jev-trader

A 24/7 market-making trading system built around Jev, TypeSafe AI's System One decision model. Jev returns calibrated, typed decisions in ~70–500 ms — inside a 300 ms block budget. This repo is the system that calls it: state engine, judgment battery, policy engine, A-S pricing, hard risk vetoes, execution, and the calibration loop.

One principle carries the whole build: Jev should not own the trading system. It should own selected judgments inside it. Code calculates the state. Jev interprets the state. Code applies policy. Execution places the order.

Architecture

MARKET DATA
     |
FEATURE ENGINE                      deterministic, pure code
     |
STATE SNAPSHOT  (< 400 tokens)
     |
JEV  — one call, six atomic judgments
  regime | direction | toxic_flow | liquidity_stressed
  quote_environment | inventory_pressure
     |
PROBABILITY VECTOR (typed answers + confidence)
     |
POLICY ENGINE                       thresholds live in code
     |
  HARD RISK RULES  ------------->   absolute veto, never delegated
     |
EXECUTION (post-only quotes)
     |
CALIBRATION LOG (state, decision, outcome triples)

Everything computable stays in code (mid, microprice, spread, imbalance, realized vol, inventory, drawdown, queue position). Everything that is a judgment goes to Jev. Atomic questions, composed in code — change a coefficient, not a prompt.

Quickstart

uv sync

# Run the paper loop in fallback mode (no API key needed)
uv run jev-trader paper --blocks 2000

# With live Jev judgments
export TYPESAFE_API_KEY="..."
uv run jev-trader paper --blocks 10000

# Realtime pacing (300 ms blocks) instead of accelerated simulation
uv run jev-trader paper --realtime

# Calibration report from the logged triples
uv run jev-trader calibrate

# List available models
uv run jev-trader models

Paper runs are deterministic for a given --seed and write JSONL calibration triples to data/calibration.jsonl by default.

Layout

Module	Responsibility
state/	L2 book math, feature engine, dense snapshot builder
judgment/	Six-question battery, async Jev client, deterministic fallback judge
policy/	Action composition (thresholds in code), fractional Kelly sizing
pricing/	Avellaneda-Stoikov reservation price, half spread, inventory skew
risk/	Hard-coded limits, absolute veto, kill switch
execution/	VenueAdapter protocol, paper venue with post-only lifecycle + fills
loop/	Nine-stage loop, block-deadline rule, fallback ladder
calibration/	JSONL triple logging, Brier / log loss / ECE / reliability / Platt

Fallback ladder

The system is autonomous only if it degrades in order:

healthy + high confidence  -> normal operation
healthy + low confidence   -> reduce size
late past block deadline   -> hold, never quote stale state
Jev unavailable            -> deterministic fallback judgments only
hard limit breached        -> kill switch, flatten, stop

Configuration

All config is env-driven (TYPESAFE_API_KEY, JEV_*). See .env.example for the full set. Risk limits are hard-coded vetoes and are never delegated to the model.

Calibration

Every block logs the triple (state, decision, outcome). jev-trader calibrate reports Brier score, log loss, expected calibration error, and a reliability curve, and fits Platt scaling so policy thresholds can be recalibrated on your venue's data — RLCD calibrates against TypeSafe's distribution, not yours.

Roadmap

See PLAN.md. Current status: Phase 0 (foundation + paper loop) complete.

Disclaimer

This software is for research and educational purposes. It is not financial advice. Trading involves substantial risk of loss. Run in paper mode and testnet before risking capital, and never disable the risk engine.