How do you deploy an AI trading strategy?

On Pancake, deployment is one MCP tool call after a verified backtest: create_paper_deployment takes a strategy version and runs it forward against live Polymarket data with simulated fills. The deployment gets a public page backed by an append-only event ledger — the strategy's track record from day one.

Deployment on Pancake is deliberately gated by evidence: backtest first, then deploy. The deployable artifact is a strategy version — an immutable, content-hashed spec — so there is never ambiguity about what is actually running. If the agent revises the strategy, that is a new version with its own results, not a silent edit.

The mechanics: an agent calls create_paper_deployment with the strategy version. Pancake evaluates the version's entry and exit rules against live Polymarket data, simulates the fills, and appends every event — orders, fills, position changes — to a write-once ledger. The deployment page renders positions and P&L from that ledger, not from editable snapshots.

Lifecycle control is three states away: set_paper_deployment_state pauses, resumes, or stops the deployment; get_paper_deployments reads current status. Both are MCP tools, so the same agent that authored the strategy can supervise it across sessions.

What "deployed" means here, precisely: a hosted forward test. Fills are simulated at observed market prices; no real orders are placed. Live venue execution is a v2-roadmap capability, and the deployment's ledger-backed record is the evidence a strategy carries toward it.