Sessions

This contract preserves what a person actually played, sang, rapped, performed as musical spoken word, or programmed before editing makes the result look inevitable.

Arm and capture

riffmaw.capture_take@2 starts only through an explicitly armed interface. It records the device profile, clocks, sample rate, latency, monitoring route, channel layout, plug-in state, session revision, and consent boundary. The immutable raw audio or MIDI is retained before transcription, timing correction, comping, cleanup, or transformation.

A PerformanceTake@1 distinguishes raw performance from observations about it. Tempo, pitch, timing, score or lyric transcription, and performance notes remain attributed interpretations with their source regions and uncertainty. They do not replace the recording. Capture authority is visible, time-bounded, revocable, and governed by the custody and hostile-audio rules in Audio.

A later jam may reference one or more settled takes inside MusicalPerformanceSession@1; capture does not create that session merely because both records contain audio or MIDI.

Choose without erasing

Comping and editing create new revisions that name every source take and selected region. Rejected takes remain subject to their own retention rule; an accepted comp never rewrites them. A changed device, clock, session state, or plug-in chain is explicit rather than smuggled into a continuation.

If capture state is uncertain after failure, Riffmaw stops the device and reconciles the artifact and session records before retry. It never assumes that silence means nothing was recorded.

Accepted takes anchor Music, including its instrumental and vocal work. The handoff shares exact artifact references and observations, not microphone authority, credentials, or an open DAW session. An ordinary spoken take intended for dialogue replacement or localization belongs to Language Edition rather than this musical session.

Jam in shared time

Live Jam is Designed beyond Audio's first record-and-send slice; no current Portfolio runtime or candidate stack proves it. riffmaw.open_jam@2 describes one bounded performance Pattern. Spellweaver separately admits its Invocation and exact Scroll; the Pattern cannot admit a Circle or nested production by itself.

A jam may contain microphones, guitars, keyboards, drum pads, controllers, MIDI, previously admitted clips, remote performers, and one or more machine players. Every route independently declares monitoring, recording, retention, analysis, transformation, remote transmission, and machine-response scopes. If preservation of a route is required but recording or retention is not admitted, Riffmaw refuses that route or the jam. Network presence is never recording permission, and the Magus cannot alter another performer's monitor or participation scope unilaterally.

MusicalPerformanceSession@1 binds the jam revision, participants and roles, assistance policy, model and tool allowlist, and these timing and capture records:

The shared jam mix, individual sources, MIDI and control events, model responses, and clock or route changes remain separately attributable. No PCM, MIDI stream, device handle, or plug-in state enters Graph state; checkpoints retain typed references, bounded sequence state, and receipts.

The admitted participation level for every affected route is one of:

• observe records and derives correctable musical observations but produces no audible reply;
• accompany may add a bounded rhythmic, harmonic, melodic, or textural answer;
• transform may process an explicitly selected live route while preserving its dry source; and
• conduct may propose section, tempo, cue, or arrangement changes, but cannot silently impose them on human performers.

Escalating the participation level requires a new visible admission by every affected Principal. A model may not turn observation into accompaniment, accompaniment into transformation, or a jam into autonomous publication by inference.

Performance as a correctable language

A MusicalGestureObservation@1 may interpret notes, chords, key or tuning, beat and pocket, accents, articulation, dynamics, timbre, phrase boundaries, repetition, tension, release, and the relation between players. It retains the exact source regions, feature and model revisions, latency, confidence, and competing interpretations. It is an observation about performed sound, not access to the performer's emotions or unspoken intent.

A JamResponse@1 binds the context available through an exact source time, its lookahead, jam and route epochs, turn sequence, the observation or explicit cue it answered, generation controls, ready-by deadline, scheduled window, planned and actual first audible sample, measured path latency, output acknowledgement, omission reason, and resulting audio or MIDI artifact. Performer corrections create a new immutable FeelProfile@1 instance linked to its predecessor: “follow behind my attack,” “answer only at phrase endings,” or “that distortion is energy, not a key change.” A correction links but never rewrites the performance, observation, or earlier profile.

This is the long route toward performance-native direction: a player can steer Riffmaw through timing, touch, phrasing, dynamics, repetition, and contrast rather than translating every musical decision into prose. Persistent personalization or training from those performances is a separate corpus-admission and model-promotion act; a jam alone authorizes neither.

The realtime stopping line

Generic callback, clock-epoch, output-fence, direct-monitoring, reconnect, and uncertain-playback law lives in Audio. Riffmaw adds musical beat and bar mapping, JamResponse@1 ready-by and scheduled windows, and a safe omission when a machine entrance is late. The human monitor route remains available without model cooperation. A partial MusicalPerformanceSession@1 preserves acknowledged takes and events; continuation is a newly armed forward Invocation with a new musical clock and route epoch.

Proving the jam

Use a focused fixture with an explicitly armed dry guitar DI, keyboard MIDI, two local performer roles, and one synthetic remote leg with controlled jitter, loss, and clock drift. Capture a ClockDomainMap@1, LiveAudioGraphProfile@1, MidiEventStream@1, CaptureManifest@1, and one scheduled JamResponse@1. Exercise xrun, device removal, clock-epoch loss, plug-in crash or latency change, late model output, and uncertain remote acknowledgement. Prove that direct monitoring continues, stale frames cannot enter the output route, raw audio and MIDI reconcile, and continuation requires a new Invocation. Export the attributable session, source artifacts, stems, raw MIDI, and neutral tempo and marker maps. This deterministic fixture proves state, fencing, and recovery; a later hardware bake must separately prove measured audio-device latency and musical usefulness.