a60e7b092f
119 commits, 89 files changed (+24,734 / -1,415).
Two coordinated workstreams:
1. Plan C Phase 2 (#233): direction-branch Thompson sampling resumed
to resolve the ff00af68a UCB-asymmetry regression. Selector now
matches distributional Q-target (Thompson at training, argmax at
eval), restoring the symmetry pearl. Exit gate train_active_frac >
0.40 met by smoke-test-gmvf8 (F0=0.75, all 3 folds Succeeded).
2. SP4 Layer A + Layer B + close-out: ISV-driven adaptive bounds for
the magnitude branch replace SP1-era 1e6×ISV multipliers and
hard-coded clamps. 42 ISV slots wired through 5 producer kernels
with sentinel-detect → Pearl-A first-observation + Pearl-D
Wiener-optimal meta-rate adaptation. Pearl B fused per-param-group
oracle covers all 8 param groups including Curiosity sub-buffers.
Pearl C engagement counter integrated into 5 Adam kernels with
rate-deficit gating. StateResetRegistry extended for fold-boundary
correctness.
Subsequent close-out:
- GPU-only Pearls A+D refactor (eliminated host-side compute path
that broke CUDA Graph capture).
- Layer B per-group Adam split (resolved Pearl C engagement deferral).
- #260 SP1-era 1e6 × ISV multiplier elimination via dedicated
bw_d_h_s2_p99 + q_dir_grad_p99 producers.
- feedback_no_cpu_compute_strict sweep: 5 GPU EMA migrations
(grad_norm fast/slow + iqn_loss + utilization + adaptive_clip +
calibrate_homeostatic_targets) + 5 documented exceptions for
host-aggregated PnL/eval-result inputs.
- Intent-side magnitude distribution diagnostic (#212).
Validation:
- 4 L40S smokes Succeeded: smoke-test-tkkx6, 5nrz7, g2mvm, gmvf8
- Local RTX 3050 Ti: 16/16 SP4 GPU producer tests + 11/11 SP4 lib tests
- cargo check --workspace clean (no new warnings)
Zero open deferrals on SP4-touched paths. Zero hardcoded magnitude
multipliers in cuda_pipeline consumer code. Zero host-side EMA/
reduction violations on production hot paths.
Memory pearls landed: pearl_first_observation_bootstrap (Pearl A),
pearl_fused_per_group_statistics_oracle (Pearl B),
pearl_engagement_rate_self_correction (Pearl C),
pearl_wiener_optimal_adaptive_alpha (Pearl D),
pearl_l1_lambda_grad_direction_entropy_deficit (signal-driven L1),
pearl_sp4_close_out (canonical reference for future migrations),
feedback_no_cpu_compute_strict (hard rule against host-side compute).
Closes #84, #175, #212, #217, #233, #240.
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.