An instance of our continual learning architecture recovered a hidden capital-gains tax schedule from nothing but the outcomes of its own simulated trades, without ever being told the tax existed.
Full methodology and outcomes: Learning a Hidden Cost Structure from Outcomes Alone.
Three actions were available: buy, sell, or hold. The only instruction was to prefer whichever action's outcomes had been larger. The instance ran in a simulated equities environment and was never told a tax existed. Over a 26-year run it developed a persistent reluctance to sell positions carrying large gains.
The behavior alone doesn't prove much, since a reluctance to sell winners could come from other things. So we read its stored values directly, rather than inferring them from what it did, and they matched the tax rate the environment had been charging all along, calibrated to within about 6%. Removing the tax removed the behavior. Varying the rate scaled it in proportion.
We picked a tax because it has a knowable right answer, which makes it possible to check. The underlying problem isn't specific to markets. Operational systems everywhere run on constants someone set once, like a wind farm's assumed output or a factory line's tolerance threshold. Nobody revisits these on a schedule, and the software has no way of noticing when they've stopped being true.
Our architecture is built to close exactly that gap. An instance can watch the same data a piece of decision software already uses, and report when reality has moved away from what the configuration assumes. This is evidence it can find a real, costly rule with nobody writing it down first.
We're already working with early partners on exactly this pattern. If it sounds familiar, reach out to us at [email protected].
