Skip to content

jev-trader

Paper market-making system where code builds the state and policy while Jev answers six atomic judgments per 300 ms block.

Projectprototype

Overview

This jev-trader is a 24/7 market-making system built around Jev, organised on one principle: Jev should not own the trading system, only selected judgments inside it. A deterministic feature engine computes mid, microprice, spread, imbalance, realised volatility, inventory, drawdown and queue position and builds a dense state snapshot under 400 tokens. One Jev call answers six atomic judgments: regime, direction, toxic flow, whether liquidity is stressed, quote environment and inventory pressure.

A policy engine with thresholds in code composes those answers into actions with fractional Kelly sizing, Avellaneda-Stoikov pricing sets the reservation price, half spread and inventory skew, and hard risk rules hold an absolute veto and a kill switch. Execution places post-only quotes through a venue adapter (a paper venue is included). Every state, decision and outcome triple is logged for a calibration report. Paper runs are deterministic per seed and can run accelerated or in real-time 300 ms pacing, with a deterministic fallback judge when no key is set.

What the decision model decides

TypeQuestionAnswers
ChoiceWhat is the current market regime?(inferred)—
ChoiceWhich direction is the market leaning?(inferred)—
Yes / noIs order flow toxic right now?(inferred)—
Yes / noIs liquidity stressed?(inferred)—

How it works

  1. Feature engine computes state from L2 data
  2. One Jev call answers six atomic judgments
  3. Policy composes actions; A-S pricing sets quotes; hard risk rules can veto
  4. Paper venue executes; triples logged for calibration

Strengths

  • Clear split of computation, judgment, policy and risk
  • Calibration loop from logged triples
  • Deterministic seeded paper runs and fallback judge

Caveats

  • Paper trading only in this repo
  • No reported performance results

Getting started

uv sync && uv run jev-trader paper --blocks 2000 (export TYPESAFE_API_KEY for live judgments).

At a glance

Type
Project
Status
prototype
Language
Python
Runtime
CLI
Has code
yes

Use cases and integrations

How curated lists describe it

Frequently asked questions

What does Jev decide?

Six judgments per block: regime, direction, toxic flow, stressed liquidity, quote environment and inventory pressure.

Can Jev override risk limits?

No. Hard risk rules have an absolute veto and a kill switch.

Does it need a key to run?

No; a deterministic fallback judge runs without one.

Related projects

  • Research wiring LAYA/Jev typed decisions into a quant stack with Claude fallback: good news reading, no market-beating edge.

    ProjectresearchPython
  • Reproducible test of viral Jev HFT claims: direction not calibrated, never under 100 ms, and no edge after trading fees.

    ServiceresearchPython
  • Local China A-share technical analysis workbench where Jev assesses eight indicator dimensions and returns Buy, Hold or Sell with evidence.

    ProjectbetaPython, TypeScript
  • Local stock decision service that sends market data and news to Jev or a local model and returns a policy-checked buy/sell/hold.

    ProjectbetaTypeScript
  • PyTorch factor-research stack with 213 factors and cost-aware backtests, plus a preview typed portfolio-decision adapter.

    ProjectbetaPython
  • Chinese open-source research and trading workbench whose JEV monitor watches futures contracts and judges entries and exits.

    ProjectbetaTypeScript