Composable transition and effect discovery

Live update: Experimental learning recipes now connect scored practice scripts to simultaneous live playback, bass swaps and high-pass sweeps. Verified phrase sync and musical-quality validation remain incomplete. Earlier sequential-only descriptions below record the prior milestone.

Version one: implemented foundation

The Learning studio Technique discovery tab exposes a versioned catalog of 225 seed concepts across base transitions, EQ, filters, structure, loops, tempo, harmony, vocals, stems, layering, samples, deck motion, effects, energy intent and unknown discovery. This is a working vocabulary, not a universal exhaustive taxonomy or 225 implemented effects.

crates/evaluation/resources/techniques-v1.json is the source of truth. Every item states private-renderer support and live-autoplay support. Current live automation supports only sequential fades. Private recipes support 16/32-beat blends, bass handoffs and high-pass sweeps. Catalog-only effects cannot become device commands.

New completed private experiments emit technique-evidence.json and a scoped MongoDB technique_evidence record. The record contains:

• Exact renderer frame/time samples and per-deck fader, EQ, filter and transport state.
• Immutable recipe/version, PCM hashes, timing revisions, target tempo and engine fingerprint.
• Declared recipe components, independently measured technical metrics and output hash.
• Explicit synthetic-control origin, empty inferred labels, null cluster and no execution approval.

The time axis is seconds at 48 kHz; the whole audition is retained along with the precise transition interval. Bars/meter remain null. No normalization may erase the original trajectory. A later trajectory embedding should use a separate normalized-progress array, retaining absolute durations and source coordinates.

Evidence layers

1. Observation: a human played A then B; source/mix identities, cue alignment, estimates, corpus revision and audio rights remain explicit.
2. Control estimate: fader/EQ trajectories inferred from original A/B and the mixed recording. Store estimator version, alignment quality, reconstruction residual and uncertainty. These are estimates, not controller telemetry.
3. Synthetic truth: our rendered recipes and the actual control trajectory. Useful for implementation tests and controlled perturbations; not automatically a good-sounding positive example.
4. Hypothesis: multi-label classifier output, similarity embedding and stable unknown cluster ID. Each label has method/provenance; uncalibrated scores are not probabilities.
5. Human annotation: reviewer, ratings, comments and corrections. Retain conflicting opinions rather than overwrite them.
6. Executable recipe: a separately versioned, bounded program compiled for a specific engine capability set. Promotion requires deterministic regression checks, musical suitability checks and representative listening evaluation.

Use multiple dimensions: LONG_BLEND + BASS_SWAP + HP_FILTER_OUT, with source/target structure and energy intent separate. Do not equate CROSSFADE with all mixing, or derive verified bars from assumed meter. Filter controls normalized by our engine are not cutoff frequencies in Hz. Echo feedback, wet level, beat delay, scope and trajectories must be stored separately when that DSP exists.

Real-mix research: verified scope

The 2022 DJ mixer-estimation paper estimates time-varying fader and EQ gains by reconstruction, reporting a collection of 5,040 mixes and 50,742 transitions. This count describes that research corpus, not newly downloaded control curves in our deployment. The earlier transition-analysis implementation and DJtransGAN are useful references for estimation and differentiable mixing.

Our existing joint catalog is metadata. Recovering controls requires correctly matched source recordings and mix audio; those audio assets and estimated trajectories are not currently acquired. A residual can reflect alignment, mastering, compression, source-version mismatch or noise—not necessarily an echo/filter/loop. Residual effect labels need validation against synthetic effects and listening examples.

Next implementation stages

1. Match a small licensed/local set of exact A/B/mix recordings, retain match confidence, and benchmark fader/EQ reconstruction against synthetic ground truth.
2. Derive normalized trajectory features; cluster with UNKNOWN/noise support and retain original examples. Evaluate cluster stability and explainability before naming techniques.
3. Add bounded synthetic effects one DSP family at a time, with deliberate timing/gain failures as labeled test perturbations. Perturbed audio is not automatically musically bad in every context.
4. Learn contextual technique ranking from observations and representative human preference checks. Split connected recording/mix identities across train/validation/test to prevent leakage. Do not train and evaluate on the same generated pair variants.
5. Apply repetition penalties only within musically eligible candidates. Kimi chooses among available techniques and learned hypotheses; it does not invent real-time control coordinates or calibrated quality percentages.
6. Release additional live styles after engine support and evaluation. Discovery labels alone never enable an effect.

Continuous improvement without continuous human work

Private practice is uncapped by default, exclusive and idle-aware, with audio retention cleanup. Technical scoring runs for every experiment; Kimi reviews batches and prioritizes hypotheses. The human review shortlist samples strong, weak and middle technical outcomes. This is a heuristic spot-check workflow, not yet a trained quality critic or calibrated uncertainty sampler. Human feedback should steer direction occasionally; it is not required for every generated artifact.

The private worker seeds 36 additional immutable parameter variants: 16/32 beats × linear/equal-power fades × early/middle/late bass handoff × dry/soft/strong high-pass exit. These are bounded synthetic experiments using existing DSP, not techniques recovered from human recordings. Kimi can prioritize these variants, and exact pair/input/version combinations are still not repeated. This expands exploration without pretending the engine supports echo, stems or loops yet.

Kimi artistic experiment advice

Every six completed private experiments, the optional Kimi batch reviewer compares measured results and a bounded set of recent human ratings. It proposes a musical intention and selects an existing parameterized recipe: duration in beats, fade curve, bass-swap placement and high-pass intensity. The prompt asks for an interpretable parameter difference, expected benefit and a listening or measurement check. It favors changing one factor at a time; this is advisory rather than an enforced experimental design.

The stored hypothesis includes the reference attempt, its recipe, the proposed recipe and both recording hashes. The worker prioritizes that pair and recipe on its next eligible pass. The Learning studio shows the intention, proposed experiment and what to check. An already completed exact experiment is skipped. Kimi's opinion neither changes the technical score nor approves a recipe for live execution. Suggestions remain hypotheses until compared with rendered evidence and occasional human feedback. This version chooses among bounded variants; it does not generate arbitrary DSP scripts or train a new neural model.