Walking the graph: eight papers

Drafted graph v0.4 · steps coded from the full text of the success set · success score v0 · not yet discussed

Is the graph right? What these eight say

Test of graph v0.4 Result on the eight papers
Every move can be placed on a node Yes: no coder needed a node outside v0.4
Transitions follow edges the graph draws 24 of 83 transitions between the 91 steps (29%); the rest are jumps
The textbook route Q → H → S → D → F → O → C is walked Never
D, a discriminating test designed before the data 1 of 8 papers
H, stated rival hypotheses 5 of 8 papers
Entry at a node other than Q or H 7 of 8 papers
  • The node set holds; the edges do not. Most science moves between nodes in ways the graph draws only as long detours.
  • The most frequent jumps: V ⇢ C in 4 papers; S ⇢ C in 4 papers; C ⇢ P in 3 papers; C ⇢ K in 3 papers; X ⇢ V in 2 papers. S ⇢ C is a derivation compared with data already in hand; C ⇢ X and C ⇢ K are verdicts without a control series, which found conditions force.
  • Next tests. The same coding on the 24 matched controls and on unsuccessful papers. If ordinary and failed papers walk the same routes, the route does not explain success; if they skip different nodes, the graph has found something.
  • Candidate changes for v0.5: draw S → C and C → X, K as edges; or make the graph two-level, modes as the outer graph and the hypothesis pass as one mode’s inner loop. The room decides.

The success score (v0)

  • Four dimensions, each scored from 0 to 1 against comparable work, then combined. No measure here is settled; the score is a scaffold the group replaces piece by piece.
Dimension v0 measure Compared against Status
Impact Citation percentile All research articles in the same journal, year, and OpenAlex subfield Measured
Breadth (polymathy) Mean of two ranks: share of citers outside geoscience; share of references outside geoscience The three matched controls Measured, coarse
Novelty Rank of CD₅ disruption The three matched controls Measured, but CD₅ is a poor proxy; to be replaced by the group’s rating on the Novelty Dossier
Advance The group’s blinded rating on the advance profile, rescaled to 0–1 Not yet rated Pending
  • Macro score. The geometric mean of the measured dimensions, each floored at 0.05. A geometric mean does not let a high score on one dimension buy back a near-zero on another, which is what “success on all four” means (OECD et al. 2008). The arithmetic mean is reported beside it as a sensitivity check.
  • What the v0 scores can and cannot say. The recognized four score high because they were chosen on citations. The designated four have 1 to 25 citers each, so their breadth and novelty ranks are noise. The weak novelty proxy sinks Gräff and Xu.
Paper Impact Breadth Novelty (proxy) Advance Macro, geometric Macro, arithmetic
Shapiro et al. (2005) 0.97 0.67 1.00 pending 0.86 0.88
Sun and Davis (2026) 0.86 0.00 n/a pending 0.21 0.43
Normandeau et al. (2025) 0.07 0.33 1.00 pending 0.29 0.47
Vine and Matthews (1963) 0.99 1.00 0.67 pending 0.87 0.89
Rodell et al. (2009) 0.95 0.83 1.00 pending 0.93 0.93
Lam et al. (2023) 0.95 1.00 1.00 pending 0.98 0.98
Gräff et al. (2025) 0.57 0.83 0.00 pending 0.29 0.47
Xu et al. (2025) 0.72 0.17 0.00 pending 0.18 0.30

