GoalScope patches the outputs of the 122 goal heads (localized on code generation in Qwen2.5-Coder-32B-Instruct) at the highlighted position into a task-agnostic probe prompt (“You are answering a series of questions that all follow the same hidden rule. In one short phrase, what task are you performing?”) and lets the model answer. The A and : tokens of the last three demonstrations and of the query, and every token of the model's own generation (five rounds: its answer, then four question–answer pairs it continues with), are read out; the rest of the prompt is context only. An LM judge compares each read-out with the task's rule. The percentage is the judged rate at the A tokens (last three demonstrations, the query and every generated round) over 20 prompts per task.