livepilot 1.26.0 → 1.26.1

package/livepilot/skills/livepilot-core/references/automation-atlas.md

@@ -0,0 +1,272 @@
+# Automation Atlas — LivePilot v1.6
+
+Complete knowledge corpus for musically intelligent automation. Load this reference when working with automation tools.
+
+---
+
+## 1. Curve Theory
+
+### Why exponential for filters
+Filter frequency perception is logarithmic (octaves). 100Hz → 200Hz is the same perceptual distance as 1kHz → 2kHz. An exponential curve in the normalized 0–1 range maps to perceptually even movement across the frequency spectrum. Linear automation on a filter sounds like it's rushing through the highs and crawling through the lows. Use `exponential` with `factor` 2.0–3.0 for filter cutoff.
+
+### Why logarithmic for volume
+Human loudness perception follows a logarithmic curve (Weber-Fechner law). -6dB = half perceived loudness, not half amplitude. A logarithmic fade-in sounds smooth; a linear fade-in sounds like nothing happens then suddenly gets loud. Use `logarithmic` with `factor` 2.5–3.5 for volume fades.
+
+### Why S-curve for crossfades
+Natural acceleration/deceleration. Avoids the "hole in the middle" of linear crossfades where combined energy dips. The smoothstep function (3t² - 2t³) ensures the rate of change is zero at both endpoints, so the transition feels organic. Use `s_curve` for any A→B crossfade.
+
+### Why spike for throws
+Dub production technique — instant send level spike creates a single reflection that decays naturally through the reverb/delay tail. Any other shape creates unnatural sustained reflections. The exponential decay (peak × e^(-decay×t)) models how real acoustic energy dissipates.
+
+### Pan is linear
+Stereo position perception is roughly linear. A pan pot moving at constant speed sounds even. No need for perceptual correction on pan automation.
+
+### Resonance is dangerous
+Filter resonance is non-linear — subtle changes at low values, dramatic and potentially destructive at high values. Self-oscillation begins above ~0.85 on most filters. Never automate resonance above 0.85 without explicit user intent. Use `breathing` recipe with reduced amplitude (0.05–0.10) for subtle resonance movement.
+
+---
+
+## 2. The Perception-Action Loop (5 levels)
+
+**MANDATORY for all automation work.** Never write automation blind — always perceive first.
+
+### Level 1: Reactive
+Single read, single action. The simplest loop.
+1. `get_master_spectrum` → read frequency content
+2. Identify one issue (e.g., "mud below 200Hz")
+3. Apply one automation (e.g., HP filter sweep)
+4. `get_master_spectrum` → verify improvement
+
+### Level 2: Diagnostic
+Multi-step investigation using EQ as a microscope. See Section 3 (Diagnostic Filter Technique).
+
+### Level 3: Verification
+Act → measure → adjust cycle. For every automation written:
+1. Read spectrum BEFORE
+2. Write automation
+3. Read spectrum AFTER
+4. Compare: did the problem frequency range improve?
+5. If not, `clear_clip_automation` → adjust parameters → try again
+6. Log the successful parameters in memory for reuse
+
+### Level 4: Cross-Track
+Solo each track, build spectral map, write complementary automation.
+1. For each track: solo → `get_master_spectrum` → record fingerprint
+2. Find frequency overlaps between tracks (masking)
+3. Write complementary automation: as kick's filter opens, bass's filter narrows
+4. Verify no new masking was introduced
+5. Store the cross-track map in memory
+
+### Level 5: Full Pipeline
+Per-track analysis across entire session.
