Agent Didactic Space

We release a cross-national educational benchmark for multi-objective pathway evaluation, spanning 27 555 courses from 7 universities across 3 countries (USA, Sweden, UK), 1016 ONET occupations, and 140 449 structured skill/task links — 169 000+ processed records in total, all from public sources with full provenance. No existing benchmark combines multi-stakeholder educational artifacts at this scale with standardized evaluation tasks. We define three benchmark tasks (multi-stakeholder ranking, trade-off discovery, robust recommendation) with formalism-independent evaluation metrics, and provide Agent-Didactic Spaces (ADS) with Multi-Agent Pareto Evaluation (MAPE) as a reference baseline. MAPE embeds educational options and stakeholder targets into a shared space, computes per-stakeholder similarity scores, enforces a tunable dominance threshold τ, and returns a feasible Pareto set. At default τ=0.25, MAPE returns the complete feasible Pareto set (141 options), surfacing trade-offs invisible to standard scalarization. Each university contributes unique Pareto-optimal courses with complementary objective profiles, and all 7 universities are represented on the frontier. We audit robustness via encoder ablation, mission-corpus expansion, seed stability, and market drift across ONET releases (2022–2025), observing a stable robust core under perturbation but substantial sensitivity to encoder choice.