Two layers.
One decision per tick.

Produces individual alpha signals. One class per recipe: indicator, HMM, ML, factor. Each takes a window of data (L1 / L2 / L3) and emits a typed direction + confidence. No fusion happens here. No portfolio decisions. Just the read.

# Layer 1 — every signal implements this contract.
class AlphaSignal(Protocol):
    def name(self) -> str: ...
    def required_data(self) -> DataSpec: ...
    def emit(self, w: Window) -> SignalOut:
        # direction ∈ {-1, 0, +1}
        # confidence ∈ [0.0, 1.0]
        ...

Fuses the per-tick output of N signals into one bet per symbol. Five built-in methods. Confidence is a first-class output, not a derived quantity. The downstream Scoreboard sorts across symbols, surfacing the top conviction reads for portfolio construction.

// Layer 2 — combinator contract (Python & C++23 share the shape).
class Combinator(Protocol):
    method:    FusionMethod   # equal/conf/vote/max/min
    def fuse(self, ss: list[SignalOut]) -> FusedOut: ...
    def rank(self, all: dict[Symbol, FusedOut])
              -> Scoreboard: ...