Curriculum Order in Evidence-Grounded Language-Model Tasks: An Automated Two-Family Pilot

Abstract

This empirical follow-up to From Training to Qualification reports an amended, fully automated ER-101 pilot testing first-pass curriculum order in two instruction-tuned language-model families. Qwen3-4B and OLMo-2-7B received matched synthetic course content, adapter initialization, eight-case optimizer-step bundles, three passes, and 450 updates under three paired seeds per family. Only first-pass bundle order differed. Four baselines and twelve trained checkpoints yielded sixteen evaluated configurations and 4,800 responses on 300 held-out synthetic cases. The equal-family ordered-minus-shuffled accuracy difference was -1.67 percentage points, with a module-stratified source-family bootstrap 95% task-sampling interval of [-3.11, -0.39]. Family means were +0.67 points for Qwen and -4.00 points for OLMo. The interval is conditional on the six trained pairs and does not estimate uncertainty over training seeds or architectures. The study used exact structured-output grading, shared task templates across partitions, no human raters, and no independent examination custody or instrument validation. It does not establish professional competence, qualification validity, or a universal curriculum effect. The earlier independent qualification protocol was not executed. The deposit includes a professionally typeset preprint and reproducibility materials: the executed synthetic instrument and schedules, analysis and scoring source, raw predictions, execution evidence, checkpoint file hashes, and manuscript source. Adapter files and base-model weights are excluded; all twelve trained adapters are retained by the author. All reported scores and the primary analysis can be reproduced from the deposited responses without model weights. AI assistance and the prospective protocol amendment are disclosed.