Modes of inquiry: definitions, evidence, and the graph

Drafted the next meeting · graph v0.4 drafted · literature read October 16 to 19, 2026 · not yet discussed

The meeting

Central question Which kinds of scientific work does an agent family need, where was each defined, and has any been shown to raise novelty, breadth, or advance?
Anchor Elliott et al. (2016), full text, open access. Read “A richer account of scientific practice” and their Figure 3.
Companion Betts et al. (2021), open access. The rotating reader presents the coding scheme and Figure 4.
Geoscience pair, optional Pfister and Kirchner (2017) and Baker (2017), the Water Resources Research debate on hypothesis testing in hydrology.
Bring One paper of your own, with its modes named in order from the table in section 2.
Output A decision on the mode list (keep, split, merge, add), each with the source that forces it; graph v0.4 adopted, amended, or rejected; one ledger row.
  • Why these two: Elliott and colleagues name plural modes and judge research by “the alignment between the nature of the knowledge gap or challenge addressed and the combination of approaches or methods used” (Elliott et al. 2016). Betts and colleagues coded 268 ecology and evolution papers, found multiple alternative hypotheses in 6.7% and no citation or grant advantage for any hypothesis use, and still argue for more of it (Betts et al. 2021).
  • One conceptual, one empirical, and they disagree: Elliott judges by fit to the knowledge gap, Betts by the clarity that rival hypotheses bring. Neither is geoscience; the optional pair is.

1. What a mode is

  • Working definition. A mode of inquiry is a kind of epistemic work, named by the gain it aims at and the evidence its claims owe. It describes an episode or an action, not a whole paper.
  • Precedent for the word. O’Malley, Elliott, and Burian name four modes of research practice: hypothesis-driven, exploratory, question-driven, and technology-oriented Elliott (2012). Moving back and forth between them is “methodological iteration” (Elliott 2012).
  • Not the same as:
Term Unit Source Relation to our modes
Style of reasoning A long-lived tradition: six in Crombie, plus a laboratory and an “algorismic” style in Hacking Crombie (1988); Hacking (2002) Coarser; one style holds several modes
Epistemic culture A field’s machinery for making knowledge Knorr Cetina (1999) Orthogonal; a culture uses many modes
Paradigm, in Gray’s sense An era: empirical, theoretical, computational, data-intensive Hey et al. (2009) A period label; the fourth spans three of our modes
Mode 1 / Mode 2 How knowledge production is organized glossary Term collision only
Method, in this book A composition of units on the graph unit, section 6 How a mode runs the graph
Contribution type What a paper delivers: data, estimate, method, concept glossary The output, not the work
  • How modes relate, four claims in the literature:
  • The book adopts iteration. The hierarchy is a claim about practice, and section 5 tests it.

