Music
Music owns the route from musical intent to an accepted musical work. It can begin with a generated sketch, a played take, MIDI, a score, a remembered moment, or an empty session. None of those methods receives privileged authorship.
Compose and arrange
riffmaw.forge_track@2 pins the music brief, admitted sources, FeelProfile@1, target form,
budget, and acceptance criteria. AI music generation may propose motifs, harmony, rhythm, timbre,
sections, or complete candidates. Human performance, manual composition, procedural systems,
sampling, and DAW editing may answer or replace those proposals. Every generated or transformed
artifact keeps its provider or tool revision, control material, seed when available, input and
output digests, cost, and source influence.
A FeelProfile@1 may begin as explicit direction or be revised from admitted, correctable
observations of a take or jam. It can describe pocket, density,
attack, space, motion, instrumental relation, tension and release, and when the machine should
listen rather than answer. Each correction creates a linked immutable revision. It never converts
a model's guess about a performance into a fact about the performer.
The Magus selects and arranges the musical matter. Structure, tempo map, meter, key or tuning, instrumental roles, transitions, repetition, and controlled variation remain editable decisions. Generated output is a candidate, not a finished track and not evidence that reuse is safe.
Language belongs to the musical request
The music-brief language, structured music-description language, lyric language, requested sung language, and any phonetic or transliteration aid are separate typed facts. Riffmaw preserves the original words. It may place an exact versioned Translation Spell for a lexical derivative, but lyric adaptation and musical fit remain Riffmaw judgment. The derivative retains its author, implementation revision, rhyme, metre, pronunciation, and meaning-loss findings; it never silently replaces lyrics merely because a generator prefers another prompt language.
Each exact music profile declares and separately verifies description understanding, lyric rendering, pronunciation, structure tags, mixed-language behavior, and instrumental mode. Altar locale and Persona language do not select these values. Model-generated or optimized lyrics are a separate authored candidate, not an invisible convenience flag and not permission to change words already approved for performance.
Singing, rap, and spoken-word used as musical material remain Riffmaw performance. Ordinary dialogue, narration, spoken dialogue replacement, captions, and spoken timed-language editions belong to Language Edition. When a translated song enters a wider language edition, Riffmaw retains its lyric and sung-performance revision while Language Edition retains source alignment, captions, edition fit, and packaging. Their handoff carries exact lyrics, performance, timing, and findings rather than merging offices.
Answer and develop
During an admitted jam, Riffmaw may generate or schedule accompaniment, counterpoint, percussion,
harmony, texture, effects, or transitions against the shared clock. A reply may use an exact recent
performance window, an accumulated submix, MIDI, an explicit cue, or the current FeelProfile@1.
The scheduled and actual entrance are both recorded so a musically useful late answer is not
misreported as realtime success.
The request declares one latency class rather than calling every machine reply realtime:
| Class | Contract |
|---|---|
| Inline DSP | Bounded processing inside a proved realtime-safe host; no model or control-plane wait enters the audio callback. |
| Streaming player | A model emits playable increments under a measured buffering and deadline contract. |
| Scheduled player | Riffmaw prepares a phrase, bar, or section ahead and fences its future entrance against the mapped musical clock. |
| Offline producer | No live deadline; generation and revision occur only after the captured session is frozen. |
Scheduled phrase- or bar-ahead accompaniment is near-live, not proof of sample-continuous realtime generation. The first credible Linux path may schedule symbolic MIDI into a deterministic synth before attempting generated audio inside the live deadline.
After capture closes, Riffmaw can develop the session without pretending the live moment remains open: select or comp takes, infer an editable tempo map, repair a bounded passage, extend an idea, generate supporting layers, arrange sections, and propose or execute a mix. Every step works from frozen artifact and session revisions. Returning to live performance requires a newly armed jam.
Production autonomy
The production request declares how far Riffmaw may finish on its own:
| Finish policy | Riffmaw may do | Required return |
|---|---|---|
assets_only |
Generate, capture, separate, clean, align, label, and package reusable musical material. | Dry and processed assets, stems where available, observations, and lineage; no implied mix approval. |
assisted_session |
Also arrange, route, automate, process, and render one or more proposed mixes. | Editable production plan and session state, stems, candidate mixes, findings, and unresolved choices. |
autonomous_master_candidate |
Also select bounded candidates, repair failed regions, mix, master, run deterministic gates, and repeat within budget. | A master candidate satisfying its automated gates, stems and receipts, or the exact criteria and budget that prevented completion; Magus acceptance remains separate. |
Autonomy means Riffmaw can close the production loop, not that one model emits a finished song in one shot. A planner can direct replaceable generators, analyzers, editors, renderers, effects, and mastering tools; deterministic probes and bounded listening judgments feed the next revision. Candidate count, wall time, compute, paid cost, similarity risk, loudness target, repair attempts, and stopping policy are declared before the loop begins. Exhaustion returns the best attributable candidates and findings when policy permits, never a quietly relabelled failure.
Each finish policy is a scope ceiling, not an authority grant. It never by itself arms capture, authorizes payment or Portal egress, admits a plug-in, permits export, accepts a master, or publishes the result. A failed bound implementation is replaced only by an exact predeclared branch in the pinned Scroll or by a new forward Invocation after settlement; a casting never silently changes its Resolution Lock.
Edit, mix, master
Editing preserves its parents: comping, timing changes, tuning, cleanup, resampling, and destructive processing each create an attributable revision. The mix owns routing, gain, pan, automation, production effects, spatial relation, and the balance among instrumental, vocal, sampled, and effected musical elements. A stem is an export of those relationships, not an independent master.
Mastering targets declared loudness, peak, dynamic, sample-rate, format, and sequencing profiles. The accepted output can include score or MIDI where present, session and edit lineage, stems, mix, and master. Deterministic probes establish the measurable facts; listening review still decides whether the work fits.
Paid generation, DAW export, and plug-in render are separate effects. Lost acknowledgement leaves the result unknown until the provider, session, or destination is inspected. A failed profile, unresolved similarity, stale source, or exhausted repair returns findings or refusal instead of a quietly degraded master.
Music passes accepted assets and mix relationships to Package, and exposes temporal events through Sync when another craft needs them.