C.29.3:10.4 - Changing model and means together

A fixed computation can be a useful constraint on the device design. A fixed device can instead suggest more suitable computational operations. Stepney (2019), §§4–5 proposes combining these directions in model-and-substrate co-design.

Kalita and colleagues (2026), §§1 and 3–4 develop the reverse direction through a bosonic-device example: capabilities of the physical system inform a computational model and language. Adapt this source contribution as the formulation return in :4.6. Developing such a language requires its own constructive repertoire; a return from realization identifies that work rather than completing it.

The analog example shows a small instance of the same design freedom. Keeping the averaging circuit and changing its encoding or decoding can supply a useful addition operation. The physical relation constrains which interpretation works.

Thermodynamic sampling and optical ML hardware extend this design choice to computations based on distributions or iterative physical evolution. Melanson et al. (2025) demonstrate sampling and matrix inversion on a small stochastic circuit. Kalinin et al. (2025) co-design an optical/electronic fixed-point computation with its learning model. Their different result and execution forms motivate the choices in :4.1–4.5. The detailed sampler, model-training and device-construction Methods supply the corresponding specialist work.