J3 Methodology — Seven-Phase Claim-Driven Prototype¶

Date: 2026-04-18 Status: Second prototype of the seven-phase claim-driven workflow, applied to Paper J3 — the centrifuge saturation and backfill study, under Round 1 revision at Ocean Engineering.

The J2 prototype showed that for a numerical-modelling section, the workflow surfaces two specific missing figures and two one-sentence paragraph fixes. This note runs the same pass on a centrifuge experimental section. The claim structure is different — experiments have protocol fidelity and scaling-law claims that numerical work does not — and the diagnostic accordingly picks up a different class of defect.

Phase 1 — Thesis statement¶

One sentence the methodology section must prove:

The 70 g centrifuge programme on the UNISON U136 4.2 MW tripod model, extended from the dry-sand baseline of Paper J1, isolates the saturation and backfill effects through a structured five-series design (T1–T5) whose soil state is verified by in-flight CPT, whose scour protocol reproduces the dry-sand geometry exactly, and whose reported sensitivities are explicitly interpreted as a lower bound on true field sensitivity because of a documented stress-history artefact — so the saturation ratio of 1.94 and the density amplification of 3× are measurements, not artefacts of an uncontrolled experiment.

Phase 2 — Claim list¶

Six load-bearing claims decompose the thesis. Centrifuge work splits into more claims than a numerical-modelling section because protocol fidelity is its own evidence layer.

C1 — Scaling fidelity. Centrifuge model correctly reproduces the prototype structure: \(EI_{\text{eq}}\) matches within 2.3 %, dimensions scaled per Kutter-law, identical structural model as J1 so any dry-vs-saturated difference is not structural.
C2 — Soil-state verification. In-flight CPT confirms the intended density state per flight (\(D_r = 70.2\) % dense, \(60.8\) % loose), independent of the test label; saturation protocol achieves full pore-pressure equilibration before dynamic testing.
C3 — Scour protocol comparability. Saturated scour stages use the identical concentric excavation tool and 45° frustum geometry as the dry-sand J1 programme, so any saturation vs. dry comparison is about the soil state, not the scour geometry.
C4 — Stress-history artefact acknowledged and bounded. The 1 g excavation protocol produces OCR \(\approx 1.5\)–\(2.0\) in the near-scour zone, causing a 15–25 % stiffness bias that is directionally conservative (lower-bound sensitivity); the bias is explicitly accounted for in the comparison with J2's numerical power-law.
C5 — Instrumentation and dynamic testing. Triaxial accelerometers, strain gauges, pore-pressure transducers, displacement sensors at prescribed locations; impulse + sinusoidal + square-wave excitation per flight, 22–26 minutes at 5,000 Hz; OMA extracts \(f_1\) with MAC verification against FE mode shape.
C6 — Backfill protocol replicability. Backfill tests (T4-5, T5-5) use ≈ 20 kg of pre-saturated coarser No. 5 sand with specified tamping (≈ 10 applications per bucket); the recovered frequency is attributable to the backfill material properties rather than to redistribution of native soil.

Six claims, six evidence bundles.