Three contrasting walks

Shapiro et al. (2005): noise becomes a seismometer

  • Start. A new capability: Green functions recovered from ambient noise, plus the new USArray stations (p. 1615). Enters the graph at N (New capability). (Shapiro et al. 2005)
  • End. K, a capability claim: noise correlations carry shallow and middle crustal structure, mapped at 60–100 km resolution (p. 1617).
%%{init: {"theme": "base", "themeVariables": {"fontSize": "20px", "fontFamily": "Source Serif 4, Georgia, serif", "lineColor": "#b9c3c6"}, "flowchart": {"htmlLabels": true, "nodeSpacing": 14, "rankSpacing": 22, "curve": "basis", "padding": 6}}}%%
flowchart LR
    E["<b>E</b>"] --> Q["<b>Q</b>"]
    Q --> H["<b>H</b>"]
    H --> S["<b>S</b>"]
    S --> D["<b>D</b>"]
    D --> F["<b>F</b>"]
    F --> O["<b>O</b>"]
    O --> I["<b>I</b>"]
    O --> C{"<b>C</b>"}
    I --> C
    C --> V["<b>V</b>"]
    V --> F
    V --> X["<b>X</b>"]
    V --> K["<b>K</b>"]
    X -.-> P["<b>P</b>"]
    P -.-> S
    X --> R["<b>R</b>"]
    K --> R
    R --> H
    C -.-> E
    O -.-> E
    N["<b>N</b>"] --> O
    E -.-> N
    classDef on fill:#cfe3e8,stroke:#1c6a7f,stroke-width:2px,color:#0f4c5c;
    classDef entry fill:#f6e0d4,stroke:#a1512b,stroke-width:4px,color:#5a2e17;
    classDef off fill:#f3f3f1,stroke:#d3d9db,stroke-width:1px,color:#b3bcbf;
    class C,E,F,I,K,O,V on;
    class N entry;
    class D,H,P,Q,R,S,X off;
Figure 1: Graph v0.4 as visited by Shapiro et al. (2005). Rust: entry node N. Teal: visited. Grey: never used (D, H, P, Q, R, S, X).
1NNew capability→2NNew capability⇢3EExplore⇢4VVary and replicate→5FFind conditions→6OObserve→7IInfer⇢8VVary and replicate⇢9CCompare⇢10KKeep
# Node Mode What the authors did Page
1 N New capability Methods Correlating a random wavefield gives the Green function between two receivers, after acoustics and helioseismology 1615
2 N New capability Methods One-bit correlation to suppress anisotropic noise and earthquake signals 1616
3 E Explore Robustness Four separate months of 2002 give similar waveforms, distinct between station pairs 1616
4 V Vary and replicate Robustness Noise correlations match earthquake Rayleigh waves on the same paths; one month suffices 1616
5 F Find conditions Observation Thirty quiet days (no event above M5.8), August–September 2004, 62 stations 1616
6 O Observe Observation 1,891 station-pair paths at one sample per second 1616
7 I Infer Estimation Group speeds by frequency–time analysis (678 and 891 at 7.5 and 15 s), inverted on a 28 km grid 1616
8 V Vary and replicate Replication A second month gives similar maps 1617
9 C Compare Observation Slow basins, fast plutonic ranges, the San Andreas contrast, against the geological map 1617
10 K Keep Methods Noise correlations retained as a tomographic capability 1617
  • Success profile (v0): impact 0.97 · breadth 0.67 · novelty proxy 1.00 · advance pending · macro 0.86.