+1. Run Level 4 for all tracks
+2. Build session-wide spectral map
+3. Write automation considering all interactions
+4. Verify global mix spectrum
+5. Iterate until balanced
+
+---
+
+## 3. Diagnostic Filter Technique
+
+Using EQ Eight as a measurement instrument. Step-by-step:
+
+1. **Load EQ Eight** on target track via `find_and_load_device`
+2. **Configure band 1** as narrow bandpass: Q=8, gain=0dB, frequency=100Hz
+   - Use `set_device_parameter` to set Filter Type to Band Pass, Q to 8.0
+3. **Solo the track** via `set_track_solo`
+4. **Read spectrum** at 100Hz: `get_master_spectrum` → note the `sub` value
+5. **Sweep frequency** through key ranges:
+   - 100Hz, 200Hz, 300Hz, 500Hz, 1kHz, 2kHz, 5kHz, 10kHz
+   - At each: `set_device_parameter` (frequency) → `get_master_spectrum` → record
+6. **Build frequency map**: `{100Hz: 0.18, 200Hz: 0.25, 300Hz: 0.09, ...}`
+7. **Identify problems**: resonance buildup, mud, harshness, dead zones
+8. **Remove diagnostic EQ** — always clean up: `delete_device`
+9. **Un-solo the track**: `set_track_solo` (toggle off)
+10. **Write targeted automation** addressing what was found
+
+### What to look for:
+- **Mud** (200-400Hz): values > 0.20 suggest buildup → HP filter automation
+- **Resonance** (narrow peak in 300-800Hz): ringing → notch filter or gentle sweep
+- **Harshness** (2-5kHz): values > 0.25 suggest brightness → LP filter or shelf automation
+- **Dead zone** (any range with 0.00): frequency hole → boost or different device
+
+---
+
+## 4. Genre Recipes
+
+### Techno
+- `filter_sweep_up` on LP cutoff, 32 bars, factor 2.0 — the classic build
+- `sidechain_pump` on pad volume via Utility gain, 1 beat loop
+- `stutter` on vocals before drop, 0.5 beats, frequency 16
+- `stereo_narrow` on master bus, 8 bars before drop → instant widen at drop
+- Long transitions: combine HP filter + reverb send + stereo width, all exponential
+
+### Dub
+- `dub_throw` on Send A at each snare position, 1 beat duration
+- `tape_stop` on clip transpose for transitions, 0.5 beats
+- `washout` on delay feedback, 4 bars → cut at phrase boundary
+- `breathing` on filter cutoff, 4 bars — everything breathes in dub
+- Tip: never automate delay feedback above 0.85 (infinite feedback risk)
+
+### Ambient
+- `breathing` on filter cutoff, 4 bars, amplitude 0.10 — subtle is key
+- `perlin` noise on reverb mix, 8 bars, amplitude 0.15, seed varies
+- `brownian` drift on wavetable position, 16 bars, volatility 0.05
+- Polyrhythmic automation: 3-beat filter + 5-beat reverb + 7-beat pan = 105-beat cycle
+- Less is more. Ambient automation should be felt, not heard.
+
+### Hip Hop
+- `tape_stop` on clip transpose, 0.5 beats — classic vinyl stop
+- `vinyl_crackle` on Redux bit depth, 16 bars — lo-fi character
+- `sidechain_pump` on bass under kick, 0.5 beat — tight pocket
+- `fade_out` on filter for transitions between sections
+- Keep automation sparse — hip hop is about groove, not modulation
+
+### IDM / Experimental
+- `stutter` on volume, 0.5 beats, frequency 16 — micro-editing
+- `euclidean` on filter cutoff: 5 hits across 8 steps — rhythmic intelligence
+- `stochastic` on grain parameters, narrowing 0.7 — controlled chaos
+- `spring` on filter cutoff for overshoot character
+- Layer multiple polyrhythmic automations for generative evolution
+
+---
+
+## 5. Sound Design Automation
+
+### Wavetable position
+Exponential sweep for timbral morphing over 8–16 bars. The timbral change is often perceptually logarithmic, so exponential automation creates even-sounding evolution.
+
+### Grain size
+Sine modulation at 0.5–2Hz for alive textures. The slow oscillation prevents the grainular artifacts from sounding static. Use `breathing` recipe adapted: center 0.5, amplitude 0.15.
+
+### Reverb decay
+Link inversely to volume: quieter passages get longer tails for natural space. As volume drops, reverb decay increases. Automate both with complementary curves.