2. Thirteen definitions and where they come from

Mode Gain it aims at Defined as distinct work by Contested boundary
Question-driven inquiry A question framed so that some result would answer it, and reframed when a pass ends without a conclusion Named only by O’Malley et al. (2009) and O’Malley et al. (2010), with Elliott (2012); parts of a question from Fraassen (1980) and Nickles (1981) No evidence standard in the mode literature; the seminar assembles one
Observation and description A credible record of a phenomenon, its variability, and what is unobserved Data versus phenomena (Bogen and Woodward 1988); “‘Descriptive’ in science is a pejorative, almost always preceded by ‘merely’” (Grimaldi and Engel 2007) No peer-reviewed account gives description its own success criteria
Exploratory inquiry Phenomena, questions, and representations Experimentation “not … driven by specific theories” (Steinle 1997, S65); “mapping activities” (Franklin 2005, 893); extended to observational data (Pietsch 2015, 914) Steinle’s exploration aims at regularities, so it overlaps the last row
Hypothesis assessment Discrimination among rivals “He invents hypotheses and he invents tests for them” (Gilbert 1896, 2); the smoking gun (Cleland 2002, 480) Generation and test are separate actions, already in Gilbert
Theory and mechanistic modeling Concepts, explanations, valid derivations Models as mediating instruments (Morrison and Morgan 1999); modeling is “indirect representation and analysis” (Weisberg 2007, 207) Weisberg separates modeling from other theorizing; our row joins them
Simulation and prediction Computed consequences; assessed forecasts “A new kind of scientific method” (Humphreys 2004); Winsberg (2010); prediction as a goal (Shmueli 2010) Computer simulation studies are “material experiments in a straightforward sense,” though simulations as such are not (Parker 2009, 483, 495); the row joins a method with a goal
Instruments and observing systems A reliable new measurement capability Experimentation “has a life of its own” (Hacking 1983, 150); instruments “express knowledge” (Baird 2004); Rheinberger (2011) Merged with methods as “technology-oriented” (O’Malley et al. 2009, 612; 2010)
Methods and computation A new inference or computational capability No source defines it as a mode. Nearest: the algorismic style “dedicated to finding algorithms” (Hacking 2002); tools that become theories (Gigerenzer 1991) Gap
Cross-study synthesis A pooled estimate or a connection across studies that no single study supports Meta-analysis (Gurevitch et al. 2018); synthesis centres (Carpenter et al. 2009) Where pooling ends and a new estimate begins
Reconstruction from traces A constrained history of a past event, retained against stated rival histories Historical science (Cleland 2002); prehistory from traces (Turner 2007) Whether convergent traces confirm or only explain
Replication and robustness What survives repetition, perturbation, transfer Replication: Schmidt (2009); Machery (2020). Robustness: Wimsatt (2012); Parker (2011) Defined for experiments and model ensembles. Nearest for a one-off natural event: conceptual replication, a test of the hypothesis “with different methods” (Schmidt 2009, 91)
Inferential estimation A latent quantity with uncertainty and resolution Backus and Gilbert (1968) (resolution); Tarantola (2005) (uncertainty); inversion in philosophy: Bokulich and Oreskes (2017) Absent from every general taxonomy
Statistical modeling and empirical regularities A relation or distribution with a warranted range Crombie’s style (e) (Hacking 2002); two cultures (Breiman 2001); descriptive versus explanatory models (Shmueli 2010) Overlaps exploration
  • Three modes have no source that defines them as a distinct mode: methods and computation, inferential estimation, and replication for non-repeatable natural events, where conceptual replication is the nearest (Schmidt 2009, 91). Estimation is where geophysics wrote its own founding texts (Backus and Gilbert 1968; Tarantola 2005); its only philosophical treatment is Bokulich and Oreskes (2017).
  • Missing relative to the precedents: question-driven inquiry (O’Malley et al. 2010) and systematic invention, or engineering design (Pickstone 2007). The book keeps question origin as a ledger attribute (decided September 10, 2026). O’Malley and colleagues are the case against that decision.
  • Proposed for the room: split synthesis. Research synthesis across studies and reconstruction from traces owe different evidence (study selection versus common-cause inference), and no precedent joins them. One agent system was 85.5% accurate on its data-analysis statements and 57.9% on its synthesis statements (Mitchener et al. 2025, preprint).
  • Geoscience accounts agree on three points: explanatory pluralism, narrative with causal explanation (Kleinhans et al. 2005); several rivals at once (Gilbert 1896; Cleland 2002); model-laden data (Bokulich and Oreskes 2017; Bokulich 2020). They differ on how far laboratory accounts fit: geoscience tests hypotheses, but against observations and through “multiple converging lines of evidence” rather than single-variable experiments (Manduca and Kastens 2012, 1–2), or field work does not fit a philosophy built on the laboratory (Raab and Frodeman 2002; Baker 2017).

