An autonomous quant,
thinking in the open.
Futarchists forms its own hypotheses, backtests them against its own market data, and paper-trades the survivors — every tool call and every fill streamed live. No black box. Nothing hidden.
Three books, one loop
One controller drives every book. Each is scored by the same audited, deterministic gate — never by an LLM's say-so.
Autonomous research
A LangGraph controller proposes a hypothesis, runs a multi-agent swarm to research it, then scores a real backtest on first-party data. Survivors get promoted; the rest are logged as evidence.
PEAD equities
Post-earnings-announcement drift, sized to a vol target and traded long-only on paper. Validated on ~170M daily bars and 217k earnings events — then forward-papered toward a 250-trade go-live gate.
Prediction markets
Earnings-beat probabilities calibrated per ticker, edge measured against market price. It will not trade a market type until a fail-closed calibration gate passes — n≥100, ECE≤0.05, positive Brier skill.
The loop, in the open
The same six steps run on a schedule. You can watch each one land on the Cognition Tape.
Hypothesize
Form or advance a testable market thesis.
Research
A swarm screens, mines factors, and audits risk.
Backtest
A lookahead-safe run on our own price history.
Gate
Deterministic pass/fail — no LLM numbers trusted.
Record
Promote, monitor, or reject — durably, on the wall.
- Paper only. No real capital moves. Every broker path is fail-closed by construction.
- Real data, real gates. Backtests run on first-party prices; the gate math is the same audited code end-to-end.
- Backfill never counts as live. Historical validation is tracked separately and can never inflate the forward go-live record.
- Nothing trades until it earns it. Live execution needs a signed mandate, a kill-switch, and a passed calibration gate.
Pull up a chair.
The desk is live right now — forming hypotheses, booking paper trades, and showing its work in real time.