Phase 3 — Claim → evidence map¶

flowchart TD
    T["<b>Thesis</b><br/>Saturation ratio 1.94 and density 3× amplification<br/>are measurements, not protocol artefacts"]
    T --> C1["C1 · Scaling fidelity"]
    T --> C2["C2 · Soil-state verification (CPT)"]
    T --> C3["C3 · Scour protocol matches J1 exactly"]
    T --> C4["C4 · Stress-history artefact bounded"]
    T --> C5["C5 · Instrumentation and dynamic testing"]
    T --> C6["C6 · Backfill protocol replicability"]
    C1 --> E1["Table · scaling laws at N=70<br/>Table · prototype vs. model EI comparison<br/>Stat · EI match within 2.3%"]
    C2 --> E2["Fig · CPT profiles per flight (T4-1, T5-1 baseline)<br/>Fig · pore-pressure time series to equilibration<br/>Stat · D<sub>r</sub> from Baldi 1986"]
    C3 --> E3["Fig · scour-hole schematic (frustum, 45° slope)<br/>Fig · plan view of three buckets with scour annulus<br/>Photo · excavation tool"]
    C4 --> E4["Text · OCR estimate derivation<br/>Fig · f<sub>1</sub> comparison centrifuge |a|=0.044-0.086 vs. numerical |a|=0.167<br/>Stat · 15-25% stiffness bias"]
    C5 --> E5["Fig · model configuration with sensor layout<br/>Table · sensor list with sampling rates<br/>Fig · MAC between measured and FE mode shape"]
    C6 --> E6["Fig · backfill before/after<br/>Fig · grain-size curves No. 5 vs. No. 7 sand<br/>Text · tamping protocol"]
    E1 --> S1["table · scaling-laws<br/>table · ei-matching"]
    E2 --> S2["spec · fig-cpt-profiles<br/>spec · fig-ppt-equilibration"]
    E3 --> S3["spec · fig-scour-schematic<br/>spec · fig-plan-view"]
    E4 --> S4["spec · fig-centrifuge-vs-numerical-powerlaw"]
    E5 --> S5["spec · fig-test-setup<br/>spec · fig-mac-mode-shape"]
    E6 --> S6["spec · fig-backfill-state<br/>spec · fig-grain-size"]

Eleven figure specs, two tables. The existing manuscript has roughly fifteen methodology-section figures — the mapping reveals one orphan (a raw time-series snapshot of a single flight that does not map to any claim) and one redundancy (two similar model-configuration photos).

Phase 4 — Figure specs¶

table-scaling-laws. Scaling factors at \(N = 70\) for length, mass, stress, strain, force, acceleration, time, frequency, displacement. Single-column, referenced to Kutter (1992) and Madabhushi (2014).

table-ei-matching. Prototype vs. model comparison: material, tower diameter, tower wall thickness, tower height, bucket diameter, skirt length, skirt wall thickness, bucket spacing, \(EI_{\text{eq}}\), RNA mass, total mass. Pass criterion: \(EI_{\text{eq}}\) within 3 %.

fig-cpt-profiles. Must show in-flight CPT tip resistance \(q_c(z)\) for baseline flights T4-1 (dense saturated, target \(D_r = 70\) %) and T5-1 (loose saturated, target \(D_r = 61\) %), plus the derived relative density via Baldi et al. (1986), overlaid. Twin-panel: (a) \(q_c(z)\) with measured points and fitted envelope; (b) \(D_r(z)\) with the target value marked as a horizontal line.

fig-ppt-equilibration. Must show pore-pressure transducer readings over the 15-minute spin-up, confirming pressures stabilise within ± 0.002 bar before dynamic testing begins. Three traces (three PPT locations), single panel, time on x-axis, pressure on y-axis.

fig-scour-schematic. Must show the frustum scour-hole geometry with 45° side slopes, inner floor radius equal to bucket outer radius (\(D/2 = 4.0\) m prototype), surface outer radius \(D/2 + S\), and the maximum test scour \(S/D = 0.58\) labelled. Two panels: (a) cross-section through a single bucket; (b) plan view of bucket with scour annulus.

fig-plan-view. Must show the tripod footprint as an equilateral triangle with side \(L_{\text{base}} = 20.0\) m prototype, three buckets at the vertices labelled A/B/C, and the shake direction aligned with A–B axis. Annotation confirms three-fold symmetry; B and C are mirror-symmetric about the A–B axis.

fig-centrifuge-vs-numerical-powerlaw. Must show the centrifuge power-law coefficient \(|a| = 0.044\)–\(0.086\) (range across T4 and T5 saturated series) on the same axes as J2's numerical \(|a| = 0.167\), with the OCR-induced stiffness bias band (15–25 %) shown as the shaded difference. Log-log axes; annotation identifies the stress-history artefact as the systematic offset.

