ENLIST: preregistered replication package for small-language-model deployment decisions

Abstract

Replication package for "Selection information is not free: a preregistered study of small-language-modeldeployment decisions" (submitted to Information and Software Technology). The study asks whether charging deployment-selection policies for the information they buy changes which policylooks best. In ten environments, every action of the four checkpoints that pass an 8 GiB memory gate was executed(zero-shot deployment, supervised fine-tuning, diagnostic fine-tuning and, in nine environments, retrieval few-shotprompting), with tail latency measured and every token metered. Policies are scored by action regret (the deployedcell against the best measured action) and by policy regret (adding what they bought and did not ship). Relative to earlier versions, this version adds:- three sensitivity studies registered before measurement: learning-rate sensitivity (108 SFT runs), a non-Qwen catalogue extension and two retriever variants, with decision-relevant cells re-measured on the latency card (preregistration sections 20-141 to 20-143);- a re-sweep of the fallback rule's threshold registered before computation (section 20-145);- two audits registered before computation (section 20-146): the registered rules re-run with selection-time inputs only, and every cell's longest served request checked against the load-tested context;- the manuscript, supplementary material and submission files as submitted to IST.Of the 21 predictions registered for these studies and audits, 19 were supported and 2 failed; both failures arereported in the paper. Contents: the measurement pipeline; the fact-derivation harness that regenerates every reported number from rawartefacts (212 facts; python3 .claude/wiki/verify.py); the full preregistration with every amendment and prediction;raw run outputs and the full token ledger; analysis outputs; and the manuscript sources with generated tables andbuilt PDFs. Not included: adapter weights, model caches and copies of public benchmarks (obtain them from theirpublished sources under their licences). RELEASE-MANIFEST-v023.md lists the reproduction commands; SHA256SUMS-v023 gives the file hashes.