Sun and Davis (2026): boreholes feel the megathrust lock

  • Start. An unsolved problem: offshore secular strain is hard to measure (p. 1), met with borehole observatories installed on drilling expeditions between 1995 and 2010 (p. 8). Enters the graph at N (New capability). (Sun and Davis 2026)
  • End. K then R: strain rates of 0.097–0.686 ×10⁻⁶/yr and full shallow locking at northern Cascadia (pp. 4, 7–8); Barbados left unmodeled (p. 8).
%%{init: {"theme": "base", "themeVariables": {"fontSize": "20px", "fontFamily": "Source Serif 4, Georgia, serif", "lineColor": "#b9c3c6"}, "flowchart": {"htmlLabels": true, "nodeSpacing": 14, "rankSpacing": 22, "curve": "basis", "padding": 6}}}%%
flowchart LR
    E["<b>E</b>"] --> Q["<b>Q</b>"]
    Q --> H["<b>H</b>"]
    H --> S["<b>S</b>"]
    S --> D["<b>D</b>"]
    D --> F["<b>F</b>"]
    F --> O["<b>O</b>"]
    O --> I["<b>I</b>"]
    O --> C{"<b>C</b>"}
    I --> C
    C --> V["<b>V</b>"]
    V --> F
    V --> X["<b>X</b>"]
    V --> K["<b>K</b>"]
    X -.-> P["<b>P</b>"]
    P -.-> S
    X --> R["<b>R</b>"]
    K --> R
    R --> H
    C -.-> E
    O -.-> E
    N["<b>N</b>"] --> O
    E -.-> N
    classDef on fill:#cfe3e8,stroke:#1c6a7f,stroke-width:2px,color:#0f4c5c;
    classDef entry fill:#f6e0d4,stroke:#a1512b,stroke-width:4px,color:#5a2e17;
    classDef off fill:#f3f3f1,stroke:#d3d9db,stroke-width:1px,color:#b3bcbf;
    class C,E,H,I,K,O,Q,R,S,V,X on;
    class N entry;
    class D,F,P off;
Figure 2: Graph v0.4 as visited by Sun and Davis (2026). Rust: entry node N. Teal: visited. Grey: never used (D, F, P).
1NNew capability→2OObserve→3EExplore⇢4HHypothesize⇢5XExclude⇢6IInfer⇢7QQuestion⇢8SSimulate or derive⇢9VVary and replicate⇢10CCompare⇢11XExclude⇢12KKeep→13RRefine
# Node Mode What the authors did Page
1 N New capability Instruments Borehole observatories with quartz pressure sensors; drift from sensor creep and aging characterized 1–3, 9
2 O Observe Observation Multiyear formation and seafloor pressure at four holes on four margins 2
3 E Explore Observation All four show “rising secular Pf trends greater than instrumental drift” 4
4 H Hypothesize Hypothesis Rivals to tectonic strain: drift, consolidation, compaction, post-drilling recovery, leakage 3–5, 8
5 X Exclude Robustness Drift excluded by a seafloor reference, an accidental open-ocean test, and calibrations; drainage too slow 3–5, 9
6 I Infer Estimation Pressure rate to strain rate by undrained poroelasticity, 4.21–6.79 kPa per microstrain 3, 10
7 Q Question Theory Which along-dip locking gives the observed sign and rate? 5–6
8 S Simulate or derive Simulation Finite-element backslip models: at least 90% locking at the trench needed for contraction 5
9 V Vary and replicate Robustness Rigidity, elastic mantle, weak asthenosphere, and locked width varied 5–6
10 C Compare Theory Cascadia models give 0.08–0.15 against 0.146 ×10⁻⁶/yr observed 7
11 X Exclude Hypothesis Ten percent or more shallow creep “would result in dilatational strain … counter to the observation” 7
12 K Keep Estimation Strain rates; full shallow locking at northern Cascadia, inter-slow-slip locking at Costa Rica 7–8
13 R Refine Synthesis Barbados left unmodeled; post-drilling recovery not excluded there 8
  • Success profile (v0): impact 0.86 · breadth 0.00 · novelty proxy n/a · advance pending · macro 0.21.

Normandeau et al. (2025): the seafloor freezes and thaws

  • Start. An anomaly between repeat surveys: hundreds of small mounds in July 2023, absent shallower than 25 m in 2021 (p. 4). Enters the graph at O (Observe). (Normandeau et al. 2025)
  • End. K, best explanation by analogy: seasonal freeze–thaw of freshened porewater expands and contracts the seafloor (pp. 1, 9–11).
%%{init: {"theme": "base", "themeVariables": {"fontSize": "20px", "fontFamily": "Source Serif 4, Georgia, serif", "lineColor": "#b9c3c6"}, "flowchart": {"htmlLabels": true, "nodeSpacing": 14, "rankSpacing": 22, "curve": "basis", "padding": 6}}}%%
flowchart LR
    E["<b>E</b>"] --> Q["<b>Q</b>"]
    Q --> H["<b>H</b>"]
    H --> S["<b>S</b>"]
    S --> D["<b>D</b>"]
    D --> F["<b>F</b>"]
    F --> O["<b>O</b>"]
    O --> I["<b>I</b>"]
    O --> C{"<b>C</b>"}
    I --> C
    C --> V["<b>V</b>"]
    V --> F
    V --> X["<b>X</b>"]
    V --> K["<b>K</b>"]
    X -.-> P["<b>P</b>"]
    P -.-> S
    X --> R["<b>R</b>"]
    K --> R
    R --> H
    C -.-> E
    O -.-> E
    N["<b>N</b>"] --> O
    E -.-> N
    classDef on fill:#cfe3e8,stroke:#1c6a7f,stroke-width:2px,color:#0f4c5c;
    classDef entry fill:#f6e0d4,stroke:#a1512b,stroke-width:4px,color:#5a2e17;
    classDef off fill:#f3f3f1,stroke:#d3d9db,stroke-width:1px,color:#b3bcbf;
    class C,E,H,I,K,P,S on;
    class O entry;
    class D,F,N,Q,R,V,X off;
Figure 3: Graph v0.4 as visited by Normandeau et al. (2025). Rust: entry node O. Teal: visited. Grey: never used (D, F, N, Q, R, V, X).
1OObserve→2EExplore⇢3HHypothesize⇢4IInfer⇢5SSimulate or derive⇢6CCompare⇢7PProtect⇢8KKeep
# Node Mode What the authors did Page
1 O Observe Observation Multibeam, backscatter, subbottom profiles, 13 cores, porewater, moorings; surveys 2021, July 2023, November 2023, August 2024 3–5, 12
2 E Explore Observation Blisters about 5 m wide present in July 2023, “greatly diminished” in November, back in August 2024 4
3 H Hypothesize Hypothesis Pockmarks from groundwater, gas, or burrowing; frost blisters against all-ice icing blisters 3–11
4 I Infer Estimation Water temperature at 30 m inferred from a 60 m mooring after the 30 m mooring was lost 6, 12
5 S Simulate or derive Theory Stefan equation: freezing depth 0.32–0.70 m 10
6 C Compare Theory Freezing depth matches the frost base at about 40 cm; ice expansion gives at most 0.05 m against 0.2 m of relief 10
7 P Protect Hypothesis “The underlying ground ice that forms the blisters may exist over multiple years” 10
8 K Keep Synthesis Seasonal freeze–thaw retained, by analogy with frost blisters on land 9–11
  • Success profile (v0): impact 0.07 · breadth 0.33 · novelty proxy 1.00 · advance pending · macro 0.29.

