tl;dr: The author critiques the rise of LLM-based tools like "Jev" that return typed values with confidence scores, arguing developers are shipping them without evals, calibration checks, or failure-mode analysis—treating confidence scores as cargo-cult justification for errors. The deeper concern isn't that AI-driven systems fail more, but that "sometimes it just sucks" is becoming an acceptable end-state for debugging, normalizing inexplicability and eroding accountability in software.