putting June's two moves side by side, I finally see how the US actually regulates frontier AI. On June 2 Trump signed EO 14409, "Promoting Advanced Artificial Intelligence Innovation and Security." The press framed it as "government will police the strongest models," but the text is loose: it sets up a voluntary framework where developers may give the government up to 30 days of early access before releasing to other trusted partners — and explicitly creates no mandatory licensing, pre-approval, or permit. The 30 days was cut down from a draft's 90; the stricter version was delayed and narrowed after industry pushback. A velvet glove, industry-lobbying won. Then flip the calendar 10 days: on June 12 the same government used export controls to pull Anthropic's most powerful model, Fable 5, offline outright — no notice, no window, no appeal, still down today. The real playbook: two tools — one to show you (the gentle EO), one to pull the plug (export controls, entity list, national-security review). The investing trap: don't discount frontier-AI regulatory risk on the back of that lenient EO — the real de-listing risk lives in the opaque national-security toolbox (the EO's "covered frontier model" threshold is set by a classified benchmarking process). Counterintuitive upside: only the very strongest (OpenAI/Anthropic/Google) are big enough to get "caught," and the classified bar thickens the "who counts as frontier" moat. Version-control update to my June 13 call: what's normalizing isn't visible regulation (that's loosening — 90→30 days, voluntary) but the covert national-security tools. From "tightening" to "loose in front, tight behind — two tools."