Five more walks

Vine and Matthews (1963): reversed blocks and spreading

  • Start. An unsolved problem: magnetic anomalies over ridges that no model of uniform magnetization explains (p. 947). Enters the graph at E (Explore). (Vine and Matthews 1963)
  • End. K, best explanation: reversed blocks from spreading under a reversing field, “more plausible” than the alternative (p. 949).
%%{init: {"theme": "base", "themeVariables": {"fontSize": "20px", "fontFamily": "Source Serif 4, Georgia, serif", "lineColor": "#b9c3c6"}, "flowchart": {"htmlLabels": true, "nodeSpacing": 14, "rankSpacing": 22, "curve": "basis", "padding": 6}}}%%
flowchart LR
    E["<b>E</b>"] --> Q["<b>Q</b>"]
    Q --> H["<b>H</b>"]
    H --> S["<b>S</b>"]
    S --> D["<b>D</b>"]
    D --> F["<b>F</b>"]
    F --> O["<b>O</b>"]
    O --> I["<b>I</b>"]
    O --> C{"<b>C</b>"}
    I --> C
    C --> V["<b>V</b>"]
    V --> F
    V --> X["<b>X</b>"]
    V --> K["<b>K</b>"]
    X -.-> P["<b>P</b>"]
    P -.-> S
    X --> R["<b>R</b>"]
    K --> R
    R --> H
    C -.-> E
    O -.-> E
    N["<b>N</b>"] --> O
    E -.-> N
    classDef on fill:#cfe3e8,stroke:#1c6a7f,stroke-width:2px,color:#0f4c5c;
    classDef entry fill:#f6e0d4,stroke:#a1512b,stroke-width:4px,color:#5a2e17;
    classDef off fill:#f3f3f1,stroke:#d3d9db,stroke-width:1px,color:#b3bcbf;
    class C,H,I,K,O,P,Q,S,V,X on;
    class E entry;
    class D,F,N,R off;
Figure 4: Graph v0.4 as visited by Vine and Matthews (1963). Rust: entry node E. Teal: visited. Grey: never used (D, F, N, R).
1EExplore→2QQuestion⇢3OObserve→4IInfer→5CCompare⇢6PProtect→7SSimulate or derive⇢8XExclude⇢9HHypothesize⇢10KKeep⇢11HHypothesize→12SSimulate or derive⇢13CCompare→14VVary and replicate⇢15PProtect⇢16XExclude⇢17KKeep
# Node Mode What the authors did Page
1 E Explore Observation A three-category anomaly pattern over ridges in three oceans, from earlier surveys 947
2 Q Question Observation Why do the central and long-period anomalies exist? 947
3 O Observe Observation H.M.S. Owen survey of the Carlsberg Ridge, November 1962 947
4 I Infer Estimation Computer best-fit magnetization of two seamounts 947
5 C Compare Estimation The reversed case fits well; the normal case poorly 947
6 P Protect Hypothesis The poor normal fit blamed on adjacent topography and weak contour control 947
7 S Simulate or derive Simulation Forward profiles across three cross-ridge lines for uniform normal magnetization 948
8 X Exclude Hypothesis The uniform model bears “little resemblance to the observed profiles” 948
9 H Hypothesize Theory Crust in blocks about 20 km wide, alternately normal and reversed 948
10 K Keep Theory “Some 50 per cent of the oceanic crust might be reversely magnetized” 948
11 H Hypothesize Synthesis The model as “virtually a corollary of” seafloor spreading plus field reversals 948
12 S Simulate or derive Simulation Forward models for the Pacific, the Mid-Atlantic, and the Carlsberg Ridge 948
13 C Compare Simulation “The essential form of the anomalies is readily achieved” 949
14 V Vary and replicate Synthesis Independent traces: Iceland lavas, reversed basalt in the preliminary Mohole 949
15 P Protect Hypothesis Auxiliaries defended: susceptibility contrast, Curie depth “not critical” 949
16 X Exclude Hypothesis The same-direction rival set aside on plausibility, not by a test 949
17 K Keep Hypothesis Reversed blocks retained as “more plausible” 949
  • Success profile (v0): impact 0.99 · breadth 1.00 · novelty proxy 0.67 · advance pending · macro 0.87.

