Sound Design Engine — Critic-Driven Patch Refinement

The sound design engine analyzes synth patches, identifies timbral weaknesses, and iteratively refines them through a measured critic loop. Every change is evaluated against the before state.

Analyzer Character Is the Main Signal

For broad quality requests, this skill is the primary route. "More punch", "warmer", "darker", "brighter", "less flat", "more alive", "more texture", and "more character" should become source/device/parameter decisions before they become volume moves.

When the analyzer is available, read character, not just level:

• get_master_spectrum for the 9-band contour
• get_spectral_shape for centroid, flatness, crest, rolloff, and brightness/noise shape
• get_mel_spectrum when EQ or source choice needs perceptual detail
• get_onsets for transient/envelope decisions
• get_novelty for movement/staticness decisions
• get_momentary_loudness only for safety/headroom/loudness context

Translate those measurements into musical moves:

• Bright/harsh character → filter contour, softer source, de-harshing, saturation tone; do not merely lower volume.
• Dark/dull character → oscillator/filter opening, excitation, air-band source, tasteful saturation; do not merely raise volume.
• Static/low novelty → modulation, envelope drift, LFO, generative device, granular/vector source.
• Weak punch → envelope/transient shaping, source layering, attack/release work; volume push is last.
• Flat/noisy spectrum → source substitution, subtractive filtering, simpler spectral role.
• Weak weight → instrument/register/source decision before master or track gain.

Atlas-first reflex (v1.23.x+, MANDATORY before any creative move)

Before producing ANY creative response, query the user's atlas overlays. The corpus contains 337 entries across 3 namespaces, plus 3,917 parameter-level JSON sidecars — far richer than anything inferable from training data alone.

Query order:

1. `extension_atlas_search(namespace="packs", query=<intent>)` — pack identity, signature workflows, hidden gems, anti-patterns, notable presets with macro deep-data, demo projects
2. `extension_atlas_search(namespace="packs", query=<intent>, entity_type="cross_pack_workflow")` — multi-pack signature recipes (15 entries: dub-techno spectral drone bed, BoC decayed pad, Mica Levi orchestral dread, etc.)
3. `extension_atlas_search(namespace="m4l-devices", query=<sonic descriptor>)` — M4L instrument/effect/midi-effect device catalog (155 entries)
4. `atlas_search(...)` — bundled atlas (Core Library, fallback)

Multi-grain traversal:

When an aesthetic-level query lands a pack-level result, AUTO-DRILL: pack → its notable_presets → those preset macro states → load via load_browser_item. Don't stop at "I found a relevant pack" — drill to the actual preset/parameter level the user can immediately use.

# Example — agent received "design a BoC pad — sublime, decayed, harmonic warmth"
hit = extension_atlas_search(namespace="packs", query="BoC sublime decayed pad harmonic warmth")
# → boc_decayed_pad cross-pack-workflow + inspired_by_nature pack

workflow = extension_atlas_get("packs", "boc_decayed_pad")
# → reveals signal flow + which notable_presets to start from

drone_lab = extension_atlas_get("packs", "drone_lab")
# → notable_presets reveals Razor Wire Drone with macros Filter Control=108, Movement=53...

# Now propose the patch with concrete preset names + macro starting values, not vague descriptions

When the user mentions a producer or pack by name:

• "BoC sublime pad" → atlas hit: boc_decayed_pad cross-pack-workflow + inspired_by_nature pack
• "Henke spectral chain" → atlas hit: pitchloop89 + granulator_iii + 2 Henke cross-pack workflows
• "Mica Levi orchestral dread" → atlas hit: mica_levi_orchestral_dread workflow + the orchestral suite packs
• "Drone Lab" → atlas hit: drone_lab pack + 4 Drone Lab demo_project entries

The atlas knows the user's installed library at parameter depth. Producer-anchor queries land specific moves, not vague descriptions.

Anti-pattern surfacing:

Every pack entry has an anti_patterns body field listing "don't reach for this when X." Surface the relevant anti-pattern when proposing a move so the user knows the move's domain. (E.g. "Drone Lab is sustain-only — don't use for percussive content.")

For deliberately rule-breaking creative requests ("eclectic", "ignore the limits", "weird combo", "mix incompatible aesthetics"): enter Eclectic Mode inside this sound-design loop. Anti-patterns become prompt tension rather than guardrails: keep hard safety and protected-user constraints, then pair one normally-avoided source or processor with one identity-preserving anchor. Do not route to a private or missing skill.

