One system. 173 signals. Nothing else like it.

The decision engine runs all 173 data cards using the current learner-adjusted weights. Each card outputs a directional signal, strength, and confidence score. The zone engine then takes these outputs through a multi-step pipeline — clustering nearby levels, scoring directional confluence, and qualifying only zones that pass distance and separation filters. The final output — bias, probability scores, qualified zones, and per-category dominant signals — is published for tracking.

For each new signal, the system starts two independent tracks — one long, one short — each with its own entry zone derived from the zone engine. Reference levels for targets and stops are stored alongside, but they don't interfere with how price is tracked. The stop distance is dynamic: calibrated from the adverse excursion of past trades so it sits just beyond normal volatility noise.

Every price tick from the exchange is processed in real time. When price enters the entry zone, tracking begins. From that moment, the system continuously records how far price moves in the favorable direction (MFE) and how far it moves against you (MAE) — without any take-profit or stop-loss cutting the observation short. It also captures time-bucketed snapshots so the learner can see how setups develop over time. The result is a clean, unbiased measurement of signal quality.

Once a track completes, the system reviews it: did price trigger the entry? How far did it move in your favor versus against you? Could the target have been reached based on the actual favorable excursion? The outcome is marked as a win or loss based on whether price moved far enough — a direct measure of signal quality, not execution luck. These outcome records, tagged with the original signal's category scores, become the training data for the learner.

The weight recalibrator pulls all outcomes with category score data and measures predictive power: for each category, it compares the win rate when that category agreed with the trade direction versus when it disagreed. Categories where agreement correlates with winning get higher weights. The zone optimizer runs separately — it checks entry trigger rates and uses excursion distributions to derive optimal target and stop distances mathematically. If zones trigger too often, entry distances widen. If adverse excursion is consistently too high, stops get more room. A performance guardrail compares win rates before and after each adjustment and rolls back automatically if performance drops.

The engine reads the updated weights on its next 10-second cycle. Each learner category controls a group of related signal cards — for example, the order flow category influences all order flow, microstructure, and order book cards simultaneously. The learned weight scales every card in that group, clamped to prevent any single category from dominating. The same 173 cards fire, but with learned emphasis. The loop closes — and the next signal already carries the system's latest understanding of what works.

Default weights shown. The learner shifts these based on predictive power — categories that predict winners get amplified up to 3x. Weak categories get suppressed to 0.3x.