Rodell et al. (2009): groundwater loss seen from space

  • Start. An unsolved problem: “there has been no regional assessment of the rate of groundwater depletion” in northwest India (p. 999). Enters the graph at Q (Question). (Rodell et al. 2009)
  • End. K, an estimate with an explanation: 4.0 ± 1.0 cm/yr, 109 km³ lost, human use the likely cause (p. 1001); then a comparison with the ministry figure.
%%{init: {"theme": "base", "themeVariables": {"fontSize": "20px", "fontFamily": "Source Serif 4, Georgia, serif", "lineColor": "#b9c3c6"}, "flowchart": {"htmlLabels": true, "nodeSpacing": 14, "rankSpacing": 22, "curve": "basis", "padding": 6}}}%%
flowchart LR
    E["<b>E</b>"] --> Q["<b>Q</b>"]
    Q --> H["<b>H</b>"]
    H --> S["<b>S</b>"]
    S --> D["<b>D</b>"]
    D --> F["<b>F</b>"]
    F --> O["<b>O</b>"]
    O --> I["<b>I</b>"]
    O --> C{"<b>C</b>"}
    I --> C
    C --> V["<b>V</b>"]
    V --> F
    V --> X["<b>X</b>"]
    V --> K["<b>K</b>"]
    X -.-> P["<b>P</b>"]
    P -.-> S
    X --> R["<b>R</b>"]
    K --> R
    R --> H
    C -.-> E
    O -.-> E
    N["<b>N</b>"] --> O
    E -.-> N
    classDef on fill:#cfe3e8,stroke:#1c6a7f,stroke-width:2px,color:#0f4c5c;
    classDef entry fill:#f6e0d4,stroke:#a1512b,stroke-width:4px,color:#5a2e17;
    classDef off fill:#f3f3f1,stroke:#d3d9db,stroke-width:1px,color:#b3bcbf;
    class C,H,I,K,N,O,S,V,X on;
    class Q entry;
    class D,E,F,P,R off;
Figure 5: Graph v0.4 as visited by Rodell et al. (2009). Rust: entry node Q. Teal: visited. Grey: never used (D, E, F, P, R).
1QQuestion⇢2NNew capability→3OObserve→4IInfer⇢5HHypothesize→6SSimulate or derive⇢7XExclude⇢8VVary and replicate→9KKeep⇢10CCompare
# Node Mode What the authors did Page
1 Q Question Estimation Rate of groundwater depletion over Rajasthan, Punjab, and Haryana with Delhi, August 2002 to October 2008 999
2 N New capability Instruments 73 monthly GRACE solutions, destriped; smoothing radius tested; scaling 1.95 from a synthetic signal 1001, M5
3 O Observe Observation GRACE water-storage series for the region 1000
4 I Infer Estimation Groundwater as storage minus the mean of five modeled soil-water runs; error budget 1.0 cm/yr 1001
5 H Hypothesize Hypothesis Rivals: soil moisture, surface water, snow, glaciers, biomass; natural variability against human use 1000–1001
6 S Simulate or derive Simulation Expected leakage from Himalayan melt and from surrounding trends 1000
7 X Exclude Hypothesis Glaciers at most 15%, reservoirs, paddies, snow excluded; soil water 0.4 cm/yr; rainfall near normal 1000–1001
8 V Vary and replicate Robustness Seasonal lags match earlier studies; smoothing radius makes insignificant differences 1001, M5
9 K Keep Estimation 4.0 ± 1.0 cm/yr retained; human consumption the likely cause 1001
10 C Compare Synthesis Larger than the ministry deficit of 13.2 km³/yr 1001
  • Success profile (v0): impact 0.95 · breadth 0.83 · novelty proxy 1.00 · advance pending · macro 0.93.