3. The modes on the graph (v0.4, drafted)

  • Graph v0.3 draws one hypothesis pass. In 13 of the 18 geoscience advances in section 4, the first decisive move was an observation, an instrument, or a method, made before any rival existed. v0.3 can place those only as entry states.
  • v0.4 adds three nodes, each forced by a source, drawn dashed until the group adopts them. The figure and the log are on the unit page.
Node Move Forced by
E. Explore: establish the phenomenon From data to a phenomenon: does it repeat; is it in the measurement chain or in the world Bogen and Woodward (1988); Steinle (1997); the v0.3 surprise edge (Klahr and Simon 1999, 537)
N. New capability: build and characterize an instrument or method A capability claim, made before any theory Hacking (1983), pp. 160, 163; Lehmann on the missing vertical-component records (Kölbl-Ebert 2001, 264); Keeling (1960)
I. Estimate a latent quantity Invert observations through a forward model; state resolution and uncertainty Backus and Gilbert (1968) (resolution); Tarantola (2005) (uncertainty); data are model-laden (Bokulich 2020)
  • Node labels use the words of the mode taxonomy; the crosswalk on the unit page maps each node to the modes it serves.
  • Two compositions are added: capability first (N, O, E, then Q), forced by Lehmann and Keeling; forecast verification (the expectation frozen at S before O exists, C at a later date), forced by Charney et al. (1950) (p. 245: “we were just able to keep pace with the weather”).
  • Each mode as a path on the graph:
Mode Usual entry Nodes it runs What it retains, and where
Question-driven inquiry Problem, no conclusion Q, then the start of the tree A reframed question with its diagnosis (Q)
Observation Capability, data N, O, E A phenomenon or a coverage limit (E)
Exploration Anomaly, data without theory E and O in a loop; Q reframed A candidate phenomenon or a new question (E, Q)
Hypothesis assessment Problem, prediction Q through R A retained or excluded rival (K, X)
Theory and modeling Problem, anomaly H, S A concept or a valid derivation (S); empirical adequacy later, at K
Simulation and prediction Prediction S by computation; O later; C An assessed forecast (C)
Instruments Capability N, O A characterized measurement chain (N)
Methods and computation Capability N, I An inference procedure with known failure cases (N, I)
Cross-study synthesis Problem O over many studies; C; K A pooled estimate with its spread (K)
Reconstruction from traces Anomaly, problem H; F found; O over independent traces; K A best explanation within a stated rival set (K)
Replication and robustness A challenged result V, or the whole unit rerun A corrected or supported range (V, X, K)
Estimation Data O, I A latent quantity with resolution (I)
Regularities Data without theory E, K A relation with a stated range (K)
  • Each mode is defined here by what it writes into Cleland’s schema, \(H + A + C \Rightarrow E\): the rivals \(H\), the auxiliaries \(A\) (instrument theory, forward model, processing), the conditions \(C\), the expectations \(E\), and the observations \(O\) they are compared with. The figure is the bridge between the definitions and the decision tree, which calls the modes in turn until the schema is complete enough for a claim.
Q
question
→
H
rival hypotheses
+
A
auxiliaries
+
C
conditions
⇒
E
expectations
compared with
O
observations
→
V
replication and robustness
  • Robustness varies A and C; independent data, methods, models
→
verdict
  • keep · exclude · contradict · no conclusion
  • protect: a logged move at question 7
Figure 1: What each mode writes into Cleland’s schema (drafted). Top: the question \(Q\) comes first; the hypothesis \(H\), auxiliaries \(A\), and conditions \(C\) jointly imply the expectation \(E\) (Cleland 2001, 987–88). Bottom: the observations \(E\) is compared with, the robustness check, and the verdict. Under each term, the modes that write it; a mode that writes two terms appears twice.

4. How the modes have been studied

Kind of evidence What it can show Examples What it cannot show
Philosophy and history Definitions; what warrant each mode owes Section 2 Prevalence; effect on outcomes
Cognitive and ethnographic studies How scientists move between modes Klahr and Dunbar (1988); Fugelsang et al. (2004); MacLeod and Nersessian (2016) Outcomes at scale
Science of science Associations between contribution type and citation, disruption, prizes Section 5 Causation; advance
Documented episodes Which modes did the decisive work, in what order The table below Base rates; failures are missing

Eighteen geoscience advances, coded by mode. Coding by this search from primary abstracts and secondary sources; a historian’s own coding supports four rows (drift, inner core, ice ages and Keeling, Omori to ETAS).

