Cost is every model call the account caused this month: replies, tool rounds and the memory extractor; comparison runs are left out. of net is that cost as a share of what the plan leaves after payment fees; the zone follows the cost work in prototype.md: safe up to 30%, watch to 60%, over to 100%, loss beyond. A month at this pace scales the window to a month (a calendar month to its own length, a range of days to 30). For a month, the zone is judged by the cost so far; for a range of days, by a month at this pace, so a 7-day window does not make everyone look safe. Cache hit is the share of input tokens read from the cache; cache write the share written to it (a new chart, a tool result, or the whole history again after a pause). After a pause counts messages sent after the cache had expired (60 minutes on Claude, 30 on OpenAI): each one rewrites the whole history at full price, so a person who asks one question every two hours costs far more per message than one who asks eight in a row. Reply length is output tokens per message: the largest single cost driver now.