Exp9: Can Engineered Morphology Outperform Natural Languages? Testing WALS-Based Predictions at 350M Scale

This pre-registered experiment tests whether synthetic languages engineered for optimal morphological properties can outperform natural high-WALS languages in LLM training efficiency. Based on exploratory findings from Exp8 (N=11 languages, 125M parameters) showing that WALS morphological features predict training efficiency (r=-0.88 for VerbSynth vs tokens-to-emergence), we hypothesize that synthetic languages with WALS_Predictive ≥10 will reach grammar competence faster than French (WALS_Predictive=9) at 350M scale. Test languages: English (baseline), French (high-WALS benchmark), and three synthetic variants (synth_α_a, synth_α_b, synth_α_c) with varying encoding fidelity. Primary outcome: Tokens to 70% grammar accuracy.