d6d846f896
Follow-up to commit1961857c2(which unwired G6/G10 from the training loop based on V7-gem measurement showing their penalties were sub-noise). That commit left the kernels, Rust wrappers, buffers, and readback accessors in place as scaffolding — intended for "easy re-wire if future evidence changes". Post-unification multi-trial results (commit69211af40+ N=5 trials this session) confirm G6/G10 decisions were sound: - Best Sharpe median: 32.4 (session baseline: ~18) — 1.8× improvement - q_gap median: 2.6 (baseline: 0.05-0.4) — 6× better action separation - sharpe_ema mean: +16.7 (baseline: often negative) With the unification landed and dramatically improved training dynamics, the G6/G10 scaffolding is now clear dead weight. Delete with confidence per V7 methodology (measure, then commit to the decision). Removed: - branch_independence_penalty kernel definition (.cu, ~50 LOC) - temporal_consistency_penalty kernel definition (.cu, ~55 LOC) - branch_indep_kernel / temporal_consistency_kernel fields + loads - branch_indep_penalty_buf / temporal_per_sample_buf / temporal_penalty_buf allocations + field declarations + Self init - compute_branch_independence() / compute_temporal_consistency() fns (~70 LOC Rust) - branch_indep_loss_value() / temporal_loss_value() readback accessors - read_gem_losses() tuple → read_gem_g12_loss() f32 (simplified call site) Kept: - G12 predictive_coding_loss + predictive_coding_backward kernels (real, measurable 6× variance reduction — commitse72885e8b,69211af40) - predictive_loss_value() readback + HEALTH_DIAG `gems [g12_predictive=X]` - unified_env_step_core shared helper (commita3b6bc2f3) Net: −263 LOC across 4 files. No behavior change — the scaffolding had been unwired since commit 1961857c2; this just removes the now-unused definitions. Verified: cargo check + smoke test pass (Best Sharpe 35.73 single-trial post-cleanup, 24.7–38.3 across 5 multi-trial runs). Memory trail: see feedback_v7_gem_methodology.md — 3-step process (identify signal / check better-form / measure empirically) applied across G6/G10/G12 produced 3 different correct decisions, validating the process. Files touched: crates/ml/src/cuda_pipeline/experience_kernels.cu (−132 / +8) crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs (−120 / +11) crates/ml/src/trainers/dqn/fused_training.rs (−10 / +3) crates/ml/src/trainers/dqn/trainer/training_loop.rs (−1 / +1) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Foxhunt
Production HFT trading system in Rust.
Architecture
The workspace contains 32 crates organized as follows:
Core Libraries (16)
| Crate | Purpose |
|---|---|
trading_engine |
Order processing, FIX 4.4, IB TWS, SIMD, RDTSC timing |
risk |
VaR, Kelly, circuit breakers, kill switches, compliance |
risk-data |
Risk data types and shared structures |
trading-data |
Trading data types |
ml |
DQN Rainbow, PPO, TFT, Mamba2, ensemble inference |
ml-data |
ML data types and feature definitions |
data |
Market data ingestion and storage |
backtesting |
Replay engine, strategy tester |
adaptive-strategy |
Ensemble execution, microstructure analysis |
common |
Shared types, resilience, error handling |
storage |
S3 and local model storage |
model_loader |
Model serialization and loading |
market-data |
Market data feed handlers |
database |
PostgreSQL access layer (SQLx) |
config |
Configuration management |
tli |
CLI commands and tooling |
Services (8)
| Service | Purpose |
|---|---|
backtesting_service |
gRPC backtesting service |
broker_gateway_service |
FIX routing, broker connectivity |
trading_service |
Core trading operations |
ml_training_service |
Model training orchestration |
data_acquisition_service |
Market data acquisition |
trading_agent_service |
Autonomous trading agents |
api_gateway |
gRPC API gateway with auth |
web-gateway |
Axum REST + WebSocket gateway |
Frontend
web-dashboard/ -- React 19 + TypeScript + Vite + TradingView charts.
Building
# Check compilation (no PostgreSQL required)
SQLX_OFFLINE=true cargo check --workspace
# Run tests for a specific crate
SQLX_OFFLINE=true cargo test -p <crate> --lib
# Clippy
SQLX_OFFLINE=true cargo clippy --workspace
ML Models
Four production model architectures on Candle v0.9.1 with CUDA:
- DQN Rainbow -- Deep Q-Network with prioritized replay, dueling heads, noisy nets
- PPO -- Proximal Policy Optimization with GAE and LSTM policies
- TFT -- Temporal Fusion Transformer for multi-horizon forecasting
- Mamba2 -- State space model for sequence prediction
Each model has a standalone trainer and a UnifiedTrainable adapter for the hyperopt pipeline.
Infrastructure
- Git: Gitea at
git.fxhnt.ai(Tailscale-only), Scaleway DEV1-S - Observability: OpenTelemetry OTLP (env
OTEL_EXPORTER_OTLP_ENDPOINT) - Database: PostgreSQL with SQLx offline mode for CI
License
Proprietary. All rights reserved.