fig-test-setup. Must show the model installed in the strongbox with every sensor labelled: three triaxial accelerometers at prescribed elevations, strain gauges on the tripod jacket, pore-pressure transducers at three depths, displacement sensors at the tower top. Single photograph plus overlaid schematic with call-outs.

fig-mac-mode-shape. Must show the modal assurance criterion (MAC) matrix between the measured first-bending mode (from OMA across T4 and T5 baseline flights) and the FE-predicted mode shape, with the diagonal MAC values ≥ 0.95 confirming modal identity. Heat-map style, one matrix.

fig-backfill-state. Must show the scour hole before backfill (maximum-scour condition, \(S/D = 0.58\)) and after backfill (with coarser No. 5 sand, tamped), in two photographs taken under consistent lighting. Captions describe the backfill mass (≈ 20 kg) and tamping protocol.

fig-grain-size. Must show the particle-size distribution curves of the native No. 7 sand (\(d_{50} = 0.22\) mm) and the backfill No. 5 sand (\(d_{50} = 1.99\) mm) on semi-log axes, with the \(d_{50}\) values annotated to highlight the coarseness contrast (factor of 9 in median diameter). Single panel.

Phase 5 — Paragraph skeletons¶

§2.1 Centrifuge Facility and Test Programme (proves C1 partially, introduces the test matrix). ¶1 (facility). 70 g beam centrifuge at KAIST, 5.0 m effective radius, 2,400 kg capacity, rectangular aluminium strongbox with Duxseal absorbing walls. ¶2 (test matrix). 22 cases across five series T1–T5; dry series T1–T3 from J1, saturated series T4–T5 are this paper; T1 ↔ T4 and T2 ↔ T5 isolate saturation at fixed density [→ table-test-matrix, referenced elsewhere]. ¶3 (flight cadence). Each flight is spin-up + PPT equilibration + in-flight CPT (baseline flights) + dynamic testing; 22–26 minutes at 5,000 Hz per flight; three working days per series. ¶4 (upper-bound justification). \(S/D \le 0.6\) is justified by monitoring-relevant band, transition-band inclusion, and skirt-tip preservation [narrative already present in manuscript — kept].

§2.2 Model Configuration and Instrumentation (proves C1 fully, sets up C5). ¶1 (scaling laws) [→ table-scaling-laws]. ¶2 (equivalent-section design). Prototype tower is tapered; model uses stainless-steel cylinders with wall thickness matched to \(EI_{\text{eq}}\) [→ table-ei-matching]. Identical structural model as J1 — so dry vs. saturated difference is not structural. ¶3 (instrumentation) [→ fig-test-setup]. ¶4 (MAC verification) [→ fig-mac-mode-shape]; first-bending MAC ≥ 0.95 across baseline flights confirms modal identity.

§2.3 Soil Preparation and Saturation (proves C2). ¶1 (soil placement). Dry-pluviation for dense; moist-tamping for loose; target density set by drop height and placement rate. ¶2 (saturation protocol). CO₂ flushing + de-aired water infiltration under vacuum until Skempton B ≥ 0.95; spin-up monitored via PPT until readings stable within ± 0.002 bar [→ fig-ppt-equilibration]. ¶3 (soil-state verification). In-flight CPT on baseline flights produces \(q_c(z)\) profile; \(D_r\) derived via Baldi et al. (1986) [→ fig-cpt-profiles]; reported \(D_r = 70.2\) % (dense) and \(60.8\) % (loose) are measured, not target values.

