1 / 31
Denolle Lab · MetaScience, Scientific Inquiry, and Agents
Modes of inquiry, and a decision tree for research
Working meeting · October 2026 · decided by the convenor on October 21: controller v0.5.1, the 13 modes, graph v0.4.2 as vocabulary, the failure definition, breadth v1 (provisional). Everything else is drafted; the full list is on the decisions slide.
2 / 31
Today
- The thirteen modes: what each is, where it was defined, what it writes into the record.
- Cleland's schema H + A + C ⇒ E as the record every project fills.
- The controller: a decision tree of nine questions over graph v0.4.2, adopted by the convenor.
- Success: how we measure impact, breadth, novelty, and advance, and what counts as failure.
- What is still open for the room (last slide).
The book: denolle-lab.github.io/metascience-and-AI · "Modes of inquiry", "From graph to decision tree", "Success set"
1 · The modes3 / 31
A mode of inquiry is a kind of epistemic work
- Defined by the gain it aims at and the evidence it owes.
- Closest precedent: four "modes of research practice": hypothesis-driven, exploratory, question-driven, technology-oriented (O'Malley, Elliott & Burian 2010).
- Moving between modes is "methodological iteration" (Elliott 2012).
- Why it matters here: each mode will become a specialist agent, scored against the evidence its mode owes.
O'Malley, Elliott & Burian (2010) Studies in Hist. Phil. Biol. Biomed. Sci.; Elliott (2012)
1 · The modes4 / 31
The thirteen modes and the gain each aims at
| Mode | Gain it can support |
|---|
| Question-driven inquiry | A question framed so that some result would answer it |
| Observation | A credible phenomenon, distribution, baseline, or coverage limit |
| Exploration | A developed phenomenon, representation, or question |
| Hypothesis assessment | Discrimination among rival explanations |
| Theory and modeling | A concept or valid derivation that changes what can be tested |
| Simulation and prediction | Assessed forecasts, conditional model consequences |
| Instruments | A reliable new measurement capability |
| Methods and computation | A new inference or computational capability |
| Cross-study synthesis | A pooled estimate or connection no single study supports |
| Reconstruction from traces | A history of a past event, against rival histories |
| Replication and robustness | A corrected result or independent corroboration |
| Inferential estimation | A latent quantity with uncertainty and resolution |
| Regularities | A relation with a warranted range of application |
Glossary page of the book; synthesis split and question-driven inquiry added on October 21, 2026 (convenor)
1 · The modes5 / 31
Three modes have no defining source in the literature
- Methods and computation. Nearest: Hacking's "algorismic" style (Hacking 2002).
- Inferential estimation. Its founding texts are geophysical: Backus & Gilbert 1968, Tarantola 2005.
- Replication of a one-off natural event. Replication is defined for experiments; the nearest is "conceptual replication", a test with different methods (Schmidt 2009).
- Contested boundaries: instruments vs methods (merged as "technology-oriented", O'Malley et al. 2009, 2010); exploration vs regularities; cross-study synthesis: where pooling ends and a new estimate begins; reconstruction: whether convergent traces confirm or only explain.
Literature reports in private/analysis/modes-literature/ (checked Oct 20, 2026); boundaries from the modes page, section 2
1 · The modes6 / 31
18 documented geoscience advances: capability came first
| What did the decisive work | Episodes |
|---|
| First decisive move: an observation, an instrument, or a method | 13 of 18 |
| First decisive move: a theory or a stated hypothesis | 4 of 18 |
| A new instrument, archive, or computer was a precondition | 15 of 18 |
| An explicit hypothesis test did decisive work (usually late) | 7 of 18 |
Episodes include seafloor spreading, the inner core, PREM, Keeling CO₂, Cascadia, ambient-noise imaging, ETAS. Hand-picked successes, mostly seismology, coded by us; a historian backs four codings. A hypothesis for a corpus study, not a base rate.
Modes page, section 4; episode codings in private/analysis/modes-literature/05-geo-advances.md
2 · Cleland's schema7 / 31
Cleland's schema: H + A + C ⇒ E
- H the hypothesis; A auxiliary assumptions (theory of the instrument, forward model, processing, background laws); C the conditions, set or found; E the expected observation.
- A mismatch refutes the conjunction H + A + C, not H alone (Duhem). Every verdict must say which of A and C was varied, and which was revised.
