Question-driven inquiry

Drafted human-editable specification · v0.1.0 · added on October 21, 2026 by convenor decision · no agent-performance claim

Epistemic purpose

  • Pose, sharpen, or reframe a research question, and pursue it through partial answers that change the aim, rather than testing a hypothesis already stated. The product is a better-posed question: what it presupposes, which alternatives count as answers, the constraints an answer must meet, and the partial answers that justified each change.

When this mode helps

Agent actions and representations

  • Write the question with its parts: the presupposition (what must be true for the question to have an answer), the contrast class (the alternatives an answer chooses among), the relevance relation, and the constraints on an acceptable answer (Fraassen 1980; Nickles 1981; Hintikka 1981).
  • Translate it into slots that can be checked, after the clinical PICO form (Richardson et al. 1995): object, contrast, observable, window.
  • When reframing after no conclusion, diagnose first: insufficient power, a measurement problem, the wrong representation, the wrong rivals, or a false presupposition. Then choose a move: change the representation or search space (Klahr and Simon 1999); widen or narrow the answer space (Elliott 2012); challenge an assumption of the literature (Alvesson and Sandberg 2011; Sandberg and Alvesson 2011).
  • Inputs: the record (the schema terms), the failed pass and its diagnosis, the budget.
  • Outputs: the question before and after, the diagnosis, the move used, and a claim record when the reframing itself is a claim (for example, that a presupposition is false).

Evidence and claim scope

  • A question is good when one can say, before the work, which outcomes would answer it, a contradiction included (Glass 2010). It is narrower than a problem agenda (Love 2008) and wider than a hypothesis, which is one candidate answer put up for test.
  • A reframing owes: both questions written in full; the diagnosis of why the pass did not conclude; which assumption or constraint changed; a search showing the new question has not already been answered, run against the question and not only a method; and a flag when the same data will answer the new question, because that pass is then exploratory, not confirmatory.
  • A false presupposition shown false is a conclusion, and so a success, not a failure.

Transitions and stopping

  • Hands the reframed question to the start of the tree: exploration when the phenomenon itself is in doubt, hypothesis assessment when rivals can now be named.
  • Proposed stopping rule (open for the room). One reframing per pass. A second only if the first produced a partial answer and changed the contrast class or a constraint so that the question now predicts an observation the old one did not; this is Lakatos’s progressive problemshift applied to questions (Lakatos 1970). Stop as a success when a pass concludes or a presupposition is shown false. Stop as a failure, recorded with both questions and the diagnosis, when a reframing changes only the wording, or after two reframings without a conclusion. No source sets the number two.

Characteristic failure

  • Reword the question without changing what would answer it; frame a broad inquiry after the fact as a narrow hypothesis test (O’Malley et al. 2009, 611); spot a gap without asking whether the assumption behind it holds (Sandberg and Alvesson 2011); check a new question’s novelty against methods only.

Human evidence and borrowing

  • Named as a mode only by O’Malley and colleagues: “open-ended and nonlinear approaches” whose inquiries “ramify as successive investigations produce partial answers that modify the original aim” (O’Malley et al. 2009, 612); and in their four modes of research practice (O’Malley et al. 2010), from which Elliott defines methodological iteration as moving “back and forth between different modes” (Elliott 2012). None of the three gives the mode an evidence standard; the criteria above are the seminar’s synthesis.
  • What a research question is. Philosophy treats the problem (a question with constraints and a demand for a solution: Nickles (1981); Klahr and Simon (1999)) and the question itself (presupposition, contrast class, admissible answers: Fraassen (1980); Hintikka (1981)). Clinical research makes evaluation usable: PICO (Richardson et al. 1995). Organization studies show that most published questions fill gaps rather than challenge assumptions (Sandberg and Alvesson 2011). For Earth science, the national surveys pose programmatic questions (National Academies of Sciences, Engineering, and Medicine 2020), from which project questions have to be derived.
  • What a research idea is. In the AI ideation literature an idea is a proposal that already holds a candidate answer and a test: problem, method, and experiment plan (Si et al. 2025; Baek et al. 2025; Wang et al. 2024), or a goal turned into hypotheses (Gottweis et al. 2026; Lu et al. 2026). An idea therefore comes after a question; its novelty is checked against methods, never against whether the question was already answered.
  • Why funders prefer hypotheses: hypothesis testing “is relatively easy to appraise” (Haufe 2013, 363); Glass (2010) defends the question as the frame for an experiment.
  • Failure as no conclusion. Most measures count null results, which are conclusions when powered. 37.3% of results presented in biomedical abstracts reach full publication (Scherer et al. 2018); null findings are mostly not written up (Franco et al. 2014); registered reports report 60% null hypotheses (Chambers and Tzavella 2022). In geophysics, 69 of 217 NSF awards from 2015 in the Geophysics program list no publication, an upper bound on non-publication, not a failure rate (NSF award API, queried October 21, 2026).
  • The actions above are design hypotheses derived from these sources, not established optimal policies.

Evaluation against past work

  • Given a project’s original proposal and its outcome report, an agent states the question in full, diagnoses why it did or did not conclude, and proposes a reframing; assessors judge whether the reframing changes what would answer the question. Candidate corpus: NSF Earth-science awards with a Project Outcomes Report but no publication.
  • Assessment definitions · Agent implementation