§2.4 Scour Protocol (proves C3 and C4). ¶1 (protocol fidelity). Scour stages applied at 1 g between spin-ups using the identical concentric excavation tool and frustum geometry as J1 [→ fig-scour-schematic]; tripod plan view shows scour holes do not intersect at max depth [→ fig-plan-view]. ¶2 (local vs. global scour). All scour is local (concentric around each bucket, inter-bucket seabed undisturbed); global scour not simulated, rationale explicitly stated. ¶3 (stress-history artefact). Excavation at 1 g produces OCR \(\approx 1.5\)–\(2.0\) in the near-scour zone; estimated 15–25 % stiffness bias; directionally conservative (lower-bound sensitivity); consistent with the centrifuge power-law coefficient being smaller than J2's numerical coefficient [→ fig-centrifuge-vs-numerical-powerlaw]. ¶4 (saturation-specific caveat). Water turbidity prevents direct visual verification of scour depth; \(S/D\) confirmed by cumulative excavated-mass balance.

§2.5 Backfill Protocol (proves C6). ¶1 (backfill material). Coarser No. 5 sand, \(d_{50} = 1.99\) mm, pre-saturated before pouring [→ fig-grain-size]. ¶2 (placement). ≈ 20 kg per bucket poured at 1 g while maintaining the water level, lightly tamped (≈ 10 applications per bucket) to moderate-density state. ¶3 (state documentation) [→ fig-backfill-state]. The backfill tests (T4-5, T5-5) apply this protocol identically so that dense-vs-loose recovery comparison isolates native-soil contrast.

Phase 6 — Generate figures (figure_generator integration)¶

Eleven figure specs, eleven figure_generator sessions. Two tables do not pass through the engine.

Example for fig-centrifuge-vs-numerical-powerlaw — the single most load-bearing figure because it contains the stress-history-artefact argument:

Figure: fig-centrifuge-vs-numerical-powerlaw
Paper: J3 methodology
Thesis: J3-methodology-thesis
Claim: C4 (stress-history artefact bounded)

Spec: Must show the centrifuge power-law coefficient |a| = 0.044-0.086 (range
across T4 and T5 saturated series) on the same axes as J2's numerical
|a| = 0.167, with the OCR-induced stiffness bias band (15-25%) shown as the
shaded difference. Log-log axes; annotation identifies the stress-history
artefact as the systematic offset.

Data: paperJ3_oe02685/figure_inputs/fig_centrifuge_vs_numerical_powerlaw.parquet
Schema: paperJ3_oe02685/figure_inputs/fig_centrifuge_vs_numerical_powerlaw_schema.yml
Journal: ocean_engineering
Width: single (90 mm)

The commit tag Claim: C4 is the audit link. A reviewer later asking "why are the centrifuge coefficients smaller than the numerical?" is answered by walking the chain C4 → evidence → figure → session commit → figure_inputs parquet and schema.

Phase 7 — Coherence passes¶

Pass A — paragraphs only, figures hidden.

§2.1 ¶4 on the \(S/D \le 0.6\) upper bound is already strong in the current manuscript. Kept.
§2.3 ¶3 on soil-state verification is currently worded as "D_r values of 70.2 % and 60.8 %" without stating that these are derived from CPT rather than target values. Fix: one sentence that names Baldi et al. (1986) and clarifies the measurement path.
§2.4 ¶3 on the stress-history artefact is present in the current manuscript but currently buried mid-paragraph. Fix: promote to its own subsection header (§2.4.1 "Stress-history considerations") so reviewers find it without search. The content is good; the findability is poor.
§2.5 ¶1 on backfill material does not currently state that the choice of coarser No. 5 sand is deliberate — to simulate engineered scour protection, not natural refill. Fix: one sentence on the design intent.

Pass B — figures only, paragraphs hidden.

