On 14 August we moved the account to the configuration we had spent the summer arriving at: one entry level, no averaging down, a flat target. The averaging-down pyramid was not switched off behind a flag. It was deleted as code, which is the only removal we now trust.
Deleting it exposed three defects. None of them would have raised an error, and two disabled the part of the strategy that carries the largest measured term in its result.
The circuit breaker cuts a position that is stuck in a confirmed uptrend. "Stuck" was derived from a related test in the sizing logic, which was a reasonable shortcut while positions had several levels.
With one level, that derivation collapses algebraically to a single comparison: price below entry. For a short position, price below entry means the trade is profitable. The flag was reporting stuck at exactly the moments the position was working.
Replayed over nine years of ETH, the consequence is not subtle. Forced exits fall from 34 to 1, and the longest stretch a position stays stuck goes from 34 days to 2,756. The strategy does not degrade; it stops existing. In the payoff decomposition the breaker is worth ×5.29, against ×1.48 for the regime filter and ×1.17 for the bare shape of the trade, so this one comparison holds up the largest term in the result.
The breaker was also held back until the book had accumulated twelve levels, a guard written when deep pyramids were the failure mode worth worrying about.
A one-level book reaches twelve levels never. The condition was not wrong; it was unreachable, which reads the same way in the code and the opposite way in production. The breaker was locked off permanently, and the log had nothing to say about it.
The third was in the interface. Closed positions were filtered by outcome equal to "WIN", which was every closed position back when the only way out was the target. A forced exit writes a different outcome and therefore appeared nowhere: not in the history, not in the counts.
The single event we cannot afford to miss is the one the screen was built to omit.
Each was correct code, written against a structure that no longer exists. Removing a feature does not remove the assumptions that were built on top of it; they stay behind, still compiling, still passing their tests, now describing nothing.
We found these by reading every rule that mentioned levels, rather than by testing the new configuration and observing that it worked. It did work. It would have gone on working, without the mechanism that makes it worth running.