Four stages, each gated on the last.

1. 1

   Stage 1: Research

   Research

   Define the problem space, data schemas, and methodology. Prototype the agent harness and policy engine.

   Focus

   • Data schemas and provenance model
   • Agent harness prototype and MCP server interfaces
   • Written research methodology
   • Initial research notes

   Exit criteria

   A documented methodology and reproducible demonstration experiments, reviewed for look-ahead and leakage risks.

2. 2

   Stage 2: Backtesting

   Planned

   Build point-in-time datasets and a simulation harness with explicit cost models and validation.

   Focus

   • Point-in-time datasets
   • Transaction-cost and liquidity models
   • Experiment registry
   • Out-of-sample validation procedures

   Exit criteria

   Reproducible simulations whose assumptions and limitations are published with every result, always labeled as historical simulation.

3. 3

   Stage 3: Paper trading

   Planned

   Forward-test research outputs against live data feeds with no real capital.

   Focus

   • Sandbox order-simulation endpoint
   • Forward-test monitoring and reporting
   • Operational runbooks and alerting

   Exit criteria

   Stable operation, with divergences from simulation analyzed and documented.

4. 4

   Stage 4: Controlled deployment

   Gated

   Possible limited, supervised use under strict policy limits, only after independent review.

   Focus

   • Independent technical and risk review
   • Legal and regulatory assessment
   • Policy limits and kill-switch verification

   Exit criteria

   Not scheduled. This stage requires the reviews above, and no outside capital is solicited or accepted.

This roadmap describes intended sequencing. Plans can change, stages can be skipped or stopped, and nothing here is a commitment or a promise of any outcome.