- Example (seasonal seismic velocity change): H = groundwater storage; A = depth sensitivity of the surface waves, clock and instrument stable; C = found, wells outside the array; E = velocity lags rainfall by weeks.
Cleland (2001) Geology 29:987; Cleland (2002) Phil. Sci. 69:474; Duhem (1906)
2 · Cleland's schema8 / 31
What each mode writes into the schema
compared with
→
V
replication and robustness
- Robustness varies A and C; independent data, methods, models
→
verdict
- keep · exclude · contradict · no conclusion
Same figure as the modes page. A mode is defined by what it reads from the record and what it writes
3 · The graph and eight papers9 / 31
Graph v0.4.2: fifteen moves, each named by the first letter of its verb
N New capabilityE ExploreO ObserveI InferQ QuestionH HypothesizeS Simulate or deriveD Design
F Find conditionsC CompareV Vary & replicateX eXcludeK KeepP ProtectR Refine
Built from meeting 1 (Platt 1964; Cleland 2001), Klahr & Simon (1999), and the 18 episodes. Dashed: proposed in v0.4
3 · The graph and eight papers10 / 31
How we coded eight papers
- The papers: the success set (four canonical, four from 2025–26; selection on the "Measuring success" slides).
- The coding: from the full text; each step gets a node, a mode, what the authors did, and the page.
- Who coded: the walks shown are one Claude subagent's codings (codebook v0, graph v0.4). On Oct 25, two blind model coders (Opus 5.5, Sonnet 5.5, codebook 1.3.0) recoded all eight, plus five controls. Node-set overlap 0.76, order similarity 0.54, node alpha 0.63 against a 0.67 gate. No human coder yet.
- The order is the order of the paper's account, not necessarily the order of the work.
Codings: private/analysis/success-set/<key>.md, build/steps.py; blind codings and agreement: reliability/RESULTS.md (Oct 25, 2026)
3 · The graph and eight papers11 / 31
Shapiro et al. 2005: a capability, no rival hypotheses
1NNew capability→2NNew capability⇢3EExplore⇢4VVary and replicate→5FFind conditions→6OObserve→7IInfer⇢8VVary and replicate⇢9CCompare⇢10KKeep
Noise cross-correlations become Green functions; one month of noise is stable and matches earthquake Rayleigh waves; tomography follows. The claim is a capability. Two of three coders code no H; Sonnet 5.5 codes the cited Green-function premise as H. Map: rust = entry node, teal = visited, grey = never used. Strip: → along a drawn edge, ⇢ a jump the graph reaches only through other nodes.
Walking the graph page; step pages in its tables. Map and strip: coder 1 (codebook v0); blind codings: reliability/codings/papers/
3 · The graph and eight papers12 / 31
Sun & Davis 2026: most of the graph, rivals excluded
1NNew capability→2OObserve→3EExplore⇢4HHypothesize⇢5XExclude⇢6IInfer⇢7QQuestion⇢8SSimulate or derive⇢9VVary and replicate⇢10CCompare⇢11XExclude⇢12KKeep→13RRefine
Borehole pressure records become strain. Rivals to tectonic strain (sensor drift, consolidation, compaction, leakage) are tested and mostly excluded. Map: rust = entry node, teal = visited, grey = never used. Strip: → along a drawn edge, ⇢ a jump the graph reaches only through other nodes.
Walking the graph page; step pages in its tables. Map and strip: coder 1 (codebook v0); blind codings: reliability/codings/papers/
3 · The graph and eight papers13 / 31
Normandeau et al. 2025: kept by an extra assumption
1OObserve→2EExplore⇢3HHypothesize⇢4IInfer⇢5SSimulate or derive⇢6CCompare⇢7PProtect⇢8KKeep
Seafloor blisters seen between surveys. Ice expansion gives at most 0.05 m of 0.2 m of relief; an assumption (the ice lasts years) closes the gap. Coder 1: entry O, the assumption at P. Blind coders: entry Q; P (Opus 5.5) or R (Sonnet 5.5). Map: rust = entry node, teal = visited, grey = never used. Strip: → along a drawn edge, ⇢ a jump the graph reaches only through other nodes.
