A frozen, paper-only decision system. Nine experts. Two statistical heads. Five gates that almost always say no.
This laboratory shows a real AI decision architecture in operation: raw market perception, expert estimates, probability fusion, priced uncertainty, and an expected-value policy that abstains unless the evidence clears the cost of acting. Nothing here is a mock-up — and nothing here has proven profitable trading. That is exactly why it is worth studying.
LIVE EXECUTION OFF · LIVE CAPITAL OFF · PAPER SHADOW ONLY · NO OPERATIONAL CONTROLS
The system below runs for real on a private server: it perceives one market (BTCTMN), predicts two horizons (30 and 60 minutes), prices its own uncertainty, and writes every decision to immutable ledgers. It has authority over nothing — every order path is structurally disabled. What remains is the interesting part: disciplined decision-making under uncertainty, measured honestly.
Most prediction systems answer “what will happen?”. A decision system must answer “is acting justified, after costs, given how uncertain I am?”. The gap between those two questions is where naive systems destroy value — trading on noise that looked like signal.
Different questions need different estimators: micro-direction, return surface, magnitude, large-move probability, path shape, execution cost, global dislocation, order-book events, maker-fill risk. A router weighs them per market state — and the lab shows the measured truth that several are highly correlated today.
Two frozen multi-output model heads — one for direction (huber MultiOutput), one for maximum-favorable-excitation (MFE) — turn 25 standardized features into quantiles and probabilities. Fusion assembles P(up), P(move > break-even), and expected magnitude bps.
A separate knowledge system (not the trading loop) turns matured outcomes into error attribution, episodes, and ranked hypotheses. 329 open hypotheses; promotion is owner-only and evidence-gated. The trading loop never reads the Brain at runtime — knowledge reaches policy only through the human owner.
Pre-registered experiments: spec frozen before evaluation, mechanical verdict gates (log-loss, Brier, ECE, reliability), then hash-chained forward shadow tracking. Rejections are kept as first-class results.
Five separate gates — EV margin, uncertainty cap, quiet-market veto, liquidity-stress veto, magnitude-clears-cost. A BUY or SELL requires all five to pass simultaneously. Today, the uncertainty gate binds essentially every minute: the machine abstains by arithmetic, not by accident.
The laboratory is a read-only projection of a system whose execution paths are structurally disabled. The constraints below are not policies that could be relaxed — they are properties of how the thing is built.
Guided Demo is deterministic and always available. Live Laboratory adds real freshness stamps and component status from the same snapshot — when the private system is unreachable, nothing breaks; the timestamps simply age, visibly.
Previous prediction results did not establish reliable net trading profitability. The demonstration this lab makes is intelligent decision-making with disciplined evidence and abstention — a machine that knows when it does not know.
All values come from one sanitized snapshot artifact, built read-only from the private system’s immutable ledgers and committed with this page. Internal identifiers are renamed; internal research texts are withheld; a machine check refuses to publish the artifact if any internal path, host, or credential-shaped string survives.