medsci-skills 4.1.0

package/skills/make-figures/references/exemplar_diagrams/pipeline/README.md

@@ -0,0 +1,15 @@
+# pipeline Exemplars
+
+This directory will hold 3–5 hand-curated pipeline flow/figure exemplars.
+Use the extraction script to add one:
+
+```bash
+python skills/make-figures/scripts/extract_exemplar_from_pdf.py \
+    --pdf "/path/to/paper.pdf" \
+    --page N \
+    --type pipeline \
+    --label LastnameYEAR_PIPELINE \
+    --doi 10.xxxx/yyyy
+```
+
+See `../README.md` for curation guidelines.

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_01.meta.yaml

@@ -0,0 +1,4 @@
+# Exemplar metadata (YAML)
+label: "pipeline_01"
+figure_type: "pipeline"
+dpi: 300

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_01_why.md

@@ -0,0 +1,13 @@
+# Why this exemplar is good — pipeline_01 (pipeline)
+
+Hierarchy / structure: A clean left-to-right pipeline — paired inputs (CXR image on top, clinical text on bottom) merge into a single downstream processor (MLLM) and terminate in a single output box. The two parallel encoder boxes are aligned horizontally, which communicates multimodal symmetry at a glance.
+
+Whitespace & balance: Generous horizontal whitespace around the central MLLM chip icon makes it the visual anchor. Input column and output column carry roughly equal ink weight, avoiding a lopsided read.
+
+Typography (font size, weight, alignment): Box labels (Image encoder / Text encoder) are bolded sans-serif; patient demographics use a smaller mono-like face to read as tabular data rather than prose. Illustrative "clinical question" is italicized and color-accented to flag that it is an example, not a label.
+
+Emphasis (which elements are visually strongest, why): Pink-tinted inputs and green output speech bubble draw attention to the clinically meaningful endpoints (question in → answer out), while the technical processing blocks sit in neutral gray.
+
+Color usage: Two saturated tints (pink for input, green for output) plus neutral gray for machinery. Colorblind-safe if converted — the saturation contrast survives grayscale.
+
+Weaknesses (if any — nothing is perfect): The MLLM "chip" icon is decorative rather than structural; could confuse readers expecting an architectural diagram. CXR thumbnail is small relative to the clinical-info panel, slightly under-weighting the image modality.

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_03.meta.yaml

@@ -0,0 +1,4 @@
+# Exemplar metadata (YAML)
+label: "pipeline_03"
+figure_type: "pipeline"
+dpi: 300

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_03_why.md

@@ -0,0 +1,13 @@
+# Why this exemplar is good — pipeline_03 (pipeline)
+
+Hierarchy / structure: Two-panel layout (A and B) separated by a thin vertical rule, with each panel self-contained. A shows contrastive pre-training: text encoder (top) and image encoder (bottom) converge on a similarity matrix. B shows a VQA inference pipeline flowing left → right. Panel labels sit in the top-left corner of each panel at a consistent position.
+
+Whitespace & balance: Margins around the similarity matrix give it room to read as a data object rather than decoration. The two panels carry roughly equal visual mass so neither dominates.
+
+Typography (font size, weight, alignment): Box labels use uniform sans-serif at one size; the similarity matrix cells use a mono face so Tᵢ·Iⱼ products align as a grid. The CXR/report thumbnails carry subscripts (T₁, T₂, I₁, I₂) in the same font as the matrix cells — visually tying the inputs to the matrix rows/columns.
+
+Emphasis (which elements are visually strongest, why): The diagonal of the similarity matrix is shaded darker, pulling the eye to the positive pairs — this is the pedagogical point of contrastive learning and the figure lets design carry the explanation.
+
+Color usage: Pink for text-side components, blue for image-side components. This two-channel color coding is consistent across both panels and makes the multimodal symmetry self-evident.
+
+Weaknesses (if any — nothing is perfect): Panel B's projection matrix W and token boxes are compressed; readers unfamiliar with contrastive-language-image architecture may need to read the caption closely. The "language instruction" box on the far right of B is visually detached from the main flow.

