Limen-Capital — Experimental SNN-HFT (Julia Brain + Rust Muscle)

Hybrid research stack: Julia neuromorphic strategy + Rust deterministic execution. Built for experimental spiking / liquid-state market research — not live capital.

This repo is a standalone application (your personal/org remote). It is not published under the Limen-Neural GitHub organization. It depends on Limen-Neural libraries when you want the shared packages; in-tree naut_core works without them.

Dependencies (git pins only)

Validated Limen-Neural packages are pulled by git + rev (Cargo / Julia [sources]). No sibling Limen-Neural/ checkout is required.

Package	Role
metabolic-ledger	Ghost wallet, ATP gates, JSONL (git pin)
LiquidCortex.jl	Preferred CUDA LSM (optional; else naut_core)
TemporalFocus (NeuroPulse.jl)	Optional relevance package
binary_wire (in-tree)	MarketPulse 120B / ReadoutPacket 88B until corpus-ipc publishes them

See docs/deps.md, docs/wire-protocol-v1.md, and docs/SECURITY.md.

Architecture

┌─────────────────────────────────────────────────────────┐
│  Julia Brain                                            │
│  MarketEncoder (28 spikes)                              │
│       ↓                                                 │
│  LiquidCortex EnsembleBrain (4 lobes × 65K)             │
│       ↓                                                 │
│  TemporalFocus / NERO relevance                         │
│       ↓                                                 │
│  16-element readout + 4 relevance scores                │
└──────────────────────┬──────────────────────────────────┘
                       │ 88-byte binary over ZMQ TCP (v1)
                       ↓
┌──────────────────────────────────────────────────────────┐
│  Rust Muscle                                             │
│  Confidence gates → fractional Kelly → metabolic-ledger  │
│  optional dydx feed → latency tracking → JSONL           │
└──────────────────────────────────────────────────────────┘

Directory structure

Limen-Capital/
├── brain/           # Julia brain (helm, reservoir, metrics, tests)
├── math/            # Hawkes / fractal / SDE / FastReflex (FeatureStream uses Hawkes)
├── execution/       # Rust muscle (wire, Kelly Decision, metabolic-ledger)
├── wire/            # NervousWire v1 docs + golden fixtures
├── docs/            # deps + wire protocol
├── scripts/         # local smoke (not CI)
├── strategy/        # JSON adapter (secondary)
└── proto/           # FlatBuffers (optional / unused v1)

Quick start

Prerequisites

• Julia 1.12+ (Pkg [sources] pins; CI uses 1.12), Rust stable (latest), network for first-time git deps
• Optional: NVIDIA GPU for full EnsembleBrain / research runs

1. Structure + unit tests

./test_integration.sh
cd execution && cargo test          # fetches metabolic-ledger (+ optional feature deps)
cd ../brain && julia --project=. -e 'using Pkg; Pkg.instantiate()'
julia --project=. test/runtests.jl
# or: ./scripts/smoke_wire_local.sh

2. Health check (full GPU brain)

cd brain/
julia --project=. test_brain.jl

3. Research experiment (metrics + artifacts)

cd brain/
julia --project=. research_helm.jl --duration 200 --seed 42 --horizon 5
# Artifacts: runs/<run_id>/{meta.json,metrics.json}
# Options: --cost-bps 1.0 --output-dir runs --record-spikes --backend naut_core

ExperimentRunner (experiment_runner.jl + metrics.jl) records walk-forward hit_rate, PnL proxy, and latency p50/p99.
Reservoir façade picks LiquidCortex when loadable (LIMEN_RESERVOIR / --backend).
MC fan-out stays off in research mode (LIMEN_MC_PATHS only for live helm).

4. Rust Decision path (gates + fractional Kelly + metabolic-ledger)

cd execution/
cargo run --release
# Binary ReadoutPacket on :5556 by default; LIMEN_WIRE=json for legacy IPC
# LIMEN_CONFIDENCE_THRESHOLD=0.15  (research default in main)

ExecutionEngine::process_signal: confidence gate → Neutral → 0.25× Kelly size (from ATP bankroll, b=0.05) → max-qty gate → ghost fill. Zero Kelly → RejectedNoEdge.

5. Julia live helm (binary wire v1)

cd brain/
julia --project=. spike_helm.jl

Signal Pipeline

Input: 120-byte MarketPulse (from Rust → Julia)

[0..8]    timestamp_ns (UInt64)
[8..64]   7 assets × (price f32, vol f32) = DNX,Quai,Qubic,Kaspa,XMR,Ocean,Verus
[64..80]  funding_rate, liquidation_vol, liquidity_delta, order_imbalance
[80..100] GPU temp/power/util, FPGA buffer load
[100..120] dydx OI delta, dydx funding rate, Qubic fields

Output: 88-byte readout (Julia → Rust)

[0..8]    tick_count (Int64)
[8..72]   16 × Float32 — aggregated lobe readout
[72..88]  4 × Float32 — NERO relevance [Scalper, Day, Swing, Macro]

Math Modules (FeatureStream)

Module	Purpose	Status
market_hawkes.jl	Self-excitation intensity λ(t)	Wired via feature_stream.jl → encoder δ + Scalper reflex
market_lsm.jl	256-neuron FastReflex	Optional (LIMEN_USE_REFLEX=1)
market_fractal.jl	Hurst / regime	Offline only (not on hot path)
market_sde.jl	GBM surprise z-score	Offline only (not on hot path)

See brain/feature_stream.jl.

Neural Dynamics

OU-SDE Membrane Model:

dV = ((V_rest - V) / τ_m + W_rec·S + W_in·u) dt + σ·dW

STDP Covariance Learning:

ΔW = η · trace_pre · trace_post

Global Inhibition ("Cortisol"):

• GPU temp > 75°C → raise spike threshold
• High funding rate → "market is tense" → fewer spikes
• Liquidation cascade → inhibit all lobes

License

MIT — see LICENSE.

Optional Limen-Neural dependencies are typically MIT/Apache-2.0 as well.

Repo hygiene

• Do not commit execution/target/, runs/, data/, .env, .mimocode/, or security-scans/.
• Golden wire fixtures live under wire/fixtures/ and are tracked.

Roadmap

• Binary NervousWire (120/88) + golden fixtures
• Reservoir façade + causal NERO + FeatureStream (Hawkes)
• ExperimentRunner metrics artifacts
• Decision path: fractional Kelly + metabolic-ledger
• Git rev pins for validated Limen-Neural deps (no sibling clones)
• Replace JSON serialization with FlatBuffers (proto/signal.fbs)
• Add WebSocket streaming for dydx (replace REST polling)
• Fix CUDA RNG (remove host-side random number generation)
• Implement stop-loss and risk management gates in Rust
• Backtest framework with historical data