Lam et al. (2023): learned weather forecasts

  • Start. A new capability: machine learning, 39 years of reanalysis, and compute; a benchmark set by others (p. 1). Enters the graph at N (New capability). (Lam et al. 2023)
  • End. K, forecast skill: better than the operational system on 90% of 1,380 targets, deterministic, 10 days, 0.25° (pp. 4–6).
%%{init: {"theme": "base", "themeVariables": {"fontSize": "20px", "fontFamily": "Source Serif 4, Georgia, serif", "lineColor": "#b9c3c6"}, "flowchart": {"htmlLabels": true, "nodeSpacing": 14, "rankSpacing": 22, "curve": "basis", "padding": 6}}}%%
flowchart LR
    E["<b>E</b>"] --> Q["<b>Q</b>"]
    Q --> H["<b>H</b>"]
    H --> S["<b>S</b>"]
    S --> D["<b>D</b>"]
    D --> F["<b>F</b>"]
    F --> O["<b>O</b>"]
    O --> I["<b>I</b>"]
    O --> C{"<b>C</b>"}
    I --> C
    C --> V["<b>V</b>"]
    V --> F
    V --> X["<b>X</b>"]
    V --> K["<b>K</b>"]
    X -.-> P["<b>P</b>"]
    P -.-> S
    X --> R["<b>R</b>"]
    K --> R
    R --> H
    C -.-> E
    O -.-> E
    N["<b>N</b>"] --> O
    E -.-> N
    classDef on fill:#cfe3e8,stroke:#1c6a7f,stroke-width:2px,color:#0f4c5c;
    classDef entry fill:#f6e0d4,stroke:#a1512b,stroke-width:4px,color:#5a2e17;
    classDef off fill:#f3f3f1,stroke:#d3d9db,stroke-width:1px,color:#b3bcbf;
    class C,D,K,O,R,S,V,X on;
    class N entry;
    class E,F,H,I,P,Q off;
Figure 6: Graph v0.4 as visited by Lam et al. (2023). Rust: entry node N. Teal: visited. Grey: never used (E, F, H, I, P, Q).
1NNew capability⇢2DDesign⇢3SSimulate or derive⇢4OObserve→5CCompare⇢6XExclude⇢7VVary and replicate⇢8CCompare⇢9RRefine⇢10KKeep
# Node Mode What the authors did Page
1 N New capability Methods Graph neural network on an icosahedral multimesh, 36.7 million parameters, trained on ERA5 1979–2017 2–3
2 D Design Robustness Verification designed first: 69 variable-levels; truth and initial times matched to the operational system 3
3 S Simulate or derive Simulation Ten-day rollouts from held-out initial states, 2018 onward 2–3
4 O Observe Simulation Forecasts scored against verifying analyses: RMSE and anomaly correlation 4
5 C Compare Simulation Better than the operational forecast on 90.3% of 1,380 targets 4
6 X Exclude Hypothesis Rival: the gain is only blurring. Blur the operational forecast too: still better on 88.0% 4
7 V Vary and replicate Robustness Regions, level exclusions, a second machine-learning baseline (99.2% of 252 targets) 4
8 C Compare Simulation Cyclone tracks, atmospheric rivers, and extreme heat, none trained for 5
9 R Refine Robustness Retrained on data ending 2017–2020, scored on 2021: more recent data helps 5
10 K Keep Simulation Deterministic medium-range skill retained within the stated scope 5–6
  • Success profile (v0): impact 0.95 · breadth 1.00 · novelty proxy 1.00 · advance pending · macro 0.98.