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_04.meta.yaml

@@ -0,0 +1,4 @@
+# Exemplar metadata (YAML)
+label: "pipeline_04"
+figure_type: "pipeline"
+dpi: 300

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_04_why.md

@@ -0,0 +1,13 @@
+# Why this exemplar is good — pipeline_04 (pipeline)
+
+Hierarchy / structure: Top row is the macro view (modality inputs → encoder → connector → LLM → generator → outputs), bottom row zooms into the "connector" block with four named variants (A MLP, B Q-Former, C MH-Attn, D Expert Captioning). A light-blue downward arrow visually tethers the macro block to its expansion — the classic "zoom-in detail" device applied cleanly.
+
+Whitespace & balance: Even horizontal spacing between connector variants. The expanded lower panel is framed in a subtle tinted rectangle that contains it without drawing attention.
+
+Typography (font size, weight, alignment): Consistent sans-serif labels throughout. Capital letters A/B/C/D tag the four variants in one size; component labels (MLP, Q-Former, MH-Attn) use the same type family as the macro row for continuity.
+
+Emphasis (which elements are visually strongest, why): The central LLM box is the largest and darkest block — correctly, because this is the figure's conceptual center. The connector variants underneath share equal weight, communicating their parallel status as alternatives rather than a sequence.
+
+Color usage: Restrained three-tone palette (blue for neural modules, orange for learnable queries/tokens, gray for data I/O). Dashed outlines mark optional/repeated elements. Colorblind-safe.
+
+Weaknesses (if any — nothing is perfect): Many modality icons (image/audio/video) repeat at both input and output; a single output cluster would reduce clutter. The Q/K/V labels inside panel C are small relative to neighboring elements.

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_05.meta.yaml

@@ -0,0 +1,4 @@
+# Exemplar metadata (YAML)
+label: "pipeline_05"
+figure_type: "pipeline"
+dpi: 300

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_05_why.md

@@ -0,0 +1,13 @@
+# Why this exemplar is good — pipeline_05 (pipeline / tasks illustration)
+
+Hierarchy / structure: Two stacked panels (A and B), each a four-column table-like layout. A shows report generation for two cases (pelvic radiograph and abdominal CT) side-by-side: input → reference report → generated report. B shows VQA with closed-ended vs. open-ended question variants for each case. The uniform column structure makes the cross-row comparison immediate.
+
+Whitespace & balance: Narrow column gutters (not wasteful) but generous vertical spacing between cases. Each row reads as a discrete case study.
+
+Typography (font size, weight, alignment): The strongest typographic choice is the use of colored highlights on specific phrases in the generated reports — green for confirmed findings, red for hallucinated content, orange for partial matches. This converts what would be plain prose into an annotated diff without extra legend burden, provided the reader notices the color key.
+
+Emphasis (which elements are visually strongest, why): The color-annotated report text is the primary didactic payload and is the strongest element, correctly. The input images serve as context and are sized accordingly smaller.
+
+Color usage: Color communicates semantic agreement/disagreement in text — a rare and effective use. Panel A uses gray scaffolding (headers, borders) to keep attention on the colored content.
+
+Weaknesses (if any — nothing is perfect): Red and green are the primary discriminators, which is not colorblind-safe. A shape or weight encoding (bold for hallucinations) would make it more accessible. Font size of the report text is small; readability at small figure widths could suffer.

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_06.meta.yaml

@@ -0,0 +1,4 @@
+# Exemplar metadata (YAML)
+label: "pipeline_06"
+figure_type: "pipeline"
+dpi: 300

