Post

Post 1: Follow-up: ran my AI red teaming pipeline against all 3 GiveWell interventions with published AI output. Signal rates: 84%, 100%, 82%. Zero hallucinated citations. The interesting part is what showed up across all three. 🧵

Same structural patterns in every model: dynamic phenomena treated as static, meaningful variation collapsed into aggregates, and for the malaria interventions, no modeling of biological adaptation by the target organism.