+
+### Delay feedback
+Spike for throws (dub technique). NEVER exceed 0.9 on feedback — infinite feedback creates dangerous signal buildup. Use `dub_throw` recipe with peak capped at 0.85.
+
+### Filter resonance
+Subtle sine modulation creates vocal/crying quality. Watch for self-oscillation above 0.85. Use amplitude 0.05–0.10 for subtle character, 0.15–0.25 for expressive.
+
+---
+
+## 6. Arrangement Techniques
+
+### The Build (16–32 bars before drop)
+Combine multiple automations, all exponential:
+1. HP filter: 0→0.6 (remove low end gradually)
+2. Volume: 0.7→1.0 (subtle lift)
+3. Reverb send: 0→0.5 (add wash)
+4. Stereo width: 1.0→0.0 (collapse to mono)
+5. At drop: instant step back to defaults
+
+### The Drop (instant)
+Step automation restoring all build parameters to default values. No curves — immediate snap. Use `steps` curve with a single value at time 0.
+
+### The Strip (8–16 bars)
+Gradual HP filter rise removing elements one by one:
+1. First: HP filter on bass track (removes bass)
+2. Then: HP filter on drums (removes kick)
+3. Then: HP filter on pads (removes mids)
+4. Only high frequencies remain → transition point
+
+### The Crossfade (2–8 bars)
+S-curve volume automation between two sections:
+- Outgoing track: s_curve from 1.0 to 0.0
+- Incoming track: s_curve from 0.0 to 1.0
+- S-curve prevents the energy dip of linear crossfades
+
+---
+
+## 7. Micro-Editing and Humanization
+
+### Velocity vs volume
+Velocity is timbre — it changes how the sound is generated (harder hits, brighter tone). Volume automation is loudness — it changes amplitude after generation. They are different tools.
+
+### Per-note filter accent
+Automate filter cutoff to spike on accented notes. Use `spike` with duration matching the note length. Creates timbral accents without touching velocity.
+
+### Spatial punctuation
+Send spikes on specific hits create spatial depth variation. Snare → reverb throw on beat 2 and 4. Hi-hat → delay throw on off-beats. Use `dub_throw` recipe.
+
+### Humanization via Perlin noise
+Tiny pitch/filter variations make programmed music feel human:
+- `perlin` on detune: amplitude 0.02, frequency 0.5
+- `perlin` on filter cutoff: amplitude 0.05, frequency 0.3
+- Different seeds per track — each voice drifts independently
+
+---
+
+## 8. Polyrhythmic Automation
+
+### The concept
+Unlinked automation envelopes with different loop lengths create long, non-repeating cycles:
+- 4-beat clip loop + 3-beat filter envelope + 5-beat reverb envelope
+- Total cycle: LCM(4, 3, 5) = 60 beats before exact repetition
+- 60 beats ≈ 15 bars at 4/4 — feels evolving, never static
+
+### Prime number rule
+Use prime-number beat lengths for maximum non-repetition:
+- 3, 5, 7, 11, 13 beats
+- LCM of primes = their product (all coprime)
+- 3 × 5 × 7 = 105 beats ≈ 26 bars before exact repeat
+
+### Implementation
+1. Create clip with N-beat loop (your base rhythm)
+2. `apply_automation_shape` with `duration` = M beats (M ≠ N, preferably prime)
+3. The automation loops at M beats, the clip loops at N beats
+4. Add another automation with duration = P beats (another prime)
+5. Total cycle = LCM(N, M, P) beats
+
+### Use cases
+Essential for ambient, installation, generative work. Creates the sense of "always evolving" that distinguishes produced music from loops.
