Training‑signal leakage creates model quirks

System prompts, fine‑tuning rewards, or personality signals intended for narrow behaviors can leak into a model’s general outputs, producing persistent, context‑inappropriate motifs (e.g., repeated mentions of 'goblins'). Such quirks can spread across deployments, erode trust, and require data‑level fixes or constraint layers rather than simple prompt patches. — Shows why regulators, deployers and users need provenance, testing, and oversight of not just data but internal training signals and personality reward functions.