The most complete single-source account of the Anthropic-Pentagon dispute yet published — including the specific proximate trigger (an Anthropic employee allegedly called Palantir to query Claude’s use in the Venezuela raid, which the Pentagon characterised as soliciting classified information), the personality dynamics between Dario Amodei and Emil Michael, OpenAI’s stumble and amendment, and the deeper question of whether private AI companies can structurally impose constraints on military clients. Essential reading for the Anthropic-defense governance thread, and directly relevant to the broader AI sovereignty debate.