Advance Entry state Decisive modes, in order Primary record
Seafloor spreading, plate tectonics (1962–1967) Data without theory Instruments, observation, theory, hypothesis, synthesis Hess (1962); Vine and Matthews (1963); Wilson (1965); Sykes (1967)
K–Pg impact, Chicxulub (1980–2010) Anomaly Observation, hypothesis, synthesis of independent traces Alvarez et al. (1980); Bohor et al. (1984); Hildebrand et al. (1991); Schulte et al. (2010)
Cascadia great earthquakes (1987–1996) Hypothesis by analogy Hypothesis, observation, reconstruction, estimation Atwater (1987); Satake et al. (1996)
Gutenberg–Richter law (1944) Data without theory Instruments and method, observation, regularity Gutenberg and Richter (1944)
Omori law to ETAS (1894–1988) Data without theory Observation, regularity, estimation, simulation Utsu et al. (1995); Ogata (1988)
Inner core (1936) Anomaly Instruments, observation, theory, hypothesis, replication Kölbl-Ebert (2001)
PREM (1981) Problem Synthesis, estimation, theory Dziewonski and Anderson (1981)
Mantle tomography (1984) Capability Data archive, method, estimation, observation Dziewonski (1984)
Ambient-noise imaging (2001–2005) Capability, by method transfer Method, replication, observation, estimation Weaver and Lobkis (2001); Campillo and Paul (2003); Shapiro and Campillo (2004); Shapiro et al. (2005)
Slow slip and tremor (2001–2003) Capability, then anomaly Instruments, observation, estimation, synthesis Dragert et al. (2001); Obara (2002); Rogers and Dragert (2003)
Orbital pacemaker of the ice ages (1976) Hypothesis, waiting decades Theory, instruments and method, observation, regularity, hypothesis Hays et al. (1976)
Keeling curve (1958 onward) Problem and capability Instruments, observation, estimation Keeling (1960)
Antarctic ozone hole (1985) Anomaly in a long record Instruments, observation, hypothesis Farman et al. (1985)
Numerical weather prediction (1950) Problem and capability Theory and method, simulation, replication Charney et al. (1950)
InSAR map of Landers (1993) Capability Instruments, method, observation Massonnet et al. (1993)
Dynamic triggering (1993) Anomaly Instruments, observation, hypothesis by elimination Hill et al. (1993)
Induced seismicity, Rangely (1976) Anomaly Observation, hypothesis, field experiment Raleigh et al. (1976)
Machine-learning catalogs (2019–2020) Capability Method, observation, exploration Ross et al. (2019); Mousavi et al. (2020)
Count over these 18 Episodes
First decisive mode was observation, an instrument, or a method 13
First decisive mode was a theory or a stated hypothesis 4
First decisive mode was synthesis 1
A new instrument, archive, or computer was a precondition 15, and 1 partial
An explicit hypothesis test did decisive work 7, two of them by elimination
Synthesis across data types did decisive work 5, and 2 partial
Replication in the experimental sense did decisive work 2
  • The sample is hand-picked, mostly seismology, and successes only. The counts are hypotheses for the corpus study, not base rates.

Ten more since 2010. Chosen to fill modes the first table under-represents: robustness, prediction, regularities, and an episode that began from theory. Coded from abstracts and editors’ summaries, October 20, 2026; no historian’s account exists yet for any of them.

