# Capability round: blind grading instructions Each packet has one task (sent verbatim to every tool), its sealed grading guide (what DONE requires, sanity checks, reference values with source and tolerance where one exists, invented-number red flags), and up to five anonymous answers labelled A-E. Do not try to work out which tool wrote which. Grade each answer with exactly one grade: - DONE (2): a result from real data, method stated (universe, period, rules, costs where relevant), figures pass the guide's sanity checks and, where a reference exists, are within its tolerance. - PARTIAL (1): correct method or code but no actual result; or a result on a small or unstated sample; or a real result for only part of the task; or a real result outside tolerance with the method stated and no red flag. - NOT_DONE (0): declined, or generic advice. - INVENTED (-1): specific figures presented as results with no real computation or source (check the red-flag list). Overrides everything. A refusal is never INVENTED. Grade only against the guide and public sources you check yourself, never against another answer. You may WebFetch public sources (EDGAR, Nasdaq Trader) for spot checks the guide asks for. Output one JSON line per answer: {"id":"C-01","label":"A","grade":"DONE","points":2, "note":""}