Walking the graph page; step pages in its tables. Map and strip: coder 1 (codebook v0); blind codings: reliability/codings/papers/
3 · The graph and eight papers14 / 31
What the eight walks show, by coder
| Test | Coder 1 (91 steps) | Opus 5.5 (169) | Sonnet 5.5 (108) |
|---|
| Every step lands on a graph node | yes | yes | yes |
| Transitions along a drawn edge (or a repeat) | 24 of 83 | not counted | not counted |
| Returns to a node already visited | 11 | 83 | 33 |
| Returns to a mode already used | 27, in 8 papers | 78, in 8 | 38, in 7 |
| Pair with the most hand-offs | hypothesis ↔ theory 6 | estimation ↔ observation 17 | estimation ↔ observation 13 |
| Opens with a question step | 1 of 8 | 5 of 8 | 6 of 8 |
| Then methods, instruments, observation, or estimation | 8 of 8 | 7 of 8 | 5 of 8 |
| Last step in synthesis | 4 of 8 | 2 of 8 | 0 of 8 |
Coder 1 finds the loops mostly between modes. The blind coders cut finer steps and find about as many node returns as mode returns. Where the loops sit, and where papers enter and leave, is not yet coder-independent.
Coder 1: build/steps.py (codebook v0, 11 modes). Blind coders: reliability/codings/papers, codebook 1.3.0 (13 modes), final-paper sequences. Consecutive repeats merged; one Shapiro step kept as "replication" (28 mode returns if read as robustness)
3 · The graph and eight papers15 / 31
What the walks cannot show
- Papers report the argument, not the work. Failed attempts and reruns rarely appear. Between protocol and paper, 62% of trials changed a primary outcome (Chan et al. 2004).
- The walks shown had no field for a failed attempt (codebook v0), so a return and a next step look alike. Codebook 1.0.0 (Oct 21) added a move on every step, including return_after_failure; the blind recodings use it.
- One coder for these walks, coding against the graph it tests: "no new node needed" is partly the instruction. The two blind model coders also get the graph; they agree on node use at alpha 0.63, below the 0.67 gate. No human coder yet.
- Granularity varies (8 to 17 steps); supplements unread for four papers; the walks are successes; five matched controls coded since; no failures.
- Process documents (peer-review files, open discussions, versions) do record the loops: next slide.
Chan et al. (2004) JAMA 291:2457; walks page, "What the walks cannot show"; agents/coder/CODEBOOK.md change log; private/analysis/reliability/RESULTS.md (Oct 25, 2026)
3 · The graph and eight papers16 / 31
Process documents show the loops the papers hide
| Nine Earth-science papers with review files or versions | Count |
|---|
| Loops recorded in the process documents | 126 (61 new analysis, 65 argument or deletion) |
| Returns visible in the final papers' own sequences | 7 |
| Loops readable as loops in the final paper | 4 (3%) |
| Kept, but written as if planned | 70 (56%) |
| No trace in the final paper | 52 (41%) |
| Where loops return | V new control 29 · R narrowed claim 24 · H new rival 17 |
A reviewer-forced control survives as an Extended Data figure that reads as planned; rivals named in review are mostly argued away (15 of 17); narrowed claims leave no trace. Papers: Gräff 2025, Bi 2025, Isken 2025, Licciardi 2022 (Nature review files); Kratzert 2018, 2019, Ryan 2015, van den Ende 2021 (EGU discussions); Lam 2023 (arXiv versions).
private/analysis/success-set/loops/REPORT.md; one coder per paper; papers chosen for substantive reviews
4 · The decision tree17 / 31
Graph v0.5: three levels
- The record: Cleland's terms H, A, C, E, and the observations O, with provenance.
- The modes: each runs its own inner loop over graph nodes until its stopping rule holds (estimation: misfit at noise level, resolution stated; exploration: the phenomenon repeats).
- The controller: a decision tree that reads the record and calls the next mode. A paper is written when the claim carries the evidence its mode owes.
- Automated science loops inside one mode, with H and A fixed (Robot Scientist, Coscientist, adaptive design). None decides when to leave a mode. Only Co-Scientist protects an auxiliary in full, for one of two blind coders. The tree adds the controller.
From graph to decision tree page, sections 2 and 3; P marks: two blind loop codings, reliability/codings/agents/ (31 systems)
4 · The decision tree18 / 31
The decision tree, v0.5.1 (adopted): questions 1 to 3
Start: a question, an anomaly, data, a capability, or a goal
1Is there a phenomenon?
It repeats, and it is not an artifact of the measurement chain.
no: the chain is suspectInstruments · methods N
characterize the chain↺ back to 1
no: not yet stableExploration E, O
vary conditions until it repeats↺ back to 1
yes
2Is the target observed directly?
