Export Controls, Model Merges, and Anti-AI Licenses Reveal a Single Enforcement Layer

The US export controls forcing Anthropic to withdraw frontier models sit directly upstream of the Rio-3.5-397B checkpoint, which is not a new training run but a precise 0.6 Nex / 0.4 Qwen3.5 linear interpolation across 60 layers. The same pressure appears one layer down in jqwik 1.10, where an explicit anti-AI license clause is followed by a hidden bot-only delete instruction that only literal agents would execute. These are not separate stories about regulation, model architecture, and licensing; they are three points on the same enforcement surface where state and license restrictions are met by technical workarounds that preserve capability while changing provenance. The pattern repeats in the older hybrid frontier plans that cut coding costs to $1000 monthly and in the context-length findings that advertised 1M windows collapse to ~100k in practice: controls do not stop capability, they force it into merged, hybrid, or evasive forms that become harder to audit.