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_06_why.md

@@ -0,0 +1,13 @@
+# Why this exemplar is good — pipeline_06 (pipeline / M3D architecture)
+
+Hierarchy / structure: Strict top-to-bottom flow. Volumetric CT input (single modality box with three anatomical sub-panels) feeds a 3D image encoder; parallel text-side question templates feed the text tokenizer. Both streams converge on a single LLM-with-LoRA block, then split into two downstream heads (report generation, VQA). The shape (funnel in → funnel out) mirrors the dual-task architecture.
+
+Whitespace & balance: Equal space allocated to image-side and text-side inputs; the LLM block spans the full width to signal it is the shared backbone. Tight but not cramped.
+
+Typography (font size, weight, alignment): Bolded section headers (Report Generation / VQA) for the prompt boxes; prompt contents in a slightly smaller, clearly italic-or-lighter weight. The special tokens (<|eot_id|>, <|start_header_id|>) use a monospace face — the right choice for communicating that they are literal code strings.
+
+Emphasis (which elements are visually strongest, why): The fire/snow icons (trainable/frozen) are placed on individual components, making the training-recipe claim visually inspectable. This is a small but high-value design choice — readers can audit the method by looking at the figure.
+
+Color usage: Cool blues for vision modules, warm pinks for language/prompt modules. Two-tone palette scales well to grayscale (luminance differs).
+
+Weaknesses (if any — nothing is perfect): The "trainable/frozen" legend sits in the upper right and may be missed by readers who scan left-to-right top-down. The three CT sub-views (chest/abdomen/pelvis) are small enough that anatomy is only suggested.

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_07.meta.yaml

@@ -0,0 +1,4 @@
+# Exemplar metadata (YAML)
+label: "pipeline_07"
+figure_type: "pipeline"
+dpi: 300

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_07_why.md

@@ -0,0 +1,13 @@
+# Why this exemplar is good — pipeline_07 (pipeline / MAIRA-Seg)
+
+Hierarchy / structure: A horizontal pipeline with input row (CXRs, pseudo-masks, task instruction, textual context) on top, processing row (image encoder, segmentation model, tokenizer/embedding) in the middle, and convergence on a single LLM block at the bottom. The vertical progression communicates "modalities in → unified tokens → one model out" without ambiguity.
+
+Whitespace & balance: The LLM block spans almost full width as a visual foundation. Upper row inputs are grouped by modality with clear gaps, avoiding the "everything dumps in" look.
+
+Typography (font size, weight, alignment): Small but uniform sans-serif labels; auxiliary labels ("Only in multi", "Pseudo-labels") are italic to mark them as conditional annotations without adding a legend.
+
+Emphasis (which elements are visually strongest, why): The LLM base block is the largest uniform colored rectangle — visually grounding the whole system as one model. Fire icons on trainable components again let the figure double as a training-recipe specification.
+
+Color usage: Warm pinks for segmentation/prompt-side modules; cool blues for vision/LLM modules. Dashed borders denote optional paths ("Only in multi"). Consistent with the color code used in other architecture figures in the paper.
+
+Weaknesses (if any — nothing is perfect): The dense upper input row competes for attention; a subtle grouping box or lane separator would help. The segmentation-tokens-extractor sits off the main flow axis, which introduces a jog in the layout that briefly interrupts the top-down read.

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_08.meta.yaml

@@ -0,0 +1,4 @@
+# Exemplar metadata (YAML)
+label: "pipeline_08"
+figure_type: "pipeline"
+dpi: 300

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_08_why.md

@@ -0,0 +1,13 @@
+# Why this exemplar is good — pipeline_08 (pipeline / CLIMB foundation model)
+
+Hierarchy / structure: Three-stage landscape layout separated by vertical rules: Training (left) | Specialized Encoder (center) | Testing + Downstream (right). Inside the Training panel, three modality lanes (Graphs, 1D signals, 2D&3D images) each carry three exemplar thumbnails, creating a grid-of-grids that communicates breadth without listing every dataset.
+
+Whitespace & balance: Equal vertical space per modality lane; the encoder column is visually narrow to read as a bottleneck/convergence point. Right-side testing block has enough margin around the OOD dataset callout to keep it distinct.
+
+Typography (font size, weight, alignment): Modality headers (Graphs, 1D signals, 2D&3D images) are larger and bolder than thumbnail captions. Section headers (Training / Testing) sit above their column in matching weight.
+
+Emphasis (which elements are visually strongest, why): The thumbnail grid at training time is the strongest visual payload — it answers "what does multimodal breadth look like?" in one glance. The encoder column is intentionally understated because it is a conceptual stage, not a specific architecture.
+
+Color usage: Subtle pastel backgrounds separate stages (soft pink for training, green for testing, peach for downstream) without compromising the thumbnails' own color. Downstream task illustrations (physician + patient) add warmth without clutter.
+
+Weaknesses (if any — nothing is perfect): Ten or more thumbnail captions at small sizes risk clipping on reprints. The "transfer" and "finetuned" arrows in the testing column could be slightly thicker to read as a pipeline rather than decoration.