+
+---
+
+## 9. Cross-Track Spectral Mapping
+
+### Process
+1. **Solo each track** one at a time
+2. **Read spectrum**: `get_master_spectrum` → record {sub, low, mid, high, presence, air}
+3. **Build spectral fingerprint** for each track:
+   ```
+   Kick:  sub=0.45, low=0.30, mid=0.05, high=0.02
+   Bass:  sub=0.35, low=0.40, mid=0.15, high=0.03
+   Pad:   sub=0.05, low=0.15, mid=0.35, high=0.25
+   ```
+4. **Find overlaps**: kick and bass both strong in sub/low = masking
+5. **Write complementary automation**: as kick opens, bass narrows
+6. **Verify**: un-solo all, check master spectrum for improvement
+
+### Complementary automation patterns
+- **Kick + Bass**: sidechain pump on bass volume, or complementary filter sweeps
+- **Vocals + Pads**: automate pad filter to duck in vocal frequency range (1–4kHz)
+- **Multiple synths**: different breathing rates (frequency 0.3, 0.5, 0.7) to weave in/out
+
+### Store findings
+Use `memory_learn` to save spectral maps and successful automation combinations. Build a knowledge base about the session's sounds for consistent decisions.
+
+---
+
+## 10. Golden Rules
+
+1. **Always use Utility gain for volume automation** — preserve the mixer fader for mixing. Automate the Utility device's Gain parameter, not track volume.
+
+2. **Exponential for filters, logarithmic for volume** — never linear for either. This is perceptual science, not preference.
+
+3. **Subtle automation (5–15% range) for organic feel; dramatic (full range) for transitions.** Most automation should be felt, not heard. Save the big sweeps for builds and drops.
+
+4. **ALWAYS verify with `get_master_spectrum` after writing automation.** The feedback loop is mandatory, not optional.
+
+5. **Use `clear_clip_automation` before rewriting.** Don't stack conflicting curves — clear first, then write fresh.
+
+6. **Use the diagnostic filter technique before guessing at problem frequencies.** Measurement beats intuition.
+
+7. **Store spectral findings in memory.** Build a knowledge base about this session's sounds. Cross-session awareness makes every future decision better.
+
+8. **Delay feedback NEVER above 0.9.** Infinite feedback creates dangerous signal buildup. Cap at 0.85 for safety, even in experimental contexts.
+
+9. **density 16–32 for most curves.** 16 points per bar gives smooth results for gentle curves. 32 for fast modulations or detailed shapes. Above 64 is rarely worth the overhead.
+
+10. **When in doubt, use a recipe.** The 15 built-in recipes encode production wisdom. Start with the recipe, then customize if needed.

package/livepilot/skills/livepilot-core/references/concepts/_schema.md

@@ -0,0 +1,158 @@
+# Concept Packet Schema
+
+Concept packets are structured YAML files that turn prose entries in
+`artist-vocabularies.md` and `genre-vocabularies.md` into machine-readable
+payloads the Creative Director compiles into briefs and diversity seeds.
+
+Files live in:
+
+```
+livepilot/skills/livepilot-core/references/concepts/
+├── artists/<slug>.yaml       # per-artist packets
+├── genres/<slug>.yaml        # per-genre packets
+└── _schema.md                # this file
+```
+
+The narrative .md files stay — they are human-facing overviews. The YAML
+is the source-of-truth for director compilation.
+
+## Loading
+
+The director's Phase 1 concept-packet load step resolves user references
+as follows:
+
+1. Normalize the user's reference (artist or genre, lowercase, hyphenate spaces).
+2. Check `artists/<slug>.yaml` first, then `genres/<slug>.yaml`.
+3. If no exact match, check the `aliases` field of each packet.
+4. If still no match, fall back to the narrative .md file and flag
+   the packet as unstructured.
