Fixes from Opus 4.6 review (C1-C6, I1, I3, I4, I8):
C1: static lastOutputs → per-instance lastOutputsForDelta member
C2: add_example() now on audio thread only (owns iml); worker reads
from mutex-protected staging area (stagedFeatures/stagedLabels)
C3: Worker reads stagedWeightsForWorker (not iml.get_weights()),
eliminating concurrent read/write on iml's MLP
C4: Worker spins on weightsPending before writing pendingWeights,
preventing double-write race
C5: RAND and CLEAR now enqueue Randomize/Clear jobs through worker
instead of directly mutating iml on the audio thread
C6: OSC callbacks stage JSON into oscStagedJson + atomic flag;
audio thread applies in process() (no recv-thread mutation)
Also fixed:
- I4: Separate pendingJob field (enqueueJob no longer overwrites currentJob)
- I8: Removed redundant swapReady atomic
- noiseLevel, cachedNovelty, cachedConfidence now std::atomic<float>
- Worker syncs examples back to iml after training via load_examples()
Replace direct-mutation threading with shadow IML:
- Background thread clones weights from main → shadow IML
- Training and perturbation operate only on shadow instance
- New weights staged in pendingWeights, swapped atomically by audio thread
- Thumbs-down now enqueues Perturb job instead of calling move_weights directly
- Audio thread applies new weights via iml.set_weights() at safe point
- Examples copied to shadow for training, results copied back as weights only
Threading invariant now fully enforced: background thread never writes to
the inference IML. Audio thread applies staged weights between inference calls.
Standalone test exercising IML inference pipeline without VCV runtime:
- Basic inference: all 12 outputs valid in [0,1]
- Input responsiveness: 12/12 outputs change between corners
- Randomize: weights change produces different output mapping
- Spread parameter: measurably different distributions at 0 vs 1
- Expressiveness: 6/6 corner pairs distinct, full range utilized
- Smoothness: 100% of sweep steps are smooth (no binary jumps)
Finding: spread sigmoid saturation effect is architecture-dependent
with small [16,24,16] network (filed as meml-l5a for investigation).
Full spec for MEMLNaut VCV Rack module based on interview:
- 2-8 configurable CV inputs, 12 raw + 5 derived CV outputs
- RL feedback via panel buttons + CV triggers with learn gate guard
- nisps-core C++ engine with background thread training
- User-configurable inference rate (block to audio rate)
- Bidirectional state transfer with companion webapp (file + OSC)
- 10-phase development plan from skeleton to distribution
- OSC pill: add status dot indicator (gray=off, green=connected,
pulsing yellow=connecting), use flexbox for reliable alignment
- OSC pill: add hover state for better affordance
- Synth panel: add separator line above OSC row so it doesn't
blend into the arpeggiator controls
- Bottom sheet: increase expanded height from 55vh to 60vh so
OSC row isn't cut off on mobile
- Help modal: update step 3 to mention the OSC pill in the
bottom bar as the primary connection method
Browser-side OSCOutput module sends parameter values over WebSocket to
a companion Deno bridge script that converts them to OSC/UDP messages.
Enables controlling SuperCollider, Max/MSP, Pure Data, TouchDesigner,
or any OSC-capable software from the NISPS playground.
- Browser module (js/synth/osc-output.js): WebSocket client with
auto-reconnect, throttle (~20fps), and dead-zone filtering
- Deno bridge (osc-bridge/bridge.ts): zero-dependency, compiles to
standalone binaries via deno compile for Linux/macOS/Windows
- Test receiver (osc-bridge/test-receive.ts): terminal dashboard
showing live OSC parameter values with bar charts
- OSC pill button in floating bar for quick connect/disconnect
- Help modal section with platform-aware download, setup guide,
and examples for SuperCollider/PD/Max
- GitHub Actions workflow for cross-platform binary builds
- OSC sends in both visual and synth modes
Design spec for input zoom, compound control axes, pinning system,
exploration noise controls, and visualization enhancements. Covers
the full parameter inventory with implementation phases and open
design questions for experimentation.
