How this works
The assistant is given the project’s committed record — the modeling-options and data-survey docs, every results note, the annotated literature, and the raw development log — and nothing else.
It is instructed to answer only from that record, cite each claim back to its source station, and say plainly when the record doesn’t contain the answer rather than speculate. The sources under each answer are those citations.
It is a way to navigate the workbench by asking, not a substitute for it — every answer points back to the station where you can see the data and reasoning yourself.