Measurability Before Power: Pre-Confirmatory Viability Screening for Sparse LLM Experiments

Experiments comparing large language model (LLM) interventions on sparse software-engineering tasks can spend most of their budget before establishing that a comparison is measurable. We adapt established multi-endpoint feasibility methodology to sparse LLM author/reviewer experiments, operationalise three viability endpoints, and report a prospectively governed case in which the screen stopped a confirmatory study before its reserved evaluation pool was consumed. Two development pilots, each with 1,320 model calls and 840 judged outcomes, failed criteria fixed before model calls. Task-level outcomes moved on 4 of 40 and 0 of 40 tasks. A replication-aware movement statistic has a positive, success-rate-dependent reference null even under zero treatment effect; evaluated at each pilot's pooled baseline, observed responsiveness was 0.36 and 0.00 times the pooled-binomial independent-arm reference null. Reviewers used a structural NO ISSUES FOUND channel on 0.0–8.3% of drafts against a prospective 10% trigger, while the planning model was near-separated or failed to fit. The prospective stop rests on the frozen endpoints, not on the reference-null comparison or the post-hoc estimability diagnostic. External construct-alignment tests and assumption-dependent operating-characteristic simulations are supporting evidence only. The external analysis supplies no positive corroboration after a post-hoc, reviewer-motivated noise null. We contribute an LLM-specific operationalisation of established feasibility methodology and a documented case in which a screen stopped a study before its reserved evidence was spent; we claim no new statistical method, causal result about critique, or external validation.