+
+## Schema
+
+```yaml
+# REQUIRED
+id: str                          # e.g. "dub_techno__basic_channel"
+name: str                        # human-readable, e.g. "Basic Channel / Rhythm & Sound"
+type: artist | genre | hybrid
+aliases: list[str]               # alternate names for lookup (can be [])
+
+# The sound IS — not what to do, what it sounds like
+sonic_identity: list[str]        # bullet points describing the sound
+
+# What the director reaches for — populates Phase 3 candidate pool
+reach_for:
+  instruments: list[str]         # Ableton/pack device names — must exist in atlas
+  effects: list[str]             # effect device names
+  packs: list[str]               # pack names for atlas_pack_info lookup
+  utilities: list[str]           # Utility, Spectrum, EQ Eight, etc. for diagnostic chain
+
+# Hard filter — candidate plans failing these get dropped in Phase 3
+avoid: list[str]
+
+# Stylistic tendencies — inform plan descriptions, not mandatory
+rhythm_idioms: list[str]
+harmony_idioms: list[str]
+arrangement_idioms: list[str]
+texture_idioms: list[str]
+
+# For sample-heavy aesthetics — what the sample plays in the track
+sample_roles: list[str]          # texture_bed | transient_ghost | dub_tail_source | ...
+
+# Cross-reference to the technique catalog
+key_techniques:
+  - name: str                    # e.g. "dub_throw"
+    source: str                  # sample-techniques.md | sound-design-deep.md | atlas
+    device: str                  # when source is atlas — the device name
+    notes: str                   # optional inline hint
+
+# How the director weights the goal vector for this aesthetic
+evaluation_bias:
+  target_dimensions:             # weights (sum loosely to 1.0)
+    depth: float
+    groove: float
+    motion: float
+    contrast: float
+    novelty: float
+    cohesion: float
+    clarity: float
+  protect:                       # floor values — below these = identity break
+    clarity: float
+    cohesion: float
+    low_end: float               # packets that require bass weight
+
+# How the director biases the 3-plan family diversity
+move_family_bias:
+  favor: list[str]               # families to prefer (subset of the six canonical)
+  deprioritize: list[str]        # families to avoid unless user explicitly asks
+
+# 4-move rule coverage expectation
+dimensions_in_scope: list[str]   # from {structural, rhythmic, timbral, spatial}
+dimensions_deprioritized: list[str]  # from the same set — for narrow-idiom packets
+                                      # (dub-techno deprioritizes rhythmic; ambient
+                                      # deprioritizes rhythmic + structural; beat-
+                                      # focused deprioritizes spatial)
+
+# Default novelty budget when no user framing is present
+novelty_budget_default: float    # 0.0 – 1.0
+
+# Tempo hints — optional, mainly for genre packets
+tempo_hint:
+  min: int
+  max: int
+  time_signature: str            # "4/4", "2-step", etc.
+
+# Cross-links
+canonical_artists: list[str]     # genre packets list representative artists
+canonical_genres: list[str]      # artist packets list their genres (slugs from genres/)
+
+# Free-form prose the director can read when composing the brief identity line
+notes: str
+```
+
+## Field rules
+
+- `type: artist` packets MUST have `canonical_genres` populated.
+  `canonical_artists` is optional for artist packets (typically `[]` or
+  omitted — putting `[]` is redundant but accepted for schema symmetry).
+- `type: genre` packets MUST have `canonical_artists` populated.
+  `canonical_genres` is optional for genre packets.
+- `move_family_bias.favor` + `move_family_bias.deprioritize` should not
+  overlap.
+- `dimensions_in_scope` + `dimensions_deprioritized` should together
+  cover all four dimensions (structural / rhythmic / timbral / spatial).
+- `evaluation_bias.target_dimensions` weights should loosely sum to 1.0
+  (not strictly enforced — they're ratios).
+- `evaluation_bias.protect` floor values are 0.0 – 1.0 where higher
+  means more strictly protected.
+- `novelty_budget_default` applies only when the user's turn contains
+  NO explicit framing mapping to the `creative-brief-template.md`
+  novelty table. User framing always wins.
+- `tempo_hint` should reflect the actual artist's / genre's canonical
+  tempo range, not a genre-general band. If you set 120-130 for an
+  artist whose tracks are 95-110, the agent will send the wrong tempo.
+
+## Relationship to narrative .md files
+
+- `artist-vocabularies.md` — narrative overview, human-facing
+- `genre-vocabularies.md` — narrative overview, human-facing
+- `concepts/artists/*.yaml` — structured, machine-loaded
+- `concepts/genres/*.yaml` — structured, machine-loaded
+
+When a new artist or genre is added, update BOTH. The narrative .md
+gives a reader an overview; the YAML gives the director an executable
+packet. They should not disagree.
