Claude Outputs Turn Illegible After Intense Agentic RL Training

A reply by @teortaxesTex addresses reports of Claude model outputs growing hard to read. The reply follows mention of missing words, run-on sentences, and awkward phrasing in recent responses, including from Opus. It states that such problems represent a natural outcome of sufficiently intense agentic RL. The packet records the reply and a generated summary of the issue but contains no separate confirmation of the described output problems. The exchange appears among visible replies on X in the supplied source lines.