Tape-strip: primal

The problem

A numeric routine sometimes needs a value that is a function of the model parameters but must not carry a gradient. The usual example is a quadrature window: the endpoint at which an infinite integration limit is clamped is a function of the component parameters, but it is a hyperparameter of the quadrature — just where to integrate — and it should not be differentiated. If live Dual or TrackedReal parameters flow into the code that picks the endpoint, the AD backend traces functions it cannot handle (for example gamma_inc_inv for a Gamma component under Enzyme) and the differentiation fails or returns NaN.

The fix

primal reduces an AD-wrapped scalar to its underlying value, and primal_distribution rebuilds a distribution from primal parameters. A quantity computed from primal parameters is a constant as far as AD is concerned, so it stays off the tape. primal_distribution has dedicated Truncated/Censored methods because their params does not round-trip through the positional constructor, and raises an ArgumentError naming the type for anything else that does not. The rebuild path carries no type bound, so a leaf that implements the univariate interface without subtyping UnivariateDistribution rebuilds too, provided its params round-trips.

using EpiAwareADTools, Distributions

primal(3.0), primal_distribution(Gamma(2.0, 1.5))

(3.0, Distributions.Gamma{Float64}(α=2.0, θ=1.5))

What primal accepts

A Real is stripped to its underlying value; a Tuple or AbstractArray is stripped elementwise into a new container of the same shape, recursing so a nesting of either bottoms out at a scalar. nothing passes through unchanged, so the absent bound of a one-sided truncated/censored distribution strips alongside its numeric siblings. That covers a distribution's nested parameter tuples and a vector-valued hyperparameter such as a grid of integration nodes. Anything else raises a MethodError naming primal, so extend it with a method for your own type rather than relying on a fallback.

Per-backend behaviour

The generic primal is the identity on a plain real, so a non-AD call keeps the value's own type. The extensions make it a genuine stop-gradient on each backend:

• EpiAwareADToolsForwardDiffExt and EpiAwareADToolsReverseDiffExt add methods that read the underlying value out of a Dual or TrackedReal.

• EpiAwareADToolsChainRulesCoreExt marks primal @non_differentiable.

• EpiAwareADToolsEnzymeExt marks it EnzymeRules.inactive.

• EpiAwareADToolsMooncakeExt gives it a @zero_derivative rule.

Upstream target

There is no single upstream owner for a cross-backend stop-gradient. This pair is deleted once a shared primitive exists that marks a value non-differentiable uniformly across ForwardDiff, ReverseDiff, Enzyme, and Mooncake. Until then hosting one sanctioned primal here is better than each consumer package carrying its own underscore-prefixed copy.