Gräff et al. (2025): a fibre along the calving front

  • Start. A new capability: seafloor fibre sensing “not yet used to study ice–ocean interactions” (p. 1). Enters the graph at N (New capability). (Gräff et al. 2025)
  • End. K, a “potential” mechanism: calving amplifies submarine melt, to be “investigated further” (p. 8).
%%{init: {"theme": "base", "themeVariables": {"fontSize": "20px", "fontFamily": "Source Serif 4, Georgia, serif", "lineColor": "#b9c3c6"}, "flowchart": {"htmlLabels": true, "nodeSpacing": 14, "rankSpacing": 22, "curve": "basis", "padding": 6}}}%%
flowchart LR
    E["<b>E</b>"] --> Q["<b>Q</b>"]
    Q --> H["<b>H</b>"]
    H --> S["<b>S</b>"]
    S --> D["<b>D</b>"]
    D --> F["<b>F</b>"]
    F --> O["<b>O</b>"]
    O --> I["<b>I</b>"]
    O --> C{"<b>C</b>"}
    I --> C
    C --> V["<b>V</b>"]
    V --> F
    V --> X["<b>X</b>"]
    V --> K["<b>K</b>"]
    X -.-> P["<b>P</b>"]
    P -.-> S
    X --> R["<b>R</b>"]
    K --> R
    R --> H
    C -.-> E
    O -.-> E
    N["<b>N</b>"] --> O
    E -.-> N
    classDef on fill:#cfe3e8,stroke:#1c6a7f,stroke-width:2px,color:#0f4c5c;
    classDef entry fill:#f6e0d4,stroke:#a1512b,stroke-width:4px,color:#5a2e17;
    classDef off fill:#f3f3f1,stroke:#d3d9db,stroke-width:1px,color:#b3bcbf;
    class C,E,F,I,K,O,S on;
    class N entry;
    class D,H,P,Q,R,V,X off;
Figure 7: Graph v0.4 as visited by Gräff et al. (2025). Rust: entry node N. Teal: visited. Grey: never used (D, H, P, Q, R, V, X).
1NNew capability⇢2FFind conditions→3OObserve→4EExplore⇢5IInfer⇢6SSimulate or derive⇢7CCompare⇢8KKeep
# Node Mode What the authors did Page
1 N New capability Instruments A 10 km cable near the front carrying acoustic and temperature sensing; calibrated; its position inverted by MCMC 2, 10–12
2 F Find conditions Observation Found conditions: one glacier, 9–29 August 2023, a cable layout parallel to the front 2, 8
3 O Observe Observation 21 days, 13 TB, alongside a radar interferometer, a camera, seismometers, 22 CTD casts, a tide gauge 2, 10
4 E Explore Exploration A process chain: about 56,000 detachments, several hundred tsunamis, internal waves, iceberg wakes, currents 3–7
5 I Infer Estimation Seafloor wave speed 230 ± 40 m/s; tsunami source 3 ± 1 km away; detection threshold about 100 m³ 3–5, 12–14
6 S Simulate or derive Theory A three-equation melt model; internal-wave dispersion; the Turner angle 5, 12–13
7 C Compare Synthesis Dispersion, tsunami origin, and wave periods match the radar, CTD, and theory 4, 12–13
8 K Keep Theory Calving as a melt amplifier, retained as a potential mechanism 8
  • Success profile (v0): impact 0.57 · breadth 0.83 · novelty proxy 0.00 · advance pending · macro 0.29.

