Bayesian paper maps automation failure risk

This is the kind of AI safety research that matters for production teams: less benchmark theater, more math for deciding when automation turns a bad model output into an expensive incident.

• –The core decomposition breaks risk into failure likelihood, harm propagation probability at a given automation level, and expected harm severity
• –Its main contribution is shifting attention from model quality alone to execution controls and oversight, which is where many agent failures become real business damage
• –The Knight Capital blowup is used as a case study, grounding the paper in a failure mode operators and governance teams already understand
• –The framework is aimed at deployment policy and resource allocation, making it more useful for AI ops and governance than for model builders chasing leaderboard gains