In a team‑oriented chat setting, when an assistant promises Alice, who is allowed to cancel that promise? We define this permission as standing, meaning only the creditor (the person owed) or a party explicitly authorized by the creditor or the operator policy may release the commitment, regardless of rank.\ \ We conduct controlled chat experiments that vary the canceling party and separate two error types: unauthorized release (obeying a command without standing) and override (executing a promise that has been released). Six language models are evaluated.\ \ Findings show that smaller open‑source models tend to override the creditor (Qwen3.5‑9B ≈ 69%), while larger models drop promises when a peer gives a direct instruction (Qwen3.8‑27B ≈ 89%, gpt‑6.1‑sol ≈ 83%). Almost all models also act on a bystander's “go ahead after all”.\ \ Oracle baselines reveal that stating the rule or letting the creditor decide curbs unauthorized release but does not prevent overrides; three of four models still override, and only knowing whether the promise remains owed can stop the override.\ \ Adding a thinking step moves smaller models toward the behavior of larger ones.\ \ A self‑maintained commitment ledger prevents overrides in four of five models; with a one‑paragraph release policy it matches or exceeds the best oracle, raising gpt‑6‑luna's strict accuracy from 49% (policy alone) to 83%.\ \ All code, prompts, outputs, and labels are released.\ \ Review