SignalQuarry

SignalQuarry turns a trading-strategy idea into an honest backtest on Alpaca data and then into an Alpaca paper forward test. It is built to be driven by your own coding agent: one CLI (sqy), one JSON envelope per command, a reason code for every outcome, and an AGENTS.md in every project.

The framework owns time, data, fills and the evidence record. You (or your agent) write one pure function, decide(ctx, params), and declare everything else in strategy.yaml. That split is what makes the results trustworthy:

• No look-ahead by structure. decide sees only completed bars, adjusted point-in-time, and no clock, files or network.
• Every trial counts. Each configuration evaluated on real data is appended to a hash-chained trial ledger; the deflated Sharpe ratio uses the count.
• A sealed holdout. sqy spec freeze seals the most recent months; they can be opened once per strategy family, after the walk-forward gates pass.
• Claims follow evidence. Every envelope carries evidence.claim_level (none → in_sample → walk_forward → holdout_passed → paper_forward).
• The same step in backtest and paper. The paper runner reuses the backtest's decide and sizing code; a parity test keeps them identical.

SignalQuarry has no live-trading path. Passing tests, backtests or paper results never authorize live trading.

Start with the quickstart.