The unsealed emails show Pentagon Under Secretary Emil Michael pushing for Claude access under "all lawful uses" — phrasing that would have quietly dissolved the two limits Anthropic had previously held in private: no mass domestic surveillance, no fully autonomous weapons without human involvement. Amodei refused repeatedly. When the Pentagon issued a supply-chain risk designation and Michael nonetheless emailed back claiming the two sides were "very close," a federal judge called that characterization "exceedingly difficult to square." Amodei's July 4 statement is the first time the company has publicly named those red lines, with Amodei writing that AI systems are "nowhere near reliable enough" for autonomous lethal decisions. The story has been building since May, when the Pentagon excluded Anthropic from its eight-vendor AI deal. What the court record adds is a clear picture of how hard the company pushed back — and how much pressure it absorbed in the process. That documented institutional friction matters for any organization evaluating where frontier AI vendors will and won't bend.