening evidence
- Daily provider-traceable catalog qualification
One cross-lane candidate engine
- Song-level proposal union and eligibility
- Recorded component ranking and pack composition
- Feedback-aware tail update and bounded expansion
Improvement over time
- Better packs because of prior feedback
- Universal scene, freshness, or Radio coverage
- Cross-session repeat control for every listener
Your taste, on cue
Return to the music that formed you. Leave with something new.
CUED is a private research beta for people who want discovery to feel personal without becoming predictable.
Request beta access →iPhone · Apple Music subscription · free private beta