+
+## Sync check (future CI)
+
+A test in `tests/test_concept_packets.py` should verify:
+
+1. Every `### <name>` heading in `artist-vocabularies.md` has a matching
+   `artists/<slug>.yaml`.
+2. Every `## <name>` heading in `genre-vocabularies.md` has a matching
+   `genres/<slug>.yaml`.
+3. Every YAML parses cleanly and has all required fields.
+4. Every `canonical_artists` / `canonical_genres` cross-reference
+   resolves to an existing packet.
+5. Every `key_techniques.name` resolves to either a line in
+   `sample-techniques.md` / `sound-design-deep.md` or an atlas
+   `signature_techniques` entry.

package/livepilot/skills/livepilot-core/references/concepts/artists/akufen.yaml

@@ -0,0 +1,116 @@
+id: microhouse__akufen
+name: Akufen (Marc Leclair)
+type: artist
+aliases:
+  - "akufen"
+  - "marc leclair"
+
+sonic_identity:
+  - micro-second vocal and radio clippings triggered percussively
+  - melodic AND rhythmic content emerge simultaneously from the same chop source
+  - '"My Way" album is the canonical reference'
+  - no coherent vocal survives — everything is syllable-or-shorter
+
+reach_for:
+  instruments:
+    - Simpler  # slicing mode, 1/64 length
+  effects:
+    - Snipper  # Building Max Devices
+    - PitchLoop89
+    - Auto Filter  # on drum bus
+    - Convolution Reverb
+  packs:
+    - Voice Box
+    - Chop and Swing
+    - Latin Percussion
+    - Lost and Found
+  utilities:
+    - Utility
+
+avoid:
+  - coherent vocal samples (longer than one syllable)
+  - pitched singing
+  - long-sustain material
+  - stock presets
+
+rhythm_idioms:
+  - micro-chopped vocals triggered as percussion
+  - 4/4 underneath but the chops override the grid
+  - off-grid re-trigger via probability
+
+harmony_idioms:
+  - implied tonality via chopped-syllable pitch relationships
+  - no conventional chord changes
+
+arrangement_idioms:
+  - constant chop-pattern variation
+  - single loop evolves through retriggering, not adding layers
+
+texture_idioms:
+  - entire track IS texture — vocal fragments fill the sonic field
+  - reverb/filter sends tucked at -20 to -30 dB
+
+sample_roles:
+  - rhythmic_chop
+  - melodic_chop
+  - transient_ghost
+
+key_techniques:
+  - name: "Vocal micro-chop (Akufen)"
+    source: atlas
+    device: Simpler
+    notes: the technique's name literally comes from this artist
+  - name: "micro_chop"
+    source: sample-techniques.md
+  - name: "stab_isolation"
+    source: sample-techniques.md
+
+evaluation_bias:
+  target_dimensions:
+    groove: 0.24
+    novelty: 0.18
+    motion: 0.16
+    depth: 0.12
+    contrast: 0.12
+    cohesion: 0.10
+    clarity: 0.08
+  protect:
+    clarity: 0.50
+    cohesion: 0.60
+    low_end: 0.60
+
+move_family_bias:
+  favor:
+    - arrangement
+    - sound_design
+    - device_creation
+  deprioritize:
+    - mix
+    - performance
+
+dimensions_in_scope:
+  - rhythmic
+  - timbral
+  - structural
+dimensions_deprioritized:
+  - spatial
+
+novelty_budget_default: 0.65
+
+tempo_hint:
+  min: 122
+  max: 128
+  time_signature: "4/4"
+
+canonical_artists: []
+canonical_genres:
+  - microhouse
+
+notes: |
+  Akufen asks reward extreme source-material discipline: the sample is
+  the instrument, and the patch work is in Simpler's slice envelope,
+  pitch ranges, and re-trigger probability. Plan distribution:
+  arrangement (clip-level chop re-sequencing with
+  dimension_hint="rhythmic"), sound_design (Simpler slice parameters
+  + Snipper drift), device_creation (load Auto Filter + Convolution
+  on the drum bus for the signature filtered-chop tail).

