State persistence, prior transfer, and warmup lifecycle. Read when working on checkpoint/, adding new checkpoint fields, debugging cold starts or stale priors, or understanding serde(default) requirements and backward compatibility rules.
Documents auto_derive.rs first-principles parameter derivation from capital and exchange metadata. Use when onboarding new assets, debugging parameter mismatches, understanding why gamma/max_position/target_liquidity have their values, or adding new derived parameters.
WebSocket management, event loop, rate limiting, reconnection, recovery, metrics, and order execution infrastructure. Use when working on orchestrator/, infra/, messages/, core/, fills/, or execution/ modules, debugging connectivity or order placement, adding message handlers, or investigating stale data and latency issues.
Documents the 9 learning feedback loops, SpreadBandit Thompson Sampling, adaptive ensemble, confidence tracking, and baseline tracker. Use when debugging learning behavior, tuning reward attribution, investigating model weight decay, or understanding how fills translate into parameter updates.
Layered risk system with monitors, circuit breakers, kill switch, and position guards. Use when working on risk/, safety/, or monitoring/ modules, debugging position limits, emergency shutdowns, spread widening, or adding new risk monitors. Covers RiskMonitor trait, severity escalation, and defense-first architecture.
Documents the additive spread composition pipeline from GLFT optimal through to final bid/ask prices. Use when debugging wide spreads, investigating spread component contributions, tuning defensive behavior, or understanding why quotes are wider than expected. Critical for incident triage.
Layer 3 optimal sequential decision-making with Bayesian belief tracking, HJB value functions, and changepoint detection. Use when working on control/, stochastic/, or process_models/ modules, debugging quote/wait/pull decisions, modifying the HJB solver, or adding action types. Covers conjugate updates, BOCD, and value of information.
Systematic analysis of model predictions vs realized outcomes. Use when computing Brier Score, Information Ratio, calibration curves, PnL attribution, or conditional calibration by regime/volatility/funding. Identifies which models are adding noise vs value.