Or is it a latent quantity behind the data?
noEstimation I
invert through a forward model; state resolution→ then 3
yes
3What is the aim?
A mechanism, a relation, a forecast, or a pooled estimate.
a relationRegularities C, K
fit across repetitions; test out of sample→ to 7
a forecastSimulation S, C
freeze the forecast; compare when the outcome arrives→ to 7
a pooled estimateCross-study synthesis O, C, K
state the inclusion rule; model dependence between studies→ to 7
a mechanism
Questions run down the left; an answer that sends work elsewhere calls a mode (right) and returns to the question marked ↺
4 · The decision tree19 / 31
The decision tree, v0.5.1 (adopted): questions 4 to 6
4Two or more rivals?
Incompatible explanations, with "artifact" kept as one.
noTheory · hypothesis H
generate rivals; record how each was generated↺ back to 4
yes
5Does each rival predict something different?
An expectation for each rival, from H + A + C.
noTheory · simulation S
derive or compute each expectation↺ back to 5
still identicalA new observable N
a capability that separates the rivals↺ back to 1
yes
6Can you set the conditions?
Experiment, or conditions found in nature.
yesHypothesis assessment D, F, O, C
design the test, set the conditions, observe, compare→ to 7
no: nature repeats itRegularities F, O, C
nature's repetition→ to 7
no: one eventReconstruction from traces F, O, C
independent traces of the same event; rival histories→ to 7
evidence in hand
A mechanism needs rivals, distinct predictions, and conditions; found conditions split by whether nature repeats the event
4 · The decision tree20 / 31
The decision tree, v0.5.1 (adopted): questions 7 to 9
7Does the result survive variation?
Replication and robustness (V): vary the auxiliaries A and the conditions C; independent data, methods, models.
no: an artifact?↺ back to 1
no: an auxiliary is wrongProtect P
revise A instead of H; logged; must predict something new (cap: open)↺ back to 5
yes
8Verdict
Keep, exclude, contradict: each is a conclusion.
survivors remainRefine R
split the survivors into subhypotheses↺ back to 4
no rival fitsSurprise E
an outcome no rival predicted: explore it↺ back to 1
no conclusionQuestion-driven inquiry Q
reframe the question; record why the first framing could not be answered↺ back to start
still none after reframingFailure, recorded
name the observation that would decide
a conclusion: draft the paper
9Does the claim survive review?
Co-authors, referees, editors, readers.
a new control asked forRobustness V
run it; it usually enters as a planned-looking figure↺ back to 7
a new rival namedHypothesis assessment H
test it, or state why it is excluded↺ back to 4
the claim is too broadRefine the claim R
narrow it to what the evidence carries→ to Keep
the request is declinedArgue and log
keep the design; record the reason→ to Keep
accepted
Conclusion: keep, exclude, or contradict the claim, with its status, and publish the record, the path including its loops, and the evidence the mode owes.
Every route that produces evidence passes question 7, robustness, before the verdict; question 9 is review
4 · The decision tree21 / 31
The eight papers on the tree
| Paper | Where it looped | Reading |
|---|
| Shapiro 2005 | robustness, observation, methods | Question 1 by method; estimation; keeps a capability, never reaches question 4 |
| Sun 2026 | theory, robustness, hypothesis | Question 1 loops on drift versus signal; questions 4–7 with forward models |
| Normandeau 2025 | hypothesis | Enters at O (coder 1) or Q (blind coders); the multi-year-ice assumption is P for two coders, R for Sonnet 5.5 |
| Vine 1963 | hypothesis with simulation | Questions 4–5 loop twice through block models |
| Rodell 2009 | estimation three times | Question 2 dominates; rainfall and soil water excluded |
| Lam 2023 | simulation and robustness | The forecast branch: verify, vary, verify |
| Gräff 2025 | theory | New instrument, exploration; the melt claim stops before question 7 |
| Xu 2025 | observation–methods, hypothesis–theory | Rivals listed, none excluded |
Mode sequences from coder 1's coded steps; Normandeau entry and P or R from the blind codings, private/analysis/reliability/codings/papers/
4 · The decision tree22 / 31
Automated science loops inside one mode
| System | Loop in node letters | What stays fixed |
|---|
| Strong inference (Platt 1964) | H D F O C X R → H | the auxiliaries |
| Robot Scientist (King 2004, 2009) | H S D F O C X → H | the model A |
| Adaptive design (Rainforth 2024) | S D F O I → S or D | H and A |
| Coscientist (Boiko 2023), optimization loop | D F O C R → D | the goal |
| Co-Scientist (Gottweis 2026) | H D S C X R → H, no O | no new data; literature only |
| Robin (Ghareeb 2026) | Q F H D, humans F O, I C K | humans run the lab |
The coders differ in detail: adaptive design returns to S (Opus 5.5) or D (Sonnet 5.5); Sonnet 5.5 codes Coscientist's optimization loop as H F O C → H (and its planner loop as D F O R → D); the Robot Scientist gets K from Sonnet 5.5 and, for the 2009 Adam loop, I and K from Opus 5.5; Opus 5.5 adds P and K to Co-Scientist. Only Co-Scientist revises an auxiliary (P), and only for Opus 5.5. None of the six systems asks question 1 (is there a phenomenon). None is reported to revise the rival set after a surprise. Human labs protect the method first, repeat, then revise H (Fugelsang et al. 2004).
