J5 Methodology — Seven-Phase Claim-Driven Prototype¶

Date: 2026-04-18 Status: Fourth prototype of the seven-phase claim-driven workflow, applied prospectively to Paper J5 — the 1,794-realisation Monte Carlo 3D limit-analysis paper. J5 has the cleanest claim structure in the portfolio (eight sampled parameters, one capacity envelope per realisation, Sobol sensitivity indices), so the workflow passes through quickly.

Running the workflow on J5 now — while the manuscript is still in planning — means the figure budget, claim list, and evidence map are locked before the 1,794-realisation output gets turned into prose. That is the mode the workflow is designed for.

Phase 1 — Thesis statement¶

Propagating realistic distributions of soil parameters and local scour geometry through a 1,794-realisation Monte Carlo ensemble of 3D limit analyses yields a population distribution of tripod suction bucket capacity whose 5th-percentile is approximately 30 % below the deterministic estimate at \(S/D = 1\), with three dominant parameters accounting for > 80 % of variance and a reproducible in-situ positioning procedure that places the Gunsan site within the distribution.

Phase 2 — Claim list¶

C1 — Sampling design is fit-for-purpose. Latin-hypercube sampling over the eight parameters covers the realistic CPT-informed parameter space without clustering artefacts; marginal prior distributions are literature-backed (Phoon & Kulhawy 1999).
C2 — The ensemble is reproducible. 1,794 realisations distributed across five PCs, per-realisation Op³ checkpoint logs, every output traceable to input hash + framework version pin, SHA-256 manifest on the Zenodo bundle.
C3 — Sobol sensitivity identifies the dominant parameters. Three of eight parameters (undrained strength gradient \(k_{s_u}\), small-strain modulus ratio \(G_0/s_u\), scour depth \(S/D\)) account for > 80 % of horizontal-capacity variance; first-order and total-effect indices confirm no material interaction among the remaining five.
C4 — The 5th-percentile gap is structural, not sampling noise. 5th-percentile capacity is 30 % below the deterministic estimate at \(S/D = 1\); the gap is robust to resampling the bottom 10 % of realisations and stable across sub-samples of size \(\ge 500\).
C5 — Gunsan in-situ positioning is auditable. Site CPT profile + bathymetric scour depth → quantile within the population distribution, with uncertainty propagation traceable from CPT measurement noise through to capacity quantile.

Phase 3 — Claim → evidence map¶

flowchart TD
    T["<b>Thesis</b><br/>1,794 MC realisations yield capacity population<br/>whose 5th pct is 30% below deterministic"]
    T --> C1["C1 · LHS sampling design fit-for-purpose"]
    T --> C2["C2 · Ensemble reproducible"]
    T --> C3["C3 · Sobol identifies 3 dominant parameters"]
    T --> C4["C4 · 5th-pct gap is structural"]
    T --> C5["C5 · In-situ positioning auditable"]
    C1 --> E1["Fig · LHS coverage across param pairs<br/>Table · prior distributions per parameter"]
    C2 --> E2["Fig · 5-PC distribution + runtime profile<br/>Table · Op³ version pin + checksum manifest"]
    C3 --> E3["Fig · Sobol first-order + total bars<br/>Fig · scatter of capacity vs 3 dominant params"]
    C4 --> E4["Fig · population VH envelope (mean + 5/50/95 pct)<br/>Table · 5th-pct gap vs deterministic per S/D bin<br/>Fig · convergence curve (5th pct vs N realisations)"]
    C5 --> E5["Fig · Gunsan site positioned in population<br/>Text · uncertainty-propagation walk-through"]
    E1 --> S1["spec · fig-lhs-coverage<br/>table · prior-distributions"]
    E2 --> S2["spec · fig-pipeline-runtime<br/>table · reproducibility-manifest"]
    E3 --> S3["spec · fig-sobol-bars<br/>spec · fig-capacity-vs-params"]
    E4 --> S4["spec · fig-vh-envelope-population<br/>table · gap-per-scour<br/>spec · fig-convergence"]
    E5 --> S5["spec · fig-gunsan-positioning"]

Seven figure specs, three tables.

Phase 4 — Figure specs¶

fig-lhs-coverage (C1). Must show Latin-hypercube coverage on three representative parameter pairs (\(k_{s_u}\) vs \(G_0/s_u\); \(S/D\) vs azimuthal extent; density vs interface \(\alpha\)), with LHS points overlaid on the uniform-grid reference. Three panels side-by-side; annotation confirms no clustering.

table-prior-distributions (C1). Eight rows (one per sampled parameter): parameter name, prior family, location, scale, literature citation. Passes as the "what was sampled" summary in one glance.

fig-pipeline-runtime (C2). Must show per-realisation runtime across the five distributed PCs as a stacked timeline (one row per PC), with the total wall-clock end-time marked. Annotation: 1,794 realisations completed in \(X\) hours total; mean per-realisation runtime \(Y\) seconds.

table-reproducibility-manifest (C2). Columns: realisation ID range, Op³ version, input-hash range, output-checksum prefix, Zenodo DOI reference. Five rows (one per PC).

fig-sobol-bars (C3). Horizontal bar chart with eight parameters, showing first-order Sobol index (light bar) and total-effect index (dark bar) side-by-side per parameter. Parameters ordered by total-effect descending. Annotation: top three account for \(\ge 80\) %.

fig-capacity-vs-params (C3). Three panels, one per dominant parameter (\(k_{s_u}\), \(G_0/s_u\), \(S/D\)), showing scatter of horizontal capacity \(H_{\max}\) vs that parameter with the remaining seven parameters averaged. Confirms the Sobol ranking visually.

