Reuse the same Objective

Define (or wrap) your f once. Pass the identical object to pilots, mixer, GLE, surrogates, polish, and anneal drivers. This is the source of the “one change serves all” property.

Prefer QMC starts when bounds are known

low_discrepancy_points (with a stable skip) gives better coverage and reproducibility than random for the same budget. Use it for mixer chains, pilots, and polish.

GLE drift is reusable

Fit (or use the default optimal-sampling) drift once; the same matrix can be used at any temperature / with any driver that accepts a GLE Move. Correctness is controlled by the covariance C you pass to sample_stationary; acceleration by the drift A.

Surrogates as first-class citizens

An additive or Chebyshev surrogate can be treated as an Obj (cheap value) or as a Move (independence proposal). The Metropolis correction reports the best value against the true objective. Charge the pilot cost to the budget; the collapse never enters the accuracy accounting.

Python array API + DLPack

When using device backends (via anneal), keep everything in the namespace supplied by the bounds array. Pass DLPack tensors for zero-copy handoff to other runtimes.

Test with the laws checker (when using via anneal)

The SaVariant / checked_with_sweep in anneal will catch neighbourhood or acceptance-law violations early. Write your custom Obj/Neigh/Move/Accept so they satisfy L1-L4; the primitives here are already designed to do so.