Debatte — Five Analysts, One Trade

Most trading agents ask a language model where the market will go. Over four sessions that has a coin-flip answer. Debatte asks a different one: five LLM analysts argue about what is happening — and none can place an order. SEPARATED VIEWS. Each analyst sees a deliberately different slice of the briefing: momentum never sees implied volatility, the vol agent never sees sector rotation. Without that, five calls to one model return five paraphrases of one analysis, and their agreement measures only the model's stability. With it, agreement between agents that saw different data is a real signal — and the aggregator pays 15% per additional agent. GUARDRAILS. Every figure an agent cites is verified against the briefing; a deviation above 1% discards the proposal and logs it. A critic scores each thesis and can veto. Revision is anonymised, so models answer arguments, not roles. Each proposal carries a machine-readable invalidation condition, so every exit traces back to its thesis. WHERE THE RETURN COMES FROM. Not from being right about direction. A Monte Carlo over 400,000 paths shows expected value tracks the variance risk premium and almost nothing else. It therefore trades only where implied exceeds realised volatility by two points, in defined-risk structures, taking profit at half. A risk guard in plain Python sizes every position and can reject any of them; a circuit breaker halts new trades at 6% drawdown. ALPACA. The official MCP server carries research — an agent choosing its own calls from a five-tool read-only allowlist. The official SDK carries the order path: multi-leg via OrderClass.MLEG at options level 3, limit-only with three repricing steps. 170 tests, unattended under systemd across the scored window. TRANSPARENCY. The dashboard shows the argument thread, per-agent attribution, and in full everything the system rejected, with reasons. Six assumptions from our own design paper did not survive measurement; each is documented where it was fixed.