All work

Research tooling / Case study

ARGUS AI

Challenge a proposed trade before it enters the simulation.

Type
Open-source engineering
Release
v2.1.0
Stack
Python 3.12 · Pydantic · Pandas / NumPy · MCP Server
Automated tests
77

System architecture

  1. 7-Agent Investment Committee & Devil’s Advocate
  2. Leakage-Free Walk-Forward Backtester (argus_ai/backtest.py)
  3. Cryptographic Trade Passports & MCP Server
Three layers from the project’s architecture blueprint

01 / Problem

Why it exists

Single-agent algorithmic trading bots suffer from confirmation bias, lookahead leakage in backtests, and blind execution during volatility traps or flash crashes.

02 / Design

How it works

Grant the Devil’s Advocate agent unilateral veto authority at >=75% disqualification risk and enforce deterministic 1.5x ATR stops in the Risk Commander.

  1. 7-Agent Investment Committee & Devil’s Advocate

    Combines Quant, News, Macro, and Options thesis generation with an adversarial Devil’s Advocate agent empowered to downsize or veto proposals.

  2. Leakage-Free Walk-Forward Backtester (argus_ai/backtest.py)

    Enforces PointInTimeGuard, deducts spread/commission/ATR-slippage costs, and compares full committee governance against an unfiltered baseline.

  3. Cryptographic Trade Passports & MCP Server

    Serializes every committee deliberation into a SHA-256 signed JSON passport with deterministic CLI replay and MCP tool exposure.

Unilateral Devil’s Advocate Veto & Non-Bypassable ATR Risk Gate
Context
Weighted averaging across bullish agents allows strong momentum signals to override severe liquidity or macro trap warnings.
Decision
Grant the Devil’s Advocate agent unilateral veto authority at >=75% disqualification risk and enforce deterministic 1.5x ATR stops in the Risk Commander.
Trade-offs
Reduces total trade frequency in choppy markets, but substantially cuts maximum drawdown and avoids bull-trap reversals.
Verification
Verified across 77 pytest tests including tests/test_backtest_engine.py, tests/test_argus_committee.py, tests/test_http_server.py, and tests/test_mcp_and_http.py.
Security and operational constraints
  • Paper trading and deterministic simulation enabled by default; live execution requires explicit environment authorization.
  • Hard circuit breaker (kill_switch.py) halts execution when daily drawdown or consecutive loss thresholds are breached.
  • SHA-256 digest verification detects any post-hoc tampering with recorded Trade Passports.
  • Any bar with timestamp > decision_timestamp immediately raises LookaheadLeakageError.
  • A Devil’s Advocate disqualification score >= 75% unconditionally vetoes execution regardless of bullish thesis scores.
  • Simulation and paper-trading are the default operating modes; no fabricated live trading returns are claimed.

03 / Evidence

How to verify it

Repository-reported / Adversarial Veto Precision

3 Trap Trades Vetoed · 77 Tests

Walk-forward simulation where Devil’s Advocate vetoed 3 flash-crash/blow-off traps ($488.37 counterfactual loss avoided)

Reproduce from the command line
git clone https://github.com/codesphere-u/ARGUS-AI.git
cd ARGUS-AI && pytest -v
python -m argus_ai.cli backtest
View source on GitHub
Next case study · AI verification Nexus Aegis Attach verifiable evidence to an agent’s proposed change.