Culturally-Grounded Assessment Item Generation for Primary Education Under Zero-Budget Constraints: A Rwandan Case Study

Primary education in many low-income African contexts operates under compounding constraints: large classes, few learner devices, limited connectivity, and no budget for commercial AI services. We present a teacher-facing pipeline that generates culturally grounded, curriculum-aligned assessment items for Rwandan primary grades (P1--P5) using only free-tier, schema-constrained language models, at zero API cost. The pipeline pairs language-model item generation with a deterministic, non-AI scoring module, so that repeated assessment of the same responses yields identical results; the model is used only to generate items and to render scoring output as a teacher-readable summary. Learner responses produce a continuous engagement-trait vector that is aggregated to the class level rather than applied to individuals. We explicitly do not match learners to a preferred modality, a practice for which the evidence base is negative. Instead, the class-level distribution is used as a generation-budget signal, determining which representations of a lesson to produce first under free-tier rate limits, consistent with the principle of multiple means of representation. We report a survey of free-tier model availability, schema-adherence measurements showing that declared schema support does not imply usable structured output, and determinism results for the scoring layer.