Monolith Research
The instinct on a new strategy is to start with the signal. The signal is the interesting part, the part with the largest visible effect on the equity curve, and the part most people want to talk about. The recommendation kept on file here is to start with the wrapper instead; the layer that has nothing to learn.
A deterministic risk shell enforces a hard cap on position size per instrument, a gross exposure ceiling across all instruments, a daily loss limit, a trailing drawdown cap measured from the equity peak, an order-rate limiter, a regime-based size reduction when volatility is unusually high, a broker heartbeat watchdog, and a manual flat command. It does not improve the signal. It improves what happens when the signal is wrong, which is what the live equity curve actually measures.
The order matters more than people usually allow. Building the shell after the signal has been calibrated is the typical mistake; by then the strategy has been fitted to a model of itself that does not include the shell, and the shell arrives later as a brake instead of as a constraint. Building the shell first turns the strategy design into a constrained optimisation problem from the outset, which is closer to how the eventual live system actually behaves. It also pre-commits the work to engineering discipline rather than to model novelty, which has the side effect of slowing the work down in the places it usually goes wrong.
This is recorded as a working principle, not a recommendation. The shell is in front. The promotion pipeline is in front of paper. Paper is in front of shadow. Shadow is in front of size. None of those layers needs a new model to be the first one built.