The public record now separates the canonical 30-word reference lexicon from larger generated corpora and later experimental benchmarks. Scale is reported without treating every generated item as an independently validated lexical unit.
“Bigger than Core-30” does not mean one canonical lexicon was simply enlarged. The papers introduce several distinct artifacts for different scientific questions.
Core-30 is the 30-word reference lexicon used in early clustering, semantic-axis, and morphological-hijacking experiments. It remains useful because it is fixed and historically traceable, but it is no longer an adequate description of the full experimental scale of VARZIN/LUXVAR.
The LUXVAR v2.2 manuscript analyzes a broader designed root inventory. It describes these roots across six semantic fields and other generator dimensions. That source-level description is distinct from Core-30’s five designed axes and is not silently collapsed into the same taxonomy.
The manuscript reports an exhaustive combinatorial generator producing about 1.3 billion forms. This is generated structure by design, not a corpus collected from human language use.
At the reported audit scale, H/Hmax = 0.9995 and the stated first-order Cramér’s V associations remain below 0.02.
Taxonomy note: Core-30 uses five designed axes; the broader v2.2 corpus description uses six semantic fields. VARZIN preserves that discrepancy as documented rather than rewriting one into the other.
Appendix E of the Morphological Hijacking manuscript introduces a fresh 20-root × 15-prefix variant: 300 words, ten times the Core-30 word count. Its frozen GPT2-small baseline was weaker than the original adversarial Core-30 condition (ARI −0.053, PHR 0.869), consistent with a milder prefix/root token-length imbalance.
With the full trained recipe, leave-one-root-out evaluation over five held-out roots and three seeds each produced pooled ARI = 0.992, SD = 0.007. The paper explicitly limits this result to the easier same-script, same-word-order condition; it is not presented as outperforming the harder cross-script 7B tests.
Appendix F builds a separate 5 roots × 6 prefixes × 12 positions = 360-word lexicon with an opaque, shuffled suffix mapping to prevent trivial numeral/ordinal leakage.
Shuffled-label control: 0.025 ± 0.006.
Shuffled-label control: 0.299 ± 0.044.
Shuffled-label control: 0.232 ± 0.048.
Boundary: this is targeted group-position classification after explicit remediation. The same paper separately reports that this success does not establish homomorphic or systematic composition.
12 structural positions, four synthetic morphemes, 30 contexts, 24 training contexts, 6 held-out contexts, and 30 instances per position/context configuration.
Across all 29 hidden-state indices, the preregistered surface-span representation was highly decodable; 25/29 indices reached BA = 1.000.
Across 29 indices × two composers, no combination met the complete preregistered positive criterion on unseen ordered pairs.