PrimePE claims lattice basis, trims perplexity

Interesting benchmark signal, but the theory is doing a lot of speculative heavy lifting.

• –The practical part is the strongest: composite-tiered frequencies, learned scaling, and VHT2 KV compression all come with concrete PPL and compression numbers.
• –The prime-vs-composite falsification is more persuasive than the grand narrative, because it suggests the useful ingredient is lattice-aware frequency placement, not primality itself.
• –The zeta-zero and “universe” framing reads as conjecture, not established mechanism; it makes for a memorable pitch, but it will need much tighter independent validation.
• –The llama.cpp integration angle matters for developers: if the gains hold, this is an inference-side improvement rather than a retraining-heavy research curiosity.
• –Net: a genuinely interesting architecture experiment with strong empirical claims, but the metaphysics should be treated as hypothesis, not conclusion.