Every run travels through the complete decision lifecycle — from raw data to executed trade to post-trade learning.
Most runs exit early. The system is designed to be cheap to scan and expensive to decide.
| Run Type | Stops At | Steps | Tokens/Run | Cost/Run | % of Runs |
|---|---|---|---|---|---|
| NO SIGNAL | Researcher Team | 6 / 16 | 5,180,000 | $18.90 | 93% |
| REJECTED | Fact-check | 7 / 16 | 5,465,000 | $20.17 | 3.5% |
| NO TRADE | Trader Agent | 9 / 16 | 5,815,000 | $21.82 | 2.7% |
| FULL RUN | Monitoring | 16 / 16 | 6,942,000 | $28.48 | 0.6% |
The compute cost of a 24/7 AI trading desk.
The Researcher must write both PUMP and DUMP theses. Deep Thinking actively argues against the trade. No one-sided decisions are possible.
The execution layer is pure code — idempotent, retried, logged. The LLM's role ends at approval. AI thinks, code executes.
Sonnet (12 agents) carries 91% of tokens. Opus (2 agents) handles only adversarial review and final approval — reasoning complexity determines the model.
93% of runs stop at the Researcher (NO SIGNAL) — $18.90 instead of $28.48. The system saves budget by deciding not to trade as cheaply as possible.
Backtest statistics, position sizing calculations, and execution run on numerical engines. The LLM reads summaries — never computes.
Every decision is digitally signed with a unique decision_id. Full token transparency: per agent, per run, per dollar.
Request a live console walkthrough and see every agent, every token, every decision in real time.
Request a Demo