Lawrencium-103/OMSF-from-DIT: OMSF

Existing AI openness frameworks (OSAID, MOF) answer the producer's question , "what must a lab release?" but not the buyer's question: "can this organization legally and operationally deploy this model in this jurisdiction?" We present the OpenModel Synthesis Framework (OMSF), a buyer-side taxonomy with four components: a six-rung Openness Ladder (L0–L5) with L1 split into sub-grades (L1a scale-capped, L1b use-restricted, L1c non-commercial); a three-tier Source Provenance Protocol (P1–P3); a three-lens deployment frame for private, enterprise, and non-profit buyers; and an African Edge & Sovereign Infrastructure (AESI) annotation for zero-egress feasibility under low-power hardware. We apply OMSF to N = 821 graded open-weight LLMs drawn from the top 1500 text-generation models on the HuggingFace Hub (snapshot 31 July 2026), after excluding quantization mirrors and library test/CI fixtures. Three findings stand out. First, repository gating is strongly associated with license restrictiveness (χ2(1, N=821) = 86.95, p = 1.1 × 10−20, φ = 0.33): 55 of 59 gated models are L1. Second, permissively licensed models are more downloaded than restricted-license models (Mann–Whitney U = 67,623, p = 0.002; median L2 = 82,267 vs. median L1 = 54,852), contradicting a widely repeated procurement heuristic. Third, only 9 of 821 graded models (1.1%) declare support for any African language. Rules-variant inter-rater agreement is κ = 0.95 (almost perfect); full human IRR is planned