7.7 A combinational experiment protocol

An experiment should compare at least four conditions: isolated ingredients, naive juxtaposition, a supplied interface, and a proposed interface. The last condition tests whether the system can discover the bridge rather than merely use one supplied by the evaluator.

Experiment: Typed interface discovery. Epistemic status: proposed benchmark design, not a completed result.
Input: two registered theories or modules, partial observations, and an interface meta-language in which the benchmark’s target bridge is expressible. A separate transformational control may deliberately omit a required interface type.
Proposal: an interface object and maps, the resulting composite, and a predicted interaction; any extension of the registered meta-language must be tagged as a transformational proposal.
Probes: isolated and paired Weil profiles, order-sensitive tests when meaningful, and counter-witnesses targeted at the proposed interface.
Admission: type correctness, preservation of declared invariants, held-out interaction, transfer to a second instance, and calibrated abstention.
Primary ablation: remove the interface description and ask whether surface association alone yields the same prediction.

The evaluation should report more than success rate. It should measure the rank of the correct interface among proposals, the number of probes required, false combinations admitted under confounding, and whether the composite is reused in a later task. Reuse is especially important: it distinguishes a one-off association from a new compositional resource.