fig-vh-envelope-population (C4). VH capacity envelope with mean, 5th, 50th, 95th percentile bands shaded. Overlay: deterministic single-point estimate as a dot. The visual gap between the dot and the 5th-percentile band is the load-bearing result of the paper.

table-gap-per-scour (C4). One row per \(S/D\) bin in \(\{0, 0.2, 0.4, 0.6, 0.8, 1.0\}\). Columns: deterministic horizontal capacity, 5th-percentile capacity, relative gap (%), 95 % CI on the gap.

fig-convergence (C4). Convergence of the 5th-percentile estimate as sub-sample size \(N\) grows from 200 to 1,794. Monotone stabilisation within \(\pm 2\) % beyond \(N = 500\) confirms the 1,794-realisation ensemble is over-sampled for this quantile.

fig-gunsan-positioning (C5). Population distribution of horizontal capacity (shaded violin or histogram) with the Gunsan in-situ state marked as a vertical line, plus a shaded band around that line from CPT measurement-noise propagation. Annotation: current site is at the \(P\)-th percentile; projected 20-year scour pushes it to the \(Q\)-th percentile.

Phase 5 — Paragraph skeletons¶

§3 Methods. ¶1 (parameter space) [→ table-prior-distributions]. Eight parameters with CPT-informed priors; sources cited (Phoon & Kulhawy 1999; local Gunsan CPT survey; DNV-RP-C212). ¶2 (sampling design) [→ fig-lhs-coverage]. LHS over eight dimensions; 1,794 realisations chosen for 5th-percentile convergence (validated by fig-convergence in §4). ¶3 (distributed pipeline) [→ fig-pipeline-runtime, table-reproducibility-manifest]. Five-PC distribution via Op³; per-realisation Op³ version pin + input-hash + output-checksum logged. ¶4 (post-processing). Each realisation → VH envelope via Op³'s standardised angular-probe routine (20 probes, matching J2 protocol).

§4 Results. ¶1 (population envelope) [→ fig-vh-envelope-population]. Mean, 5/50/95 percentile bands. Visual gap between deterministic dot and 5th-percentile band is the headline. ¶2 (Sobol sensitivity) [→ fig-sobol-bars, fig-capacity-vs-params]. Three dominant parameters, > 80 % of total variance; five effectively deterministic. ¶3 (scour dependence) [→ table-gap-per-scour]. Gap widens with \(S/D\); at \(S/D = 1\), 5th-pct is ~30 % below deterministic. ¶4 (convergence audit) [→ fig-convergence]. 5th-percentile estimate stable beyond \(N = 500\); 1,794 realisations is oversampling for this quantile (deliberate — preserves power for downstream reliability calculations in Paper A).

§5 In-situ positioning. ¶1 (procedure). CPT profile → soil-parameter posterior → Monte Carlo query → capacity quantile. ¶2 (Gunsan application) [→ fig-gunsan-positioning]. Site at ~40th percentile today; projected 20-year scour pushes to ~20th percentile. ¶3 (uncertainty propagation). CPT measurement noise propagates to \(\pm X\) percentile uncertainty on the site position. Text-only walkthrough.

Phase 6 — Generate figures (figure_generator integration)¶

Seven figure_generator sessions + three tables. fig-vh-envelope-population is the hero figure; produce it first so the paper's headline visual is settled before downstream figures are wired in.

Figure: fig-vh-envelope-population
Paper: J5
Thesis: J5-main-thesis
Claim: C4 (5th-percentile gap is structural, not sampling noise)

Spec: VH capacity envelope with mean, 5th, 50th, 95th percentile bands
shaded; deterministic single-point estimate overlaid as a dot. Visual
gap between the dot and the 5th-percentile band is the central result.

Data: paperJ5/figure_inputs/fig_vh_envelope_population.parquet
Schema: paperJ5/figure_inputs/fig_vh_envelope_population_schema.yml
Journal: computers_geotechnics
Width: single (90 mm)

Phase 7 — Coherence passes¶

Pass A — paragraphs only. - §3 and §4 must cleanly separate sampling design (§3) from sampling diagnostics (§4). The convergence audit belongs in §4, not §3. - §5 (in-situ positioning) currently reads as an application rather than a method extension. Decide: is §5 part of the headline ("we can position Gunsan") or an afterward case study? The manuscript brief treats it as a contribution; the paragraph skeleton should reflect that.

Pass B — figures only. - fig-sobol-bars and fig-capacity-vs-params together establish C3. If the Sobol bars land, the scatter panels are a double-check; consider merging into a single figure with the bars as the main panel and the scatter insets as thumbnails. - Hero figure candidate: fig-vh-envelope-population vs table-gap-per-scour. The figure shows one scour depth well; the table shows all scour depths numerically. Both should exist but the figure is the one reviewers remember.

Actionable items before drafting:

#	Item	Type	Estimated cost
1	Build `fig-vh-envelope-population` as the hero figure	New figure	1 session
2	Build `fig-sobol-bars`	New figure	1 session
3	Build `fig-convergence` (confirms N=1,794 over-samples)	New figure	1 session
4	Build `fig-gunsan-positioning` (site-context)	New figure	1 session
5	Decide §5 scope (application vs headline contribution)	Structural	30 min
6	Consider merging `fig-sobol-bars` + `fig-capacity-vs-params`	Figure cull	30 min

Five figure_generator sessions + two small decisions. Roughly one working day.

What the J5 pass surfaced¶

Unlike J2 and J3 (which are manuscript-rewrite situations), J5 is a clean prospective pass — the workflow lets the figure budget get locked before drafting, which means every figure has an audit trail and no orphans should appear in R1 review. The §5 scope decision (application vs headline contribution) is the one item a reviewer will probe; resolving it now is a reviewer-risk reduction.