Investor due-diligence answer register

What's verified: STUD verifies structure and internal consistency of the register, not the truth of its contents. Specifically, STUD verifies: the submission is a JSON object of at most 200000 characters with at most 200 items; company and round_stage are non-empty strings; items is a non-empty array where every item has a non-empty id, category, question, and status plus a string answer; ids are unique; every status is exactly one of: answered, na, pending; every category is exactly one of: financing, technology, product, market, business_model, team, financials_returns; every answered item has an answer of at least min_answer_chars characters after trimming whitespace (default 40, clamped to 1-2000); each required category has at least min_per_category answered items (default 1, clamped to 1-20); at least min_items_total items are answered overall (default 7, clamped to 1-200); no item in a required category is left pending; evidence_refs, when present on an item, is a list of non-empty strings; and category_coverage is present as a list where each answered_count equals the answered count recomputed from the items. That recount is the ONLY arithmetic checked: no valuations, totals, or financial figures inside the answers are computed or verified. Names in required_categories outside the seven-name list are dropped, and an empty selection falls back to all seven. STUD does NOT verify that any answer is true, accurate, or complete, that evidence_refs point to real documents, that the questions match what a specific investor will ask, or that meeting the length floor makes an answer substantive. company, round_stage, company_context, and known_gaps are operator context only and never change the machine verdict.