Re-apply fixes from ddb6b77 that were lost in the arpeggiator restore:
- play-btn: align-self: flex-start (pin to top)
- Play drawer: trigger on .play-btn:hover instead of parent :hover
(prevents preset select from opening the drawer)
- preset-select: align-self: flex-start (pin to top)
Restore the original arpeggiator as the default sequencer. ShapeSeq
code is preserved but gated behind a URL feature flag (?shapeseq=1):
- Arpeggiator import, state, and all wiring fully restored
- ShapeSeq modules loaded via dynamic import() only when flag is on
- ensureShapeSeqInit() creates engine/viz/UI lazily on first audio start
- animate loop routes inputs to ShapeSeq only when flag + playing
- setOutputMode shows/hides ShapeSeq container only when flag is on
- HTML: ShapeSeq container added but hidden by default
- CSS: ShapeSeq styles added (inert when container is hidden)
All 13 ShapeSeq modules in playground/js/shapeseq/ are untouched.
To test ShapeSeq: add ?shapeseq=1 to the URL.
- Remove Arpeggiator import and all arp references from a-app.js
- Add ShapeSeq imports (ShapeSeqEngine, StepVisualizer, ChainBuilderUI,
getDefaultBus) and lazy initialization on first audio start
- In synth mode: show ShapeSeq container (circular viz + chain builder),
hide flow-field particles. Visual mode unchanged.
- Route joystick or hand tracking inputs to shapeSeq.setSequenceInputs()
each frame; render step viz at 60fps
- Add HTML: #shapeseq-container with #shapeseq-viz canvas and
#shapeseq-chain wrapper, sequencer transport controls (play/stop,
tempo, step count)
- Add CSS: full-screen flex layout, viz takes upper 45%, chain builder
scrollable below, hidden by default
- Quick play button wires to ShapeSeq start/stop instead of arp
Closes meml-6ry. This completes ShapeSeq Phase 1.
Four modules that build on the Layer 0 foundation:
- primitive.js: base class with category system (generator/processor/
timing/converter), param schema validation, boundary enforcement
(clamp/wrap/scaled), symbolic process() interface, state management
for freeze, and applyBoundary() utility for delta control
- clock.js: AudioContext lookahead scheduling (25ms interval, 100ms
window) replacing setTimeout arpeggiator. Reads pattern descriptions
for per-step swing (timeOffset) and ratchet (subdivisions). Emits
seq.noteOn/noteOff/step/loopStart via event bus.
- projection.js: composable transform chain with 5 transforms
(VelocityCurve, GateThreshold, RangeMap, OctaveFolder, StutterMap)
and 3 presets (expressive, percussive, fullRange). Pitch quantization
deliberately excluded (handled by Interval Lock primitive).
- step-viz.js: Canvas2D circular step visualizer with even angular
spacing for any step count. Pitch→radius, velocity→node size,
accent→color. 60fps-friendly with pre-allocated coordinate buffers.
Tap interaction for step toggling.
Five independent modules with no interdependencies:
- event-bus.js: namespaced pub/sub (seq.*/ml.*/ui.*) with wildcard
subscriptions and automatic AudioContext/performance.now timestamps
- prng.js: seedable mulberry32 PRNG with pure-functional API, fork()
for independent per-primitive streams, serializable state for freeze
- pattern.js: symbolic pattern description data structure with
create/clone/merge/validate, additive and multiplicative merge modes
- param-map.js: maps fixed 16-output MLP to variable-count primitive
params via linear interpolation, with optional per-param min/max scaling
- seq-iml.js: factory for second WasmIML instance (2 inputs, 16 outputs,
[16,16,16] hidden layers) following existing imlJoy/imlHand pattern
All modules are ES modules with no build step. PRNG, pattern, and
param-map are port-ready (explicit state, typed arrays, no closures).
- Change drawer hover trigger from .synth-quick-controls:hover to
.play-btn:hover ~ so hovering preset selector no longer opens it
- Add .play-drawer:hover to keep drawer open when mouse moves to it
- Set align-self: flex-start on play-btn and preset-select to prevent
vertical layout shift when drawer appears
Hand tracking at 30fps was writing 126 params per frame to the C15 ring
buffer (capacity 512), flooding it and starving arpeggiator noteOn/noteOff
messages — causing stuck/dropped notes.
Three-part fix:
- Arpeggiator now runs in a dedicated Web Worker with direct
SharedArrayBuffer access, bypassing the main thread entirely for
note timing. Worker setInterval isn't subject to main thread jank.
- RingBufferWriter uses CAS (Atomics.compareExchange) for multi-producer
safety — both main thread (params) and worker (notes) write safely.
- routeOutputs throttles synth parameter sends: dead-zone filter
(0.2% change threshold) + rate cap (~20fps), reducing ring buffer
pressure from ~3800 msg/s to ~50-100 msg/s of actual changes.
