ε-Sufficient Context for Bounded Readers: Definitions and a Non-Monotonicity Result for Code Editing

Context retrieval for a language model that edits code is evaluated almost entirely by end-task pass rate. There is no agreed definition of what context is sufficient, and the natural baselines, program slicing and the principle of total evidence, assume a reader for whom more information never hurts. Language models are not such readers: their accuracy falls as input grows. This note fixes definitions that take that into account. A change request is a contract (Δ, D, I) with an explicit frame condition and an oracle that approximates it. A context K is ε-sufficient for an editor A when A's success on K is within ε of its success on the best context in a stated candidate family, not on the full repository. Under a stylised coverage-times-degradation model we prove, and mechanise in Lean 4, that ε-sufficiency is not closed under superset, with a drop that grows with the amount of irrelevant material, while it is closed under superset for a lossless reader. We state the forward slice as a requirement on the oracle, give three design properties that would make the minimal context computable, and report a mutation-based measurement on a real Go repository of how often a body-only change is caught only by tests outside the changed package, together with the token cost of including them. Finally we describe a planned benchmark, SuffBench, whose instances carry observation-based oracle slices. No language-model experiments are reported.