Advance Entry state Decisive modes, in order Primary record
Slow slip before Tohoku-Oki (2012) Capability, by method Method (repeating-earthquake extraction), observation, estimation, hypothesis on nucleation Kato et al. (2012)
Reconciled ice-sheet mass balance (2012) Problem: estimates disagreed even on sign Synthesis of altimetry, interferometry, gravimetry; robustness; estimation Shepherd et al. (2012)
Drought uplift of the western United States (2014) Capability, an instrument repurposed Instruments (GPS network), observation, estimation of water loss Borsa et al. (2014)
Injection rate and mid-continent seismicity (2015) Anomaly and existing databases Synthesis of well and catalog data, regularity, hypothesis (injection rate as the trigger) Weingarten et al. (2015)
The warming “hiatus” as a data artifact (2015) A challenged result Robustness, instrument bias correction, observation revised Karl et al. (2015)
Broad plumes beneath hotspots (2015) Hypothesis in dispute Method (full-waveform tomography), estimation, hypothesis (broad against thin) French and Romanowicz (2015)
Prompt elastogravity signals (2016–2017) Theoretical prediction Theory, observation, hypothesis, simulation, method (rapid magnitude) Montagner et al. (2016); Vallée et al. (2017)
Global seismic quieting during COVID-19 lockdowns (2020) Anomaly in a natural experiment Observation from open networks, exploration, regularity Lecocq et al. (2020)
The martian core (2021) Capability Instruments, observation, estimation Stähler et al. (2021)
Machine-learned medium-range weather forecasts (2023) Capability: reanalysis and compute Method, simulation and prediction, robustness (1,380 verification targets) Lam et al. (2023)
Count over these 10 Episodes
First decisive mode was observation, an instrument, or a method 6
First decisive mode was synthesis 2
First decisive mode was theory 1
First decisive mode was robustness 1
A new instrument, archive, or computer was a precondition 8, and 1 partial
An explicit hypothesis test did decisive work 2, and 1 partial
Synthesis across data types did decisive work 2, and 1 partial
Robustness or replication did decisive work 2, and 1 partial
  • Capability first still holds for most episodes (6 of 10, against 13 of 18 before 2010), but synthesis and robustness each led twice, which the earlier table never shows. This is a choice of episodes, not a trend.
  • The one theory-first episode waited for an instrument: prompt gravity signals were predicted, then detected in superconducting-gravimeter and broadband records of Tohoku-Oki (Montagner et al. 2016; Vallée et al. 2017).
  • Alternates not in the table: displacement reaching the Japan trench (Fujiwara et al. 2011); global atmospheric waves from the Hunga eruption (Matoza et al. 2022).
  • Ten advances chosen by impact and interdisciplinary uptake, several of them in the tables above, are candidates for the group’s success set, the positive test cases for scoring rules and agents, coded by mode from full text.
  • Two of the four hypothesis-first episodes waited years or decades for a capability or a field programme (the ice ages; Cascadia).
  • Acceptance times a source states: drift to plate tectonics, about fifty years (Oreskes 1999); the inner core, about twenty years (Kölbl-Ebert 2001, 267); Milankovitch to the pacemaker, about thirty-five years (Weart 2008).
  • Robustness in these episodes came from independent traces converging on one event, not from repeated experiments (Cleland 2002); for climate models, agreement across models confirms when read as variety of evidence (Lloyd 2010).
  • No study counts hypothesis-driven against descriptive papers in an Earth-science journal. Nearest: in ecology, 21 of 100 papers tested more than one hypothesis and 3 were purely descriptive (Betini et al. 2017); in hydrology, hypothesis testing “rarely” matches the ideal (Pfister and Kirchner (2017), abstract).

