Trading-RIL — Research Intelligence Layer

Open source peer-reviewed signal validation for AI trading bots.

Before your trading bot places any order, Trading-RIL runs a peer review of the signal using three independent research engines. Like an academic paper reviewed before publication — except it runs in 45 seconds and costs ~$0.03.

What It Does

Every trading signal gets reviewed by three layers before execution:

Layer	Engine	What It Adds
Layer 0	Bull/Bear Debate (TradingAgents inspired)	Structured debate between optimist and pessimist agents
Layer 1	OASIS Crowd Simulation	50 trader agents argue for 5 rounds → dissent_rate
Layer 2	MiroFish Swarm Intelligence	Swarm sentiment direction + confidence score

Output: gate0_pass (True/False) + confidence_modifier (-0.05 to +0.05) + research_brief (human-readable)

The Key Signal Nobody Else Has: dissent_rate

dissent_rate = 0.18  → Strong consensus → lower threshold (more trades)
dissent_rate = 0.43  → Market confused → raise threshold (sit tight)

High dissent means the market is genuinely split. Traditional sentiment scoring misses this entirely.

Quick Start

pip install trading-ril

from trading_ril import RILGate0

# Layer 0 only (Bull/Bear debate). No external state required.
gate = RILGate0(anthropic_api_key="sk-ant-...")

result = gate.evaluate(
    instrument="MGC",           # Micro Gold Futures
    strategy="MACRO_GOLD",
    signal_confidence=0.93,
    news_context="Fed holds rates. Gold edges higher on dollar weakness.",
)

print(result["gate0_pass"])           # True
print(result["confidence_modifier"])  # -0.010 (research agrees - ease threshold)
print(result["research_brief"])       # "MGC MACRO_GOLD | Research:Overweight | Crowd:n/a | Swarm:n/a | ..."

Enabling all 3 layers (optional)

By default Trading-RIL runs Layer 0 only (Bull/Bear debate). Layers 1 (OASIS crowd) and 2 (MiroFish swarm) require pre-computed daily signals stored in your own Firestore project — populated by trading-ril-morning once per day before market open.

To enable Layers 1 + 2, pass firestore_project:

gate = RILGate0(
    anthropic_api_key="sk-ant-...",
    firestore_project="your-gcp-project-id",  # reads tah2_oasis_signals + tah2_mirofish_signals
)

And populate your Firestore once per day (typically 8am market local time):

trading-ril-morning --instruments MGC MCL MES MNQ MHG MNG \
                    --firestore-project your-gcp-project-id

Cost note: the morning batch costs ~$0.08 (50 agents × 5 rounds for 6 instruments). Per-signal Layer 0 review is ~$0.03. Total: $0.38/day for 10 trades. Without the morning batch, you only pay the per-signal cost ($0.03/trade).

As a Claude Skill / MCP Server

Add to Claude Code (stdio, runs locally):

claude mcp add trading-ril -- uvx trading-ril-server

Or connect to the hosted MCP server (HTTP):

claude mcp add --transport http trading-ril https://trading-ril-1073730545783.us-central1.run.app/mcp

The hosted server runs Layer 0 only (no Firestore — keeps it stateless and free for everyone). To get all 3 layers from the hosted server, pass your own firestore_project to the tool:

Review this trade with my Firestore: MGC, MACRO_GOLD, confidence 0.93,
firestore_project=my-gcp-project-id, news: "Fed holds rates"

(Your GCP project's Firestore needs tah2_oasis_signals and tah2_mirofish_signals collections — populated by trading-ril-morning. The hosted server's service account does not have read access to your Firestore — it uses your project's own credentials via the MCP tool argument.)

Then in Claude Code, ask:

Review this gold trade: MGC, MACRO_GOLD strategy, confidence 0.93,
news: "Fed holds rates steady"

Supported Instruments

Symbol	Name	Typical Drivers
MGC	Micro Gold	Fed, DXY, inflation, ETF flows
MCL	Micro Crude Oil	OPEC, EIA, demand
MES	Micro S&P 500	Earnings, Fed, VIX
MNQ	Micro Nasdaq	Tech earnings, rates
MHG	Micro Copper	China PMI, EV demand
MNG	Micro Natural Gas	Weather, storage
Custom	Any instrument	You define the context

Morning Batch Mode

Run once before market open. Populates all six instruments:

trading-ril-morning --instruments MGC MCL MES MNQ MHG MNG

Outputs dissent_rate + direction + confidence per instrument. Gate 0 reads these instantly — no simulation overhead per trade.

Research Foundation

Trading-RIL synthesizes five academic frameworks:

Paper	Institution	Contribution
TradingAgents	Tauric Research	Bull/Bear debate framework
FinMem	Stanford/IEEE	3-layer memory architecture
FinAgent	Multimodal AI	Dual-level reflection
OASIS	Oxford/KAUST/Imperial	Social simulation engine
ContestTrade	FinStep-AI	Competitive agent ranking

Cost

Component	Daily Cost
Morning batch (6 instruments)	~$0.08
Per-signal review (10/day)	~$0.30
Total	~$0.38/day ≈ $11.50/month

Built entirely on open source. MIT licensed. No per-seat fees. No cloud lock-in.

Who Built This

Trading-RIL is the open source release of the research layer built for TAH-2 — a personal commodity futures trading system on IBKR PRO by Vinay Sharma.

TAH-2 paper trades on GLOBEX micro futures (MGC, MCL, MES, MNQ, MHG, MNG). Every signal goes through this peer review before execution.

Contributing

PRs welcome. Areas that need help:

• Additional instrument contexts (crypto, equity, FX)
• More agent personas for OASIS simulation
• Integration examples (Alpaca, Interactive Brokers, Binance)
• Backtesting framework

License

MIT. Do whatever you want with it. If it makes you money, consider contributing back.