Hand tracking via webcam as alternative to joystick input:
- Right hand tracks 14 derived features (palm XY, finger curls, spread,
roll, pitch, pinch) through a separate 14-input MLP
- Left hand gesture recognition (1 finger = thumbs up, 2 = thumbs down)
with 400ms debounce hold
- Split-zone PIP display (no camera feed, skeleton only) with dashed
divider and cross-zone dimming
- Dual IML architecture: independent imlJoy (2 inputs) and imlHand
(14 inputs) with pointer swap, preserving training data per mode
- Dev panel (?devmode=true): draggable/collapsible floating panel with
sliders for MediaPipe confidence thresholds, smoothing, gesture hold
time, world landmarks toggle, and live feature bar monitor
- Visual presets always route to joystick IML (prevents dimension mismatch)
- Race condition guard on input mode switching
Compile nisps-core C++ MLP to WASM (36KB) and use it as the ML engine
in the playground, replacing the JavaScript port for inference, training,
and weight manipulation.
- Add extern "C" WASM bindings with spread-aware drawWeights/moveWeights
- WasmIML class is a drop-in replacement for the JS IML
- Inference runs on main thread via WASM (fast, synchronous)
- Training runs in a Web Worker with its own WASM instance (non-blocking)
- Interactive training (thumbs-up, train button) no longer freezes UI/audio
- Preset loading and state restore still use sync training
4 tiers of progressive complexity (Beginner 15 params → Expert 126),
13 presets total with per-param min/max/curve overrides that bias
distributions without clamping extremes. Preset dropdown in UI,
persisted to localStorage, supports ?preset= URL param.
SVF count was 7 (actual 9), Gap Filter section was missing (6 params),
FB Mix count was 10 (actual 9) — causing cascading mislabeling of all
synth sections from SVF onwards. Also sync volume/tempo changes from
bottom sheet back to quick-play controls.
- Glass-style modal explains NISPS, how learning works, and what to expect
- Documents all controls: touch, keyboard, and gamepad bindings
- Shows automatically on first visit (localStorage flag)
- Dismissible via close button, "Got it", overlay click, or Escape
- Reopenable via ? button at bottom right
- Map A=Train, B=Clear Examples, X=Randomize, LB=Thumbs Down, RB=Thumbs Up
- Add periodic gamepad polling fallback for environments where
gamepadconnected event doesn't fire reliably (Steam Deck, some Linux browsers)
- Show "Press any gamepad button to connect" hint on page load
When synth controls (volume, tempo, etc.) have focus after dragging,
keyboard shortcuts 1/2 for thumbs down/up were intercepted by the
range input. Now all four UIs skip shortcut handling when an input,
select, or textarea element has focus.
Relocated the output mode toggle from a floating top-right position into
the bottom floating bar where it's more discoverable. Removed the
duplicate toggle from the expanded sheet. Compact pill-toggle-sm style
fits the toolbar layout.
Merge Examples/RL tabs into single unified toolbar across all four UIs.
Users can now freely mix supervised learning (Add Example + Train) with
RL feedback (thumbs up/down) without switching modes. Param bar dragging
is always enabled.
Add Web MIDI input module (js/synth/midi-input.js) for external MIDI
controllers — routes note on/off to C15 synth, CC 1/2 to joystick.
Add gamepad module (js/ui/gamepad.js) with auto-detection, deadzone,
and axis normalization across all UIs.
Fix synth parameter sync: routeOutputs() now called after trainModel()
in onThumbsUp and loadState to prevent stale params on first joystick
move. Add separate Clear Examples button (dataset only, keeps weights).
- Fix RL buttons stuck high on desktop: media query was overriding
bottom to 196px, now matches mobile 92px
- Add floating Visual/Synth toggle at top-right of immersive view,
synced with the existing sheet toggle
Add ?spread=0-1 URL param that controls weight initialization scaling,
RL noise scaling per layer, noise cap, and weight decay to prevent
sigmoid output saturation. At spread=0 (original behavior) weights are
uniform [-1,1] and outputs polarise near 0/1. At spread=1 weights use
Xavier scaling (1/sqrt(fan_in)), noise is proportionally reduced, and
10% weight decay per thumbs-down prevents unbounded magnitude drift.
Also fix randomise to re-inject current joystick position and re-run
inference before routing outputs, eliminating the jump on first
joystick move after randomise.
Defaults: tame=1, spread=0.6 across all app variants.