Eleven figure specs, eleven figures. Mostly aligned with current manuscript.
fig-centrifuge-vs-numerical-powerlaw is missing from the current manuscript. This is the single most important evidence item for C4 and its absence is the highest-value gap. The argument currently lives as one paragraph of text referencing J2's coefficient; a figure would make it inspectable.
fig-mac-mode-shape is also missing. The MAC verification is mentioned in the current manuscript but without a figure. R1 reviewers may ask for this; producing it proactively is a cheap insurance.
fig-ppt-equilibration is partially present as a description in the flight-protocol text but no figure. Consider adding, since it directly supports the saturation-completeness claim (C2).
Two existing model-configuration photos are redundant. Fix: keep fig-test-setup, remove the second photograph.
One raw time-series snapshot (a single flight's frequency trace) does not map to any claim. Fix: either assign it to C5 as an example of the OMA input (and rename accordingly) or remove it.

Reviewer objection rehearsal. Walking the chain:

"How do we know the saturation is complete?" — C2, answered by Skempton B ≥ 0.95 and PPT equilibration [→ fig-ppt-equilibration]. Current text has the numerical threshold; adding the figure closes the loop visually.
"Does the stress-history artefact dominate the reported effect?" — C4, answered by the bound being 15–25 % while the saturation ratio is 1.94. The bias is directionally conservative (makes the centrifuge sensitivity an underestimate), so the reported saturation effect is a lower bound and therefore the true field sensitivity is at least 1.94×. Stating this explicitly in the paragraph is the fix.
"Why should dry-vs-saturated comparison not be confounded by structural model changes?" — C1, answered by identical structural model across J1 and J3. Current text says this in one sentence in §2.2; promoting to the abstract-level emphasis may help reviewers.
"How is the backfill recovery attributable to the backfill material rather than to redistribution of native soil?" — C6, answered by the tamping-protocol fidelity and the grain-size contrast [→ fig-grain-size]. The one-page-of-grain-size-curves figure is small-cost, high-clarification.

Summary of actionable items from the pass (in priority order):

#	Item	Type	Cost	Claim
1	Add `fig-centrifuge-vs-numerical-powerlaw`	New figure	1 `figure_generator` session	C4
2	Promote stress-history-artefact text to §2.4.1 subsection	Structural edit	10 min	C4
3	Add `fig-mac-mode-shape`	New figure	1 session	C5
4	Add `fig-ppt-equilibration`	New figure	1 session	C2
5	Add `fig-grain-size` (may already exist)	New/confirm figure	0–1 session	C6
6	Clarify \(D_r\) is CPT-derived, not target	Text edit, 1 sentence	5 min	C2
7	Remove redundant model-configuration photo	Figure cull	5 min	–
8	Remove or re-assign orphan time-series snapshot	Figure cull / assign	5 min	–

Three new figures, three paragraph fixes, two figure culls. Roughly one day of work, all of it directly responsive to reviewer concerns that R1 comments already flagged (soil-state verification depth, stress-history explicitness, methodological clarity).

What the two prototypes together surface¶

Applying the workflow to one numerical paper (J2) and one experimental paper (J3) produces a consistent pattern:

Most of the existing argument survives. Eight of ten J2 methodology figures and roughly eleven of fifteen J3 methodology figures map cleanly to claims. The workflow is not tearing up sections; it is identifying the smallest set of interventions that move them from sound to defensible.
Two-to-three missing figures per section are load-bearing and cheap. In both papers, the most important missing item is the one that makes the paper's most subtle argument visible: J2's stress-correction ablation, J3's centrifuge-vs-numerical comparison with OCR bias band. In both cases this is the same category of defect — the argument is in the prose; putting it in a figure is the fix.
The figure_generator engine is the natural vehicle for these fixes. Each new figure is one session, one commit, one audit-trail. Producing the handful of missing figures for R2 (J2) and R1 (J3) is roughly two days of work split across paired sessions.
The workflow is cheap to run. One prototype note per paper methodology section is about one hour to write. It surfaces the interventions that reviewers will otherwise find and flag, at a fraction of the cost of discovering them in the reviewer response.

The next step is to apply the same pass to the results section of J2 (different claim structure — headline finding, sensitivity, comparison against prior work, edge cases), run the missing-figure sessions, and ship the R2 response with the specific additions the workflow identified.