Everything so far used one curve family. But seed decline sometimes genuinely isn't a symmetric probit — a resistant sub-population flattens the tail, a control ceiling caps the start. The lab's Model Workbench fits several published forms to the same counts and ranks them — and the ranking rule matters more than the menu.
AICc scores each model by fit minus a penalty for parameters (with a small-sample correction that matters at monitoring-data sizes). The discipline it enforces: a fancier curve must buy its extra parameters with genuinely better likelihood. When ΔAICc between two models is small, the honest verdict is a tie — and ties go to the simpler model, not the newer paper.
The workbench's most instructive behavior is a deliberate one: some model pairs are near-indistinguishable on realistic data, and the engine correctly refuses to pick a winner. Learning to read "the data cannot tell these apart" as a result — not a failure — is the concept.