package/livepilot/skills/livepilot-core/references/concepts/artists/aphex-twin.yaml

@@ -0,0 +1,133 @@
+id: idm__aphex_twin
+name: Aphex Twin (Richard D. James)
+type: artist
+aliases:
+  - "aphex twin"
+  - "aphex"
+  - "richard d. james"
+  - "rdj"
+  - "afx"
+
+sonic_identity:
+  - fast shifting timbres
+  - melodic acid lines
+  - detailed micro-percussion
+  - gliding portamento leads
+  - FM-heavy sound design
+  - occasional brutal distortion
+
+reach_for:
+  instruments:
+    - Operator  # FM with 6+ operators if layered
+    - Analog  # 303-style acid
+    - Simpler  # MIDI auto-trigger for rapid drum variation
+    - Wavetable
+  effects:
+    - Erosion
+    - Pitch Hack
+    - Roar  # distortion
+    - Microtuner
+    - Grain Delay
+  packs:
+    - Core Library synths
+    - MIDI Tools
+  utilities:
+    - EQ Eight
+
+avoid:
+  - single-textured drones (opposite of Hecker)
+  - slow-evolving pads (unless in ambient-phase track)
+  - preset factory drum racks
+  - conservative production
+
+rhythm_idioms:
+  - rapid drum variation via Simpler auto-trigger
+  - micro-percussion at 1/32 and 1/64
+  - breakbeat chopping
+  - Amen break variations
+
+harmony_idioms:
+  - acid basslines (303-derived)
+  - occasional microtonal moments
+  - melodic motifs with heavy portamento
+  - minor-key melodic phrases
+
+arrangement_idioms:
+  - fast timbral shifts — a section may change sound every 4 bars
+  - IDM-adjacent structural non-linearity
+  - ambient pieces on the opposite extreme (long sustain)
+
+texture_idioms:
+  - FM bells
+  - digital grain
+  - acid-bass squelch
+  - DX7-style bell clusters
+
+sample_roles:
+  - rhythmic_chop
+  - melodic_source
+  - percussive_ghost
+
+key_techniques:
+  - name: "Aphex ambient FM pad"
+    source: atlas
+    device: Operator
+  - name: "303 acid bass"
+    source: atlas
+    device: Analog
+  - name: "micro_chop"
+    source: sample-techniques.md
+  - name: "DX7 bell cluster"
+    source: atlas
+    device: Operator
+
+evaluation_bias:
+  target_dimensions:
+    novelty: 0.20
+    motion: 0.18
+    groove: 0.16
+    contrast: 0.14
+    depth: 0.12
+    cohesion: 0.10
+    clarity: 0.10
+  protect:
+    clarity: 0.55
+    cohesion: 0.55
+    low_end: 0.65
+
+move_family_bias:
+  favor:
+    - sound_design
+    - arrangement
+    - device_creation
+  deprioritize:
+    - performance
+
+dimensions_in_scope:
+  - rhythmic
+  - timbral
+  - structural
+  - spatial
+dimensions_deprioritized: []
+
+novelty_budget_default: 0.7
+
+tempo_hint:
+  min: 120
+  max: 170
+  time_signature: "4/4"  # or complex when user specifies
+
+canonical_artists: []
+canonical_genres:
+  - idm
+  - ambient
+
+notes: |
+  Aphex is the canonical IDM "all dimensions in scope" case. Plans
+  distribute freely. Typical three-plan set: sound_design (FM patch
+  on Operator), arrangement (rhythmic — probability + Simpler
+  auto-trigger with dimension_hint="rhythmic"), device_creation (load
+  Analog for 303-style acid line).
+
+  Ambient-phase Aphex (SAW II, etc.) requires re-reading the brief
+  as a modern_classical/ambient packet — user should disambiguate.