Goal 11 — Scale-up & final comparison (gate G11)

[duckyquang]

Sub-goals

• 11.0 Corpus v3 + pairs at 10–100× (fresh scaling of the whole generation pipeline; budget decision)
• 11.1 Scaling curves per channel: does the EML/AST gap grow, shrink, or invert with data? (resolves the transformer-fairness caveat empirically)
• 11.2 Full three-track evaluation (simplification, ATP, SR) at scale; parameter-efficiency hypothesis tested explicitly (small GNN vs larger transformer at matched accuracy)
• 11.3 Frontier-LLM reference comparison, verifier-normalized (verified-correct rate vs claimed-correct rate)
• 11R Reserved repair pass

Gate G11

Every headline number has both denominators, multi-seed variance, and a scaling caveat resolved or stated.