One score drives advisory, credit and recovery routing. Here it is as an explicit function of eight measured inputs: angle-encoded onto qubits, measured as Z and ZZ expectations, squashed to a probability, and then attributed feature-by-feature with Shapley values (global-consistent, additive) and a LIME local surrogate (fast, local, with a stated fidelity).
Simulation notice: the quantum stage is a classical emulation of a variational circuit (product-state angle encoding with ZZ couplings). No quantum hardware is used and no quantum speed-up is claimed. Shapley values are a Monte-Carlo permutation estimate; LIME reports its own local R² so you can see when the linear surrogate is not trustworthy.
orient() flips features where low is good (EC, THI), so every θ runs from "bad" at 0 to "good" at π. The ZZ terms are the entangling layer: they let moisture change what salinity means, and heat change what microbial activity means, instead of adding effects independently.
Efficiency: Shapley values must sum to f(x) − f(baseline). The gap above is pure Monte-Carlo noise — raise the permutation count to shrink it.
SHAP answers "how much of the distance from the baseline farm does this feature own?" — additive and consistent, but it costs 1920 model evaluations. LIME answers "if I nudge this farm slightly, which knob moves the score fastest?" — one small regression, valid only inside the σ neighbourhood.
A sign disagreement is a real signal, not a bug: it means the response is curved between the baseline and this point, so the linear surrogate cannot represent it. Widen σ or trust the Shapley value.