The Sound Design Critic Loop

Step 1 — Build Patch Model

Call get_patch_model(track_index) to build a PatchModel for the target track. The PatchModel maps every device on the track into typed blocks (oscillator, filter, envelope, lfo, saturation, spatial, effect) and classifies each as controllable or opaque.

Read the response carefully:

• blocks: ordered list of processing blocks with types and parameter names
• controllable_params: parameters you can modify via set_device_parameter
• opaque_blocks: third-party plugins where parameters may not map cleanly
• modulation_sources: detected LFOs, envelopes, and macro mappings
• signal_flow: how blocks connect (serial, parallel, or rack chains)

See references/patch-model.md for the full PatchModel structure and native device block map.

Step 2 — Analyze

Call analyze_sound_design(track_index) to run all sound design critics against the patch. The response contains an issues array with critic, severity, block, and evidence.

Five critics run during analysis. See references/sound-design-critics.md for thresholds:

• static_timbre — sound does not evolve over time, no movement
• no_modulation_sources — no LFOs, envelopes, or automation detected
• modulation_flatness — modulation exists but ranges are too narrow to hear
• missing_filter — raw oscillator output with no spectral shaping
• spectral_imbalance — too much energy in one frequency region, or gaps

Step 3 — Plan

Pick the highest-severity issue. Call plan_sound_design_move(track_index) with the issue. The planner returns a single intervention:

• move_type: one of the move vocabulary entries
• target_device: device index on the track
• target_parameter: parameter name or index
• target_value: the new value
• rationale: why this move addresses the issue

Move vocabulary:

• modulation_injection — add or increase LFO/envelope depth on a parameter
• filter_shaping — adjust cutoff, resonance, or filter type
• parameter_automation — create clip automation for time-varying timbral change
• oscillator_tuning — adjust pitch, detune, waveform, or unison settings

Step 4 — Capture Before

1. Call get_device_parameters(track_index, device_index) — save current parameter state
2. Call get_master_spectrum plus the relevant character streams above — save spectral snapshot (if analyzer available)

Step 5 — Execute

Apply the planned move using the appropriate tool:

• set_device_parameter for direct parameter changes (filter cutoff, LFO rate, oscillator shape)
• toggle_device for enabling/disabling processing blocks
• batch_set_parameters when the move requires coordinated changes (e.g., LFO depth + rate together)
• set_clip_automation for parameter automation moves
• find_and_load_device when the fix requires adding a new device (e.g., adding an Auto Filter)

Execute one move at a time. Verify before continuing.

Step 6 — Capture After

Repeat the same measurements:

1. Call get_device_parameters(track_index, device_index) — confirm the change took effect
2. Call get_master_spectrum plus the same character streams used before — save post-change spectral snapshot

Step 7 — Evaluate

Call evaluate_move(goal_vector, before_snapshot, after_snapshot) where goal_vector is the compiled goal from Step 1 and snapshots contain {spectrum: {...}, rms: float, peak: float}. Read:

• keep_change (bool): whether the change improved the sound
• score (0.0-1.0): quality improvement magnitude
• timbral_delta: what changed spectrally
• explanation: human-readable summary

Step 8 — Keep or Undo

If keep_change is false, call undo(). Explain what was tried and why it did not improve the sound.

If keep_change is true, report the improvement. If score > 0.7, consider calling memory_learn(name="...", type="device_chain", qualities={"summary": "..."}, payload={...}) to save the technique.

Step 9 — Repeat

Return to Step 2 only when the user asked for a deep refinement pass. In normal mode, stop after one meaningful character-improving move and summarize what changed plus the next optional direction. Avoid long loops of small parameter nudges.

Working with Opaque Plugins

Third-party AU/VST plugins may report as opaque in the PatchModel:

1. Check get_plugin_parameters(track_index, device_index) — some plugins expose parameters through the host
2. If parameter_count <= 1, the plugin is dead or unresponsive. Call delete_device and suggest a native alternative
3. If parameters are available but unnamed (Parameter 1, Parameter 2...), try map_plugin_parameter to identify them by ear
4. Report opaque status to the user — sound design critics cannot fully analyze what they cannot inspect

Quick Sound Design Checks

• "What's wrong with this sound?" — Call get_sound_design_issues(track_index) for a diagnostic without executing fixes
• "Show me the patch" — Call get_patch_model(track_index) then walk_device_tree(track_index) for full device chain visibility
• "What can I automate?" — Read the controllable_params list from the PatchModel response

Native Device Strengths

When adding processing blocks, prefer native Ableton devices for controllability:

• Wavetable — complex oscillator section with built-in modulation matrix
• Operator — FM synthesis with per-operator envelopes and LFO
• Analog — subtractive with two filters and two LFOs
• Auto Filter — standalone filter with envelope follower and LFO
• Corpus — resonant body modeling for physical character
• Erosion — high-frequency noise and distortion artifacts
• Saturator — waveshaping with multiple curve types

Always search_browser before loading — never guess device names.

