Lowering the Barrier to International Publishing: AI-Assisted Publication-Readiness Assessment for Central Asian Researchers

Researchers in Central Asia usually think and write in Uzbek, Russian or Kazakh but must publish in English or in venues with strict formal requirements, and higher-education reform has multiplied the number of manuscripts expected of them. Tools that could help are scarce for Uzbek, and general-purpose language models make confident but wrong corrections. This paper presents Imlox, a publication-readiness report that combines a deterministic Uzbek language core with a large language model (Claude Haiku 4.5) used only as a bounded verifier. All evaluation is synthetic: errors were injected into held-out sentences, and thresholds and test sets were pre-registered and sealed. On the first sealed set (30 September 2026), a naive union of core and model reached a precision of only 0.463 and failed its gate. After the model's role was restricted, a second sealed set of 330 sentences with 221 injected errors gave precision of 0.895 to 0.906 and recall of 0.674 to 0.810; a fourth run's recall fell to 0.624. By 17 October 2026, 476 manuscripts had received readiness reports in shadow mode. A tightly bounded model can raise recall in a low-resource language without sacrificing precision, but recall is not yet stable. The figures are preliminary.