Orchestrator-workers and an evaluator loop with subagents

Give each subagent one narrow job and a description that says when to use it, then let the main session plan, delegate, draft and loop. document-reader gets Read, Grep and a hard rule to quote and to say "not stated in this document"; report-critic gets Read only, so it can't edit the draft it judges. The decision that matters most is how you answer the exercise's questions: count from the session transcript with a small script rather than from the model's summary, and run a fixed sectioning version of the same job to compare against.