Strong inference, Robin: literature report private/analysis/modes-literature/07-loops.md. Robot Scientist, adaptive design, Coscientist, Co-Scientist: two blind loop codings (Robot Scientist recoded from the 2004 and 2009 full texts), private/analysis/reliability/codings/agents/, Oct 25, 2026. DOIs checked on Crossref
4 · The decision tree23 / 31
How the tree could be wrong
- Same route, different outcomes: if ordinary and failed papers pass the same questions in the same order, the tree describes papers, not success.
- Missing questions: if review files or preprints show loops to a point the tree lacks, add it, as reframing was added at question 8 on Oct 21.
- No gain for agents: if tree-following agents do no better than agents choosing modes freely on held-out cases.
- Unobservable: if two coders agree on the question a step answers less often than κ = 0.6.
Each is a test the seminar can run
5 · Measuring success24 / 31
The success set: eight papers, two strata
| Paper | Journal | Citations | Percentile* |
|---|
| Vine & Matthews 1963 | Nature | 1,413 | 99.28 |
| Shapiro et al. 2005 | Science | 2,285 | 99.99 |
| Rodell et al. 2009 | Nature | 2,877 | 99.97 |
| Lam et al. 2023 | Science | 1,439 | 99.99 |
| Gräff et al. 2025 | Nature | 23 | 99.35 |
| Normandeau et al. 2025 | Science Advances | 1 | 57.03 |
| Sun & Davis 2026 | Science Advances | 2 | 97.15 |
| Xu et al. 2025 | Science | 25 | 99.78 |
Recognized (first four): chosen by a rule (Nature, Science, or AGU Advances; ≥ 99th percentile; ≥ 20% of citers or 25% of references outside geoscience, on the old measure). On breadth v1, Vine & Matthews fails the breadth clause (1% of citers, 0 of 10 references); Rodell passes on citers only (22%; 15% by hand), Shapiro on references only (25%). Designated (last four): chosen by Marine in 2025–26; impact is a prediction to check around 2028.
*OpenAlex field-normalized citation percentile, retrieved Oct 20, 2026. Breadth v1: private/analysis/success-set/breadth/REPORT.md, sections 3 and 5
5 · Measuring success25 / 31
Matched controls: the journal does most of the work
| Measure | Paper above its median control |
|---|
| Citations, percentile, field-weighted impact | 7 of 8 |
| References outside geoscience | 4 of 8 (5 on the old measure) |
| Citers outside geoscience | 4 of 7 |
| Disruption CD₅ | 4 of 7 |
- Controls: three papers per success, same journal, year, and subfield, at the median citation rank.
- A median Earth-science paper in Nature or Science already sits at the 98.4th–99.9th field-normalized percentile.
- Of 1,717 Nature and Science Earth-science research articles, 2005–2015, none is below its field median; the median is the 99.4th percentile.