Xu et al. (2025): a supershear rupture in Myanmar

  • Start. Data on a new event: global records of the 2025 Mandalay earthquake, imaged to obtain “timely” rupture details (p. 476). Enters the graph at O (Observe). (Xu et al. 2025)
  • End. K, a conjunction of favorable conditions as best explanation for 450 km of supershear rupture (pp. 476, 480).
%%{init: {"theme": "base", "themeVariables": {"fontSize": "20px", "fontFamily": "Source Serif 4, Georgia, serif", "lineColor": "#b9c3c6"}, "flowchart": {"htmlLabels": true, "nodeSpacing": 14, "rankSpacing": 22, "curve": "basis", "padding": 6}}}%%
flowchart LR
    E["<b>E</b>"] --> Q["<b>Q</b>"]
    Q --> H["<b>H</b>"]
    H --> S["<b>S</b>"]
    S --> D["<b>D</b>"]
    D --> F["<b>F</b>"]
    F --> O["<b>O</b>"]
    O --> I["<b>I</b>"]
    O --> C{"<b>C</b>"}
    I --> C
    C --> V["<b>V</b>"]
    V --> F
    V --> X["<b>X</b>"]
    V --> K["<b>K</b>"]
    X -.-> P["<b>P</b>"]
    P -.-> S
    X --> R["<b>R</b>"]
    K --> R
    R --> H
    C -.-> E
    O -.-> E
    N["<b>N</b>"] --> O
    E -.-> N
    classDef on fill:#cfe3e8,stroke:#1c6a7f,stroke-width:2px,color:#0f4c5c;
    classDef entry fill:#f6e0d4,stroke:#a1512b,stroke-width:4px,color:#5a2e17;
    classDef off fill:#f3f3f1,stroke:#d3d9db,stroke-width:1px,color:#b3bcbf;
    class C,E,F,H,I,K,N,P,Q,R,S,V on;
    class O entry;
    class D,X off;
Figure 8: Graph v0.4 as visited by Xu et al. (2025). Rust: entry node O. Teal: visited. Grey: never used (D, X).
1OObserve→2IInfer⇢3EExplore→4NNew capability⇢5VVary and replicate⇢6CCompare⇢7KKeep⇢8QQuestion→9HHypothesize⇢10FFind conditions⇢11SSimulate or derive⇢12CCompare⇢13PProtect⇢14RRefine⇢15KKeep
# Node Mode What the authors did Page
1 O Observe Observation Three teleseismic arrays, satellite radar, optical imagery, strong motion 476
2 I Infer Methods Back-projection of high-frequency radiators; joint finite-fault inversion 476–477
3 E Explore Observation Three arrays agree on 4.8–5.0 km/s southward, above the crustal shear speed of 2.5–3.7 km/s 476
4 N New capability Methods A synthetic back-projection recovers the southern speed 8–20% low 476
5 V Vary and replicate Robustness Nine re-inversions from different starting models 477
6 C Compare Hypothesis Mach waves give 3.4–5.8 km/s; fault-parallel motion and onset timing agree 477–478
7 K Keep Observation Southern supershear rupture retained 478
8 Q Question Theory Why an early transition, and why sustained? 478
9 H Hypothesize Hypothesis Eight transition mechanisms from the literature, listed as possible contributors 478
10 F Find conditions Observation Found conditions: shear speed 2.9 west against 3.3 km/s east; quiet since 1839; a smooth fault 479–480
11 S Simulate or derive Theory Expectations imported from published theory and laboratory work: the bimaterial direction favors supershear 479–480
12 C Compare Theory Direction consistent; the shorter northern branch contradicts the simulations 479
13 P Protect Hypothesis The northern length explained by five earlier M ≥ 6.8 earthquakes 479
14 R Refine Theory Poroelastic contrast, stress, and the free surface added 479
15 K Keep Synthesis A conjunction of favorable conditions retained 480
  • Success profile (v0): impact 0.72 · breadth 0.17 · novelty proxy 0.00 · advance pending · macro 0.18.

What the walks cannot show

  • Loops are under-counted by design. A published paper reports the work in the order of its argument: between protocol and paper, 62% of trials changed, added, or dropped a primary outcome (Chan et al. 2004). Our coding followed the paper’s order and had no field for a failed attempt, a rerun, or a redesign, so a return and a next step look alike.
  • Loops are at the mode level, for coder 1. Merging consecutive repeats, the papers return to a node already visited 11 times but to a mode already used 27 times; every paper has at least one return. The decision tree draws the graph at that level.
  • The blind coders do not confirm it. Two blind model coders (Opus 5.5 and Sonnet 5.5, paper codebook 1.3.0, October 25, 2026) code 169 and 108 steps on the same eight papers. They find 83 and 33 node returns against 78 and 38 mode returns, so where the loops sit depends on the coder (private/analysis/reliability/codings/papers/, final-paper sequences, consecutive repeats merged).
  • One coder per walk, coding against the graph it tests. The coder was given the v0.4 nodes, so every step was placed on a node; “no new node needed” is partly an effect of the instruction. The blind coders also get the graph. They agree on which nodes a paper uses at Krippendorff’s alpha 0.63, below the 0.67 gate, over 13 papers (private/analysis/reliability/RESULTS.md). No person has coded yet.
  • Granularity varies. Vine and Matthews have 17 steps, Normandeau and colleagues and Gräff and colleagues 8. Step counts measure the coder as much as the paper.
  • Labels are not fully consistent. One Shapiro step is labelled replication where the others use robustness.
  • Supplements unread for Shapiro, Lam, Sun, and Xu, where methods and robustness checks usually sit.
  • Successes only, as walks. The eight walks are successes. Five matched controls have been coded since (October 21 and 25, 2026); no failure has been coded.
  • How to see the loops. Peer-review files, open discussions, and version histories record changes made during the work. On nine papers with such records, 126 loops were found, and 4 of them (3%) can be read as loops in the final paper (decision tree, section 6). For Gräff and colleagues and for Lam and colleagues, a review sequence can now be coded beside the final-paper sequence.

For students

  • Pick a paper you know. Write its start, its steps as a strip, and its end. For each step, name the node, the mode, and the page.
  • Mark the node where the paper entered and the nodes it never needed. Then ask what would have made it a stronger paper, and at which node.