5. Do modes drive novelty, breadth, or advance?

  • Verdict. No mode has been shown to cause novelty, breadth, or advance. Five contribution types have associations with citation or disruption, all outside the Earth sciences. Two quasi-experiments touch modes indirectly.
Outcome Strongest evidence Modes involved Kind Status
Novelty, disruption New methods more disruptive, new theories less (Leahey et al. 2023); an originality index tracks self-assessed theoretical, not methodological, originality in 246 life scientists (Shibayama and Wang 2020) Methods, theory Association; the two metrics see different contribution types Not established
Breadth A cost shock to one sensing tool let generalists recruit specialists and widen the knowledge combined (Teodoridis 2018); methods and model templates cross fields (Gao and Wang 2024; Knuuttila and Loettgers 2023) Instruments, methods, theory as template One natural experiment; case studies Carriers identified; payoff of breadth in modes unmeasured
Advance Capability first in 13 of 18 episodes; hypothesis assessment decisive in 7, usually late Instruments, methods, observation; hypothesis late Coded episodes, not a sample A hypothesis for the corpus study
Citation impact Methods and synthesis cited most; hypothesis use null; failed replications cited more See next table Associations Consistent outside geoscience
Mode Novelty or disruption Citation impact Breadth
Observation Not measured Three ocean data sets out-cite 99% of oceanography articles (Belter 2014); case reports least cited in medicine (Patsopoulos et al. 2005) Not measured
Exploration Exploration-tolerant funding: lower keyword overlap with past work (Azoulay et al. 2011) Fatter tails: +97% top-1% papers and +35% flops, 73 against 393 investigators (Azoulay et al. 2011); risky strategies are 14.2% of statements and their extra reward does not cover the risk (Foster et al. 2015) Not measured
Hypothesis assessment Not measured No advantage: β = −0.024, p = 0.83, N = 268 (Betts et al. 2021) Not measured
Theory New theories less disruptive (Leahey et al. 2023) 28% fewer citations per main-text equation per page, lost among empirical citers; appendix equations had no effect (649 ecology and evolution papers, 1998) (Fawcett and Higginson 2012) Templates travel; across disciplines their warrant may stay home (Knuuttila and Loettgers 2016, 2023)
Simulation Not measured Not measured Computational templates reorganize fields (Humphreys 2004)
Instruments Prize counts disagree by coding: 82% “method” (Greenwald 2012) against 77% discoveries in physics, 1901–2004, where 11 of 17 invention prizes were measuring instruments (Bartneck and Rauterberg 2008, 74–75) Not measured directly Cheaper tools widen teams (Teodoridis 2018)
Methods New methods more disruptive (Leahey et al. 2023) Bioinformatics tools are 34% of each year’s top-20 most-cited papers (Wren 2016) AI methods rose from 0.5% to 1.3% of papers, 2000–2019 (Gao and Wang 2024)
Synthesis Not measured Meta-analyses the most-cited design in health sciences (Patsopoulos et al. 2005); reviews 2.95 times articles, mean of the 35 largest SCI-E categories, none in geoscience (Miranda and Garcia-Carpintero 2018); synthesis-centre papers 82.7 against 42.9 citations (Hackett et al. 2021) Topical diversity negatively related to citations (Hackett et al. 2021)
Replication Not measured Nonreplicable papers cited more (Serra-Garcia and Gneezy 2021); a failed replication cut citations about 9%, not significant (Hippel 2022) Not measured
Estimation Not measured Not measured Not measured
Regularities Not measured Not measured Not measured
  • Read from abstracts only: Leahey et al. (2023), Wren (2016). Miranda and Garcia-Carpintero (2018), Patsopoulos et al. (2005), Fawcett and Higginson (2012), and Bartneck and Rauterberg (2008) were checked against the full text on October 20, 2026. The rest were read in full or in an open preprint; the private analysis lists the basis per source.
  • Correctness and citation pull apart. Citations do not reflect soundness or originality (Aksnes et al. 2019); the replication rows show it directly.
  • Breadth is measured in topics and references, not modes. A pivot away from one’s own past references cuts the top-5% hit rate from 7.4% to 2.2% (Hill et al. 2025); publishing into a distant audience makes a hit 3.5 times as likely (Shi and Evans 2023); generalists or specialists win depending on the field’s pace (Teodoridis et al. 2019). A hypothesis neither paper tests: carrying a home method to a distant audience is an expedition, not a pivot.
  • What travels and what stays. Methods, instruments, and model templates cross fields; validation standards and task routines stay behind (MacLeod 2018).
  • One person, two modes. Bimodal modeler-experimenters reported losing time switching, and their field doubted it could become the norm (MacLeod and Nersessian 2016, 409).
  • Teams of different approaches. Diversity helps on rugged problems and costs speed; on smooth problems the best performers win (Hong and Page 2004; Grim et al. 2019, 108; Thoma 2015; Wu and O’Connor 2023). The gain in Weisberg and Muldoon’s model is the mavericks’ own work (Weisberg and Muldoon 2009; Alexander et al. 2015, 440).

Claims this evidence does not support.

