Allocating AI Translation Compute for Marginal Educational Access

This standalone research report develops a reproducible global portfolio model for allocating AI-assisted adaptation of open mathematics resources according to additional comfortable, usable educational access. It joins dated population evidence to exact language, variety, script, territory, learner, curriculum, accessibility, and overlap profiles; evaluates natural-language and evidence-bounded interlanguage architectures; assigns needs-specific first products; and reports low, base, and high workload sensitivities without converting token counters into prices, energy, or observed learning gains. The frozen package contains 475 population observations registered through 81 source records from 58 distinct authority labels. It scores 143 exact natural-language/profile interventions, admits 134 to the cardinal comparison, and supplies a Top 10, Top 100, needs layer, interlanguage-overlap crosswalk, Indonesian program audit, source register, scripts, and deterministic validation receipts. The corrected Top 10 begins with Bahasa Indonesia, Bangladesh Bangla, Telugu, Indian Bengali, and Vietnamese. Research synthesis, analysis, drafting, and reproducibility work used OpenAI Codex gpt-5.6-sol, Ultra. Source authors and public-project contributors retain their original credits. First-party paper and data are licensed CC BY 4.0; first-party code is MIT licensed; third-party works retain their own terms. Standalone publication with version DOI 10.5281/zenodo.22172596 and a dedicated public GitHub repository. The record contains mixed first-party CC BY 4.0 material, MIT-licensed code, and separately noticed third-party facts or references; file-level notices control. Token counters are accounting units, not FLOPs, energy, GPU time, monetary cost, or weekly-plan consumption. Nested Indonesian accounting views must not be added.