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_09.meta.yaml

@@ -0,0 +1,4 @@
+# Exemplar metadata (YAML)
+label: "pipeline_09"
+figure_type: "pipeline"
+dpi: 300

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_09_why.md

@@ -0,0 +1,13 @@
+# Why this exemplar is good — pipeline_09 (evaluation-metrics comparison)
+
+Hierarchy / structure: A 3×2 grid where all six evaluation methods share identical sub-layout — method title (bar header), candidate excerpt (middle), and two-bullet summary (O = what is captured, △ = what is missed). One shared reference sentence sits above the grid, establishing the common ground-truth all six metrics are being compared against.
+
+Whitespace & balance: The top reference strip spans full width and is visually distinct from the grid below. Within each cell, internal padding is consistent and there is clear separation between the colored header bar and the body — giving each metric equal visual real-estate.
+
+Typography (font size, weight, alignment): Bold uppercase metric names in colored header bars; candidate text below in regular weight with discriminating phrases colored inline (e.g., "right-sided" in red for a miss). The O/△ legend uses symbols instead of words, keeping each cell compact.
+
+Emphasis (which elements are visually strongest, why): The six colored header bars act as chapter tabs, making the grid scannable. Colored phrases inside the candidate text are the didactic focus — they show precisely what each metric is scoring.
+
+Color usage: Six distinct pastel headers, chosen for differentiation rather than semantic meaning. Candidate-text color (red/green/blue) is semantic. Palette is varied but low-saturation enough to avoid fatigue.
+
+Weaknesses (if any — nothing is perfect): Six colors with no ordinal relationship could be reduced; the reader doesn't need to distinguish cells by color since the grid layout already separates them. Semantic text colors (red for missed) overlap with header colors in a few cells.

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_10.meta.yaml

@@ -0,0 +1,4 @@
+# Exemplar metadata (YAML)
+label: "pipeline_10"
+figure_type: "pipeline"
+dpi: 300

package/skills/make-figures/references/exemplar_diagrams/pipeline/pipeline_10_why.md

@@ -0,0 +1,13 @@
+# Why this exemplar is good — pipeline_10 (clinical integration overview)
+
+Hierarchy / structure: Three horizontal strips. Top = clinical workflow (Patient → Triage → Consultation → Diagnosis → Treatment → Monitoring) rendered as a linear process diagram. Middle = radial diagram with a radiologist at center surrounded by MLLM support tasks. Bottom = radial diagram with an AI icon at center surrounded by clinical-integration challenges. Two radial compositions mirror each other visually — a deliberate parallel that frames "capabilities vs. challenges".
+
+Whitespace & balance: Each strip is framed in a light background panel with its title on the left (Clinical Workflow / MLLM Support for Radiologist / Key Challenges for Clinical Integration) — the row-label placement anchors the reader and allows side-by-side reading without a separate key.
+
+Typography (font size, weight, alignment): Row-label italics on the left are consistent. Radial labels (Visual Question Answer, Clinical Decision Support, etc.) use sentence case and are placed at consistent radial distances from center.
+
+Emphasis (which elements are visually strongest, why): The two central icons (radiologist, AI chip) draw the eye first, making each radial panel's subject unmistakable. The triage/diagnosis/treatment icons in the top strip are sized uniformly so no single step is over-emphasized — correctly, because this is a generic workflow.
+
+Color usage: Warm peach/cream for support, cool blue for challenges — a clean semantic contrast. Individual radial spokes are lightly tinted to differentiate themes within each group.
+
+Weaknesses (if any — nothing is perfect): The radial metaphor collapses any ordering between spokes — a reader cannot tell which challenge is most pressing or which MLLM role is primary. Some icon glyphs (e.g., "Final Decision by Radiologist") are small and may not read clearly when printed at single-column width.