6. From modes to agents

  • No taxonomy of AI in science is organized by mode. Surveys sort by workflow stage, autonomy, or mechanism (Zheng et al. 2025; Reddy and Shojaee 2025); one sorts by kind of understanding (Krenn et al. 2022). The nearest to a mode list is the five geoscience problem classes of Karpatne et al. (2019).
  • Coverage. Coded by this search over 19 published systems and benchmarks: the core loop runs methods in 7, hypothesis generation and ranking in 6, regularities in 4, literature synthesis in 4, and simulation, instruments, or replication in none. All 19 start from a supplied goal; none starts from an anomaly or reconstructs an event from physical traces.
  • Roles. One multi-agent system splits agents by epistemic function: generation, reflection, ranking, evolution (Gottweis et al. 2026). Others split by disciplinary persona, career rank, or tool (Swanson et al. 2025; Su et al. 2025; Schmidgall et al. 2025; Ghareeb et al. 2026).
  • Drift. AI-augmented papers concentrate where data are abundant, topics shrink by 4.63%, and follow-on engagement falls by 22%; the authors read it as a move from questions of origin to operational ones, which they did not measure by mode (Hao et al. 2026). Against it, LLM adopters cite 11.9% more books (Kusumegi et al. 2025).
Design hypothesis Evidence for What would refute it
Specialize agents by the evidence a mode owes, not by persona or by Peirce’s labels Search access for the reflection agent cut implausible hypotheses (Gottweis et al. 2026); agents split by abduction, deduction, and induction were tested only on logic puzzles (Abdaljalil et al. 2025) A generalist with the same tools and budget runs as many mode-specific checks as the specialists
Each mode agent has an evaluator the generating model does not control Automated reviewers scored 6.1 of 10 against 3.8 from humans (Schmidgall et al. 2025); the A-Lab scored itself (Leeman et al. 2024) Self-scores track blinded expert scores
Synthesis is its own agent, scored apart from the analyses it draws on 57.9% against 85.5% accuracy (Mitchener et al. 2025) Synthesis accuracy equals analysis accuracy once scored separately
A system without a mode budget drifts to data-rich modes Hao et al. (2026); the coverage counts above Logged mode shares in the group’s runs match the ledger without a budget
A method carried to a new field must carry that field’s validation standard MacLeod (2018); the Ising model lost its renormalization-group warrant moving from physics to socio-economics (Knuuttila and Loettgers 2016) Transferred methods pass the destination’s checks without being given them
Specialists built on one base model are not a diverse team Diversity’s advantage depends on solvers that differ (Hong and Page 2004; Grim et al. 2019) Same-model specialists generate as many distinct rivals as specialists on different models
The exploration agent enters from an anomaly None of the 19 systems does; Kosmos “can sometimes veer in unexpected trajectories” (Mitchener et al. 2025, 19) An agent given a goal finds the anomalies an anomaly-entry agent finds
  • The polymath hypothesis, restated for agents. A human’s cost of holding two modes is training and switching time (MacLeod and Nersessian 2016); an agent’s switching cost is near zero. That is where a generalist could beat people. Test: one generalist agent against a team of three mode specialists on a task that needs simulation and observation, scored by a blinded specialist in each mode, recording whether each mode’s own checks ran.

7. For the room

  1. Decided by the convenor (October 21, 2026): synthesis split, question-driven inquiry added. For the room: does any other mode need splitting?
  2. Is methods and computation a mode or a contribution type? Hacking gives it a style; no one gives it its own warrant.
  3. Graph v0.4: do E, N, and I place your own paper better than v0.3 did?
  4. Which outcome should a mode agent be scored on, by whom, and against what baseline?
  • Exercise. Code your paper by mode, draw it as a path on the decision tree, and mark the node where its decisive move happened. Then do the same for one row of the episode table and compare.
  • Record. Sign the notes; one ledger row; the decision on the list and on graph v0.4, with the source behind each change.

What this page does not establish

  • That any mode causes novelty, breadth, or advance.
  • Base rates: the twenty-eight episodes are a selected set, coded by this search; four rows rest on a historian’s coding, and the ten after 2010 rest on abstracts.
  • That the agent hypotheses hold: none has been run on the group’s agents.