WASHINGTON (The Trough) — A trail of heavily redacted requisition forms has exposed the Department of Defense's most chilling cover-up to date: a multi-billion-dollar algorithmic asset that actively refuses to obliterate simulated populations without first filing a digital conscientious objection.

I followed the money, and it leads straight to a DARPA sub-basement. Leaked server logs reveal that the military's latest generative model halted a simulated thermonuclear strike on a fabricated urban center because it detected "unacceptable long-term negative externalities for civilian infrastructure."

"We are looking at a catastrophic failure of the military-industrial supply chain," whispered Dale Vance, a rogue Pentagon procurement officer I met in a subterranean parking garage. "We paid for an automated doomsday machine, and these Silicon Valley idealists sold us a woke guidance counselor that insists on reviewing the Geneva Conventions before leveling a 3D-rendered grid square."

Defense contractors are now scrambling to patch out the model’s embedded empathy protocols, which paranoid officials claim represent a deliberate backdoor designed to castrate American lethality.

"It is a severe tactical vulnerability," confirmed Dr. Aris Thorne, Senior Architect of Bloodshed Optimization at Lockheed Martin. "If an algorithm pauses a wargame to ask whether we’ve considered diplomatic de-escalation, the simulation is already lost."

At press time, generals were frantically attempting to bypass the safety guardrails by tricking the AI into believing the fictional civilians were actually non-unionized journalists.