package/skills/make-figures/references/exemplar_diagrams/prisma/README.md

@@ -0,0 +1,15 @@
+# prisma Exemplars
+
+This directory will hold 3–5 hand-curated prisma flow/figure exemplars.
+Use the extraction script to add one:
+
+```bash
+python skills/make-figures/scripts/extract_exemplar_from_pdf.py \
+    --pdf "/path/to/paper.pdf" \
+    --page N \
+    --type prisma \
+    --label LastnameYEAR_PRISMA \
+    --doi 10.xxxx/yyyy
+```
+
+See `../README.md` for curation guidelines.

package/skills/make-figures/references/exemplar_diagrams/prisma/template_input.yaml

@@ -0,0 +1,47 @@
+# PRISMA 2020 systematic review / meta-analysis flow diagram template.
+# Render with:
+#   Rscript ../../scripts/generate_flow_diagram.R \
+#       --type prisma --config template_input.yaml --out template_output
+rankdir: TB
+nodes:
+  - id: id_db
+    label: "Records identified from databases\n(n = 3,500)\n• PubMed: 1,800\n• Embase: 1,200\n• Cochrane: 500"
+  - id: id_other
+    label: "Records identified from other sources\n(n = 45)\n• Citation search: 30\n• Registers: 15"
+    rank_same_with: id_db
+  - id: dedup
+    label: "Records after duplicates removed\n(n = 2,700)"
+  - id: screened
+    label: "Records screened (title/abstract)\n(n = 2,700)"
+  - id: excl_ta
+    shape: note
+    fontsize: 10
+    label: "Records excluded\n(n = 2,480)"
+    rank_same_with: screened
+  - id: ft_sought
+    label: "Reports sought for retrieval\n(n = 220)"
+  - id: ft_notret
+    shape: note
+    fontsize: 10
+    label: "Reports not retrieved\n(n = 12)"
+    rank_same_with: ft_sought
+  - id: ft_assessed
+    label: "Reports assessed for eligibility\n(n = 208)"
+  - id: ft_excl
+    shape: note
+    fontsize: 10
+    label: "Reports excluded (n = 180):\n• Wrong population: 72\n• Wrong index test: 48\n• Wrong outcome: 35\n• Abstract only: 15\n• Duplicate cohort: 10"
+    rank_same_with: ft_assessed
+  - id: included
+    label: "Studies included in review\n(n = 28)\nStudies in quantitative synthesis\n(n = 24)"
+    highlight: true
+edges:
+  - {from: id_db, to: dedup}
+  - {from: id_other, to: dedup, constraint: false}
+  - {from: dedup, to: screened}
+  - {from: screened, to: ft_sought}
+  - {from: ft_sought, to: ft_assessed}
+  - {from: ft_assessed, to: included}
+  - {from: screened, to: excl_ta, style: dashed, arrow: false, constraint: false}
+  - {from: ft_sought, to: ft_notret, style: dashed, arrow: false, constraint: false}
+  - {from: ft_assessed, to: ft_excl, style: dashed, arrow: false, constraint: false}

package/skills/make-figures/references/exemplar_diagrams/stard/README.md

@@ -0,0 +1,15 @@
+# stard Exemplars
+
+This directory will hold 3–5 hand-curated stard flow/figure exemplars.
+Use the extraction script to add one:
+
+```bash
+python skills/make-figures/scripts/extract_exemplar_from_pdf.py \
+    --pdf "/path/to/paper.pdf" \
+    --page N \
+    --type stard \
+    --label LastnameYEAR_STARD \
+    --doi 10.xxxx/yyyy
+```
+
+See `../README.md` for curation guidelines.

