Autonomic Alpha Foundry: Agentic Quantitative Research Platform

[!NOTE]

Executive Overview

Attribute	Details
Status	🧪 Functional Prototype & Research Engine (Active on nur)
What It Is	Quantitative research platform where the research seat is occupied by an autonomous agent that forms hypotheses, acquires data, runs studies, and delivers trustworthy verdicts.
Core Problem & Solution	Quantitative strategies routinely overfit historical noise. Foundry enforces family-wide error control (Bonferroni / Holm corrections), measures selection cost via 252-day embargoes, and refuses lookahead-polluted evidence rather than scoring it badly.
Tech Stack	Python, Microsoft Agent Framework, FastAPI, PostgreSQL 16, Model Context Protocol (MCP), SvelteKit, WebSockets
Key Highlights	Family-wide error control across all historical tests, 7 specialized collaborating agents, AG-UI live approval gates.

Autonomic Alpha Foundry is a quantitative research platform where the research seat is occupied by an autonomous agent. Its primary product is trustworthy verdicts, not profitable ones — an honest negative result proves the statistical governance system is working.

1. Core Statistical Philosophy: Honest Verdicts & Error Control

Modern algorithmic trading platforms frequently fail because multiple testing generates false discoveries. Foundry incorporates rigorous mathematical safeguards directly into the agent's decision loop:

Refusing Evidence Rather Than Scoring Badly

If two trading strategies exhibit identical return curves and t-statistics, but one strategy reads a corporate fundamental figure that did not exist on the day of the trade, Foundry refuses the evidence outright. Scoring corrupted evidence "badly" would imply the system could be trusted with approvals; refusal is the only scientifically valid response.

Measuring the Cost of Selection

Foundry splits data into folds separated by a strict 252-day temporal embargo between selection and evaluation. In validation studies, signals with apparent in-sample t-statistics of +2.22 collapsed to -1.28 out-of-sample:

Mean in-sample t:       +0.94
Mean out-of-sample t:   +0.00
Cost of selection:      +0.94 t

The system openly surfaces how much of an apparent edge is merely the statistical cost of picking the winner.

Family-Wide Error Control (FWER)

As the autonomous agent registers hypotheses, the required statistical threshold rises automatically. While a single hypothesis requires a t-statistic of $1.96$ ($p < 0.05$), registering multiple candidate studies raises the hurdle to $2.236$ and beyond under Holm-Bonferroni correction.

2. Multi-Agent Orchestration Architecture

Built on top of the Microsoft Agent Framework, the platform coordinates seven specialized agents over an Agent-to-Agent (A2A) protocol:

                  ┌─────────────────────────────────┐
                  │      Manager Agent (V3)         │
                  │ (Intent Routing & Conversation) │
                  └───────────────┬─────────────────┘
                                  │
         ┌────────────────────────┼────────────────────────┐
         ▼                        ▼                        ▼
┌──────────────────┐    ┌──────────────────┐    ┌──────────────────┐
│ Strategy Advisor │    │  Data Specialist │    │  ML Specialist   │
│ (Pipeline Coord) │    │ (Wide CSV / TTL) │    │ (Training Runs)  │
└──────────────────┘    └──────────────────┘    └──────────────────┘
         │                        │                        │
         ▼                        ▼                        ▼
┌──────────────────┐    ┌──────────────────┐    ┌──────────────────┐
│   Optimization   │    │    Evaluation    │    │    Inference     │
│ (Hyperparameter) │    │(Sharpe/Drawdown) │    │(Live Prediction) │
└──────────────────┘    └──────────────────┘    └──────────────────┘

3. Tool Integration via Model Context Protocol (MCP)

Operations are cleanly decoupled through MCP servers:

• Data fetchers, execution simulators, and model registries operate as standard MCP tools.
• Replacing synthetic data feeds with live institutional market feeds requires swapping an MCP server with zero changes to the multi-agent reasoning logic.

4. AG-UI Real-Time Streaming & Approval Envelopes

The frontend connects via WebSockets using a custom AG-UI protocol:

• Structured Envelopes: Every message adheres to an immutable JSON envelope schema (event_type, correlation_id, agent_name, payload).
• Human-in-the-Loop Approval Gates: When an agent attempts an expensive compute action (e.g. 50-trial hyperparameter optimization), the UI dynamically summons an interactive approval gate widget into the chat stream.

5. Persistence & Relational Schema

State is maintained in PostgreSQL 16 across 17 application-owned tables:

• training_proposals: High-level architectural specifications.
• models: Central registry tracking statuses (staging, deployed, archived).
• backtests & trade_logs: Equity curves stored as indexed JSONB.
• training_metrics: Time-series loss and validation telemetry.

6. Technical Stack Summary