pipeline. audio in → FFT feature extraction (rms, spectral-flux onset, sub/bass/mid/high bands, beat phase with a BPM lock) → a shared feature bus → whichever system is live → canvas. the frame budget is 16ms; anything slower gets cut.

systems. 50+ of them: boids, physarum, Chladni plates, reaction-diffusion, strange attractors, Fourier epicycles, oscilloscope, supershapes, double pendulum, domain-warped fields. each takes the same feature bus and the same five universal faders (speed · count · size · trail · chaos).

color. cosine palettes, rgb(t) = a + b·cos(2π(c·t + d)), tuned so t=0 sits near-black; visuals live on dark stages. chaos is a curve, not a gain: v ↦ v^(1/chaos) on intensity features, so high chaos stays graded instead of clipping to strobe.

prompt engine. you type "dark warehouse, strobing grid" and new visuals are generated from scratch: a system, its parameters, and a palette, validated and hot-swapped without dropping a frame. shipped as an Electron app.

regime. a 3-state Gaussian HMM on SPY log returns labels the market calm, normal, or turbulent, with a cross-sectional breadth override for concentrated selloffs the index hides. turbulent halves momentum and reversal, upweights low-vol, and shrinks size. risk only ever scales down.

signals. cross-sectional 12-1 momentum; short-horizon reversal gated on positive 12-1 momentum (a falling-knife veto from an N=5,284 event study); a low-vol tilt ranked on hourly-bar realized volatility; gradient-boosted ML alpha with meta-labeling at a bounded ensemble weight; and cointegration pairs: Engle-Granger with Benjamini-Hochberg FDR control, an Ornstein-Uhlenbeck spread ds = κ(μ−s)dt + σdW, s-score entries and exits.

the renormalization. every signal is z-scored, then rescaled against the regime it fires in before the ensemble votes. same signal, different regime, different meaning.

the gate. every order clears a pure, deterministic risk gate: drawdown kill-switch, daily-loss and CVaR breakers, beta, correlation and concentration caps. an LLM committee is held to the same evidence bar as any signal and is currently benched.

sizing & validation. HRP + Black-Litterman portfolio construction; stops scaled by HAR-RV / GJR-GARCH volatility forecasts; execution priced with an Almgren-Chriss cost model. every change ships through a preregistered paired walk-forward and a Deflated Sharpe Ratio on purged CV; DSR ≥ 0.6 or it doesn't ship.