Interfaces or Bodies? A Preregistered Test of ε-Sufficient Context for Language-Model Code Editing in Go (Registered Report, Stage 1 protocol)

Coding agents decide, many times a day, how much of a repository to show a language model, and the decision is made by heuristic or by end-task scores that confound localisation, context size, and model quality. A recent note defines ε-sufficient context: a context is sufficient for an editor when the editor's success on it is within ε of its success on the best context in a stated family, and proves that under a bounded reader this property is not closed under superset. We propose a preregistered, paired, multi-model study that operationalises the definition as a non-inferiority test. Holding localisation fixed at the gold write-set on at least 200 fail-to-pass instances mined from a fixed, ranked list of Go repositories after each model's training cutoff, we compare (H2) a compiler-computed context consisting of the write-set and the exported interfaces of its dependencies against the same package set with full bodies and against the transitive slice; (H1) we estimate each model's degradation curve by padding a fully covered base context to fixed absolute loads up to 950k tokens with three padding types, with a pre-specified equivalence test so that a flat curve is a reportable result; (H3) we measure, descriptively, whether showing the forward slice reduces breakage of passing tests; and (H4) we report tokens per arm. Coverage is defined operationally per arm by checking that the gold patch is expressible from the context. Three models: a dense cheap model at ten samples per cell on every arm, an open-weights model at temperature zero, and a frontier model at vendor-default sampling on a fixed subset. A 24-instance pilot on one repository (2,360 verified samples, three models, about 61, published with all data) shaped this design: it fixed the margins from measured per-instance variance, made a task-only screen mandatory after the mid-tier model solved 14 of 24 instances with no repository context, extended the load range after padding to 200k produced no consistent effect, and set the sampling depth after a quarter of the dense model's cells disagreed across samples. The instances, slices, contexts and verdicts are released as SuffBench v1.0. Stage 1 protocol of a Registered Report, published as a preregistration before any confirmatory model call. The preregistration file with dated amendments is attached.