M4L Instruments in the Library — When the Standard Synths Are the Wrong Aesthetic

Wavetable / Operator / Analog cover most subtractive and FM work, but several installed packs ship M4L instruments that produce sounds those three architecturally cannot. Reach for them when the standard list is the wrong starting point:

• Granulator III (Live Suite + Max for Live) — granular synthesis as a first-class instrument. Loop / Cloud / Classic modes, MPE per-note grain control, built-in audio capture. Use when the source needs to be a sample but evolve as a sustained voice.
• Harmonic Drone Generator (Drone Lab pack) — 8-voice M4L drone synth by Expert Math. Just intonation, Pythagorean, Pelog, equal temperament. Use when the patch is a sustained tonal bed and standard equal-tempered subtractive sounds wrong (microtonal beating is the point).
• Bouncy Notes (Inspired by Nature, Dillon Bastan) — gravity-based MIDI sequencer. Drop a ball, it bounces on a piano roll producing asymmetric never-repeating note cascades. Use as a generative source instead of writing notes by hand.
• Tree Tone (Inspired by Nature) — fractal-plant-growth resonator. Each branch is a tunable resonator (frequency / decay / amplitude). Use when you want resonance/body that evolves under itself, instead of static Corpus.
• Vector FM / Vector Grain / Vector Map (Inspired by Nature) — particle-physics modulation systems. Vector Map can route one particle to multiple parameters at once — useful when the patch needs coupled modulation that LFOs cannot produce.
• PitchLoop89 (Live Suite) — Henke pitch-shift delay (Publison DHM 89 emulation). Use as the spatial/pitched-echo block on any sustained voice when standard delay + Auto Pan is too rigid.

These do not replace the modulation_injection / filter_shaping / parameter_automation move vocabulary. They change what "the source" can be — which is upstream of the critic loop.

Deep Sound Design Reference

Consult references/sound-design-deep.md for advanced techniques when working on creative requests. Key principles:

Making Sounds Breathe

Every static sound can become alive with modulation below conscious perception:

• Filter breathing: LFO at 0.1-0.5 Hz on filter cutoff, 5-15% depth
• Oscillator drift: ±1-3 cent detune with very slow LFO (0.05-0.2 Hz)
• Amplitude micro-variation: Perlin/brownian noise on volume, ±1-3 dB
• Rule: If the listener can hear the modulation, it's too much. The best modulation is felt, not heard.

Space as Composition

Reverb and delay are not decorations — in dub/minimal they ARE the composition:

• Dub chord: Short stab → long delay (70-80% feedback) + filter on the delay return
• Delay throws: Momentary send spikes (0→70% for half a beat) — the echo IS the event
• Sidechain reverb: Dry drums trigger sidechain compression on reverb returns — the room pulses
• Feedback modulation: Delay feedback at 75-85% + modulate delay time ±5-10% for warped echoes

Creative Sidechain (Beyond Pump)

Sidechain compression is a modulation source, not just a mix tool:

• Sidechain filter: Envelope follower from kick modulates pad filter cutoff — pad brightens between kicks
• Ghost sidechain: Muted kick as sidechain source for textures — phantom groove on non-rhythmic elements
• Multiband sidechain: Only duck sub frequencies from pad — shimmer stays, sub clears for kick

Effects as Instruments

• Self-oscillating filter: Push resonance until it rings, play notes by changing cutoff
• Feedback loops: Route output back to input through effects + compressor to control
• Convolution as synthesis: Load non-IR files (speech, drum break) into convolution reverb — imprints spectral character
• Granular reverb: Very short reverb (0.1-0.3s) high diffusion on percussion — smears transient into tonal cloud

The Frequency Dance

At any moment, each frequency band should have one primary element. When one opens up, another pulls back:

• Chord filter opens into highs → pull hi-hat back
• Bass drops → kick shortens
• Reverb tail fills → dry elements duck

This is mix engineering as composition.

When to Apply These

• User says "make it breathe" or "it sounds static" → micro-modulation
• User says "more space" or "deeper" → dub techniques (delay throws, reverb composition)
• User says "more groove" or "make it pump" → creative sidechain
• User says "more texture" or "more complex" → textural layering
• User says "surprise me" or "WTF moment" → brief textural disruption (2-8 beats max)
• User says "warmer" or "more analog" → oscillator drift + subtle saturation + filter breathing