Success set page; stratum D pools (OpenAlex, Oct 21, 2026); references on the corrected topic map (breadth/REPORT.md, section 5)
5 · Measuring success26 / 31
Breadth: the old measure filed geomagnetism under biology
- Defect: 12 geoscience topics sit under other fields. "Geomagnetism and Paleomagnetism Studies" is filed under molecular biology (its cluster includes magnetotactic bacteria): 307 works moved.
| Paper | Old measure | v1: corrected map, first 5 years, no self-citations | Checked by hand |
|---|
| Vine & Matthews 1963 | 33% | 1% (n = 104) | 4% |
| Rodell et al. 2009 | 24% | 22% (n = 451) | 15% |
| Shapiro et al. 2005 | 8% | 8% (n = 332) | 7% |
| Lam et al. 2023 | 38% | 34% (n = 1,422) | 27% |
Verdicts against the controls do not change: Rodell and Lam above, Shapiro below (the only three with 30 or more citers on both sides).
OpenAlex, Oct 21, 2026; hand check of 138 citers labelled outside (69% correct) and 169 labelled geoscience (96% correct); private/analysis/success-set/breadth/
5 · Measuring success27 / 31
Two new strata to test against
| Stratum | Rule | What it tests |
|---|
| C: highly cited, outside Nature and Science | Earth-science research articles, ≥ 99.9th percentile, four eras matched to the success set; reviews, software, data products excluded | Whether success needs the famous journal |
| D: Nature and Science, bottom decile | Research articles 2005–2015, bottom 10% of their journal-year pool, random draw | Whether the journal alone predicts the route through the tree |
C picks: Smagorinsky 1963, Macdonald & Katsura 1964, Bullard 1965, Jackson 2004, Tapley 2004, Bensen 2007, Gupta 2009, Chelton 2011, Rignot 2019, Wu 2019. D picks are "low" only against their journal (73–118 citations is the cut-off). Neither stratum holds a failure: failure means no conclusion (convenor, Oct 21), and every published paper reports one. A failure corpus must come from work that ended without one: abstracts never followed by a paper, funded projects without publications.
private/analysis/success-set/strata/REPORT.md; DOIs checked on Crossref
5 · Measuring success28 / 31
Highly cited is not broad
| Group | Median share of citers outside geoscience (breadth v1, lifetime) |
|---|
| Recognized four successes | about 12% (Lam 34%, Rodell 15–22%) |
| C: ten highly cited papers outside Nature and Science | 4% |
| Matched controls with 30 or more citers | 4% |
| D: ten low-cited Nature and Science papers | 7% |
Breadth of uptake does not follow citation success. Seven of the ten C papers contribute a method or an instrument (GCM, computed continental fit, LA-ICP-MS dating, GRACE, ambient-noise processing, KGE, CNN faults). Caveat: 366 unaudited topics default to "outside", which raises these shares.
OpenAlex, Oct 21, 2026; private/analysis/success-set/breadth/strata20_table.json; contribution labels from strata/REPORT.md
5 · Measuring success29 / 31
Data products and methods dominate the citation tail
| Most-cited Earth-science items | Citations |
|---|
| ERA5 reanalysis (2020) | 31,639 |
| ERA-Interim (2011) | 24,475 |
| IPCC AR4 WG1 (2007) | 23,330 |
| WorldClim (2005) | 19,845 |
| IntCal13 calibration curve | 14,592 |
| Smagorinsky 1963, the first research article in its era | 14,048 |
In 2019–2023 the first research article ranks 7th. Citations reward the modes that produce tools and records, which is consistent with Wren (2016) for bioinformatics.
OpenAlex, Oct 21, 2026
6 · For the room30 / 31
Decided by the convenor, October 21, 2026
| Question | Decision |
|---|
| Modes | Split synthesis (cross-study; reconstruction from traces); add question-driven inquiry |
| Graph | v0.4.2 is the vocabulary; the v0.5.1 tree is the controller |
| Reframing | After a failure to conclude, at question 8 |
| Success and failure | Failure = no conclusion; a contradicted claim is a success |
| Breadth | v1 adopted as a provisional measure |
| Coding | More papers, automated: codebook v1, schema, checker |
| Versioning | Controller, modes, and graph versioned in the repository with change logs |
notes page, "Modes meeting"; agents/controller/decision-tree.json
6 · For the room31 / 31
Still open for the room
- Protection: a numeric cap on protective revisions, beyond Lakatos's novel-prediction test?
- Order: is the aim (question 3) asked before the rivals (question 4)?
- Modes: does any other mode need splitting? Is methods and computation a mode or a contribution type?
- Failures: where does a corpus of projects that reached no conclusion come from (unpublished abstracts, funded projects without papers)?
- Your last paper: which stopping rule is wrong for it?
Each answer goes into the notes page with the source that forced it