Targets

A target is an interval on the raw model output — this one abstraction covers probability cutoffs, regression goals, and rating ladders.

Target.probability(op="<=", value=0.04)      # under the 4% PD cutoff (via logit)
Target.probability(range=(0.2, 0.8))         # inside a probability band
Target.raw(op=">=", value=1.5)               # raw margin / regression units
Target.raw(range=(-1.2, 0.5))

Probability targets require a SIGMOID-link model and are converted once via the logit; open endpoints (0/1) map to infinities.

Rating ladders

Target.bands solves one model compilation against several intervals — the "price of each grade":

ladder = exp.explain(x, target=Target.bands({
    "A": (0.00, 0.01),
    "B": (0.01, 0.03),
    "C": (0.03, 0.07),
}))
# {"A": Counterfactual | Infeasible, "B": ..., "C": ...}

The AIM is compiled once and only the score bounds are swapped per band, so an N-band ladder costs one compilation plus N solves.