After the fact guard, lines of the basis section that carry field names or status identifiers get one model rewrite; what is left is replaced from a table of existing wording or dropped. The body is only reported and the answer is never failed for this check. Source fixes: the 5 / 15-minute accuracy value reaches the model in words, the Argala lookup keeps the Chinese verdict only, and four prompt lines no longer name a card key. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8