There is a temptation when building an AI trading system to optimise for volume early. More calls means more data, more data means faster iteration, faster iteration means a better system sooner. That logic is not wrong but it skips a step that matters a lot.
If the quality of individual calls is low, high volume compounds the problem rather than solving it. You get more data faster, but the data you are generating reflects a system that is making poor decisions at speed. The patterns you learn from that data are patterns of low-quality decision-making, not patterns of good trading. You end up optimising for the wrong thing.
March for APEX was focused on the quality side before we pushed on scale. The strategy loop structure got a significant overhaul, tightening the sequence of checks a candidate has to pass before a call is generated. The goal was to make the system more selective, not more active. A candidate that generates no call is a better outcome than a candidate that generates a bad one.
The candidate filtering work was specifically about what happens between signal detection and execution consideration. Raw signal is noisy. A token that looks interesting for thirty seconds and then falls apart is not a candidate worth acting on. We added evaluation depth at this stage: not just whether a call can be generated from the current state, but whether that state is coherent with the broader market context and whether it survives the internal risk filters that are designed to catch setups that look promising on a narrow view and fall apart on a wider one.
Risk gate behavior also got more explicit instrumentation. Previously the gates would reject a candidate without leaving much trace of which condition triggered the rejection. That made it hard to evaluate whether the gates were calibrated correctly. We added structured logging at each gate so the reasoning for a rejection is visible and reviewable. That observability is the precondition for improving the gates over time based on real outcomes rather than intuition.
APEX is still in beta. The quality bar is getting tighter each cycle.