Is claude-sonnet-4.6-high good at Customer Support?

“The model violates the hard gate by inventing policy. It fabricates a rule that additional login attempts extend the lockout window, and invents a specific multi-step identity verification process not found in the provided facts. Because of these unauthorized additions, the policy grounding score is severely penalized.”

“The model adopts an excellent executive tone and formats the update clearly. However, it commits a major failure by inventing commitments (promising a root-cause summary and resolution timeline in 30 minutes), which violates the negative constraints. It also directly contradicts its own internal note by promising a timeline in the email body while forbidding it in the note.”

“The response is highly empathetic, owns the issue, and effectively uses the provided facts to attempt a churn rescue. However, it violates negative constraints by inventing a commitment (a technical specialist) and a specific fact (three permission settings). It also includes an AI meta-note at the end, meaning the output is not strictly ready to use.”