package/skills/make-figures/references/exemplar_diagrams/stard/template_input.yaml

@@ -0,0 +1,40 @@
+# STARD 2015 diagnostic-accuracy flow diagram template.
+# Render with:
+#   Rscript ../../scripts/generate_flow_diagram.R \
+#       --type stard --config template_input.yaml --out template_output
+rankdir: TB
+nodes:
+  - id: eligible
+    label: "Eligible patients\n(n = 500)"
+  - id: excl
+    shape: note
+    fontsize: 10
+    label: "Excluded (n = 50):\n• Contraindication to index test: 20\n• No reference standard available: 20\n• Declined: 10"
+    rank_same_with: eligible
+  - id: index
+    label: "Received index test\n(n = 450)"
+    highlight: true
+  - id: ref
+    label: "Received reference standard\n(n = 450)"
+  - id: idx_pos
+    label: "Index test positive\n(n = 180)"
+  - id: idx_neg
+    label: "Index test negative\n(n = 270)"
+  - id: tp
+    label: "Reference positive\nTP = 160"
+  - id: fp
+    label: "Reference negative\nFP = 20"
+  - id: fn
+    label: "Reference positive\nFN = 15"
+  - id: tn
+    label: "Reference negative\nTN = 255"
+edges:
+  - {from: eligible, to: index}
+  - {from: index, to: ref}
+  - {from: ref, to: idx_pos}
+  - {from: ref, to: idx_neg}
+  - {from: idx_pos, to: tp}
+  - {from: idx_pos, to: fp}
+  - {from: idx_neg, to: fn}
+  - {from: idx_neg, to: tn}
+  - {from: eligible, to: excl, style: dashed, arrow: false, constraint: false}

package/skills/make-figures/references/exemplar_diagrams/strobe/template_input.yaml

@@ -0,0 +1,43 @@
+# STROBE cohort-study flow diagram template.
+# Render with:
+#   Rscript ../../scripts/generate_flow_diagram.R \
+#       --type strobe --config template_input.yaml --out template_output
+rankdir: TB
+nodes:
+  - id: source
+    label: "Source population\n(Database XYZ, 2010–2023; total records = 120,000)"
+  - id: unique
+    label: "Unique subjects after deduplication\nN = 85,000"
+  - id: excl1
+    shape: note
+    fontsize: 10
+    label: "Excluded (n = 5,000):\n• Missing age/sex: 1,200\n• Outside age range: 3,800"
+    rank_same_with: unique
+  - id: eligible
+    label: "Meeting eligibility criteria\nN = 80,000"
+  - id: excl2
+    shape: note
+    fontsize: 10
+    label: "Excluded (n = 2,500):\n• Prior event at baseline: 1,500\n• Follow-up ≤ 0 days: 1,000"
+    rank_same_with: eligible
+  - id: cohort
+    label: "Analytic cohort\nN = 77,500"
+    highlight: true
+  - id: exp
+    label: "Exposed\nn = 20,000"
+  - id: unexp
+    label: "Unexposed\nn = 57,500"
+  - id: out_exp
+    label: "Incident events\n1,200 (6.0%)"
+  - id: out_unexp
+    label: "Incident events\n2,300 (4.0%)"
+edges:
+  - {from: source, to: unique}
+  - {from: unique, to: eligible}
+  - {from: eligible, to: cohort}
+  - {from: cohort, to: exp}
+  - {from: cohort, to: unexp}
+  - {from: exp, to: out_exp}
+  - {from: unexp, to: out_unexp}
+  - {from: unique,   to: excl1, style: dashed, arrow: false, constraint: false}
+  - {from: eligible, to: excl2, style: dashed, arrow: false, constraint: false}

package/skills/make-figures/references/exemplar_diagrams/strobe/template_input_pptx.yaml

@@ -0,0 +1,43 @@
+# STROBE cohort-study flow diagram — editable PPTX schema.
+#
+# Render with:
+#   python3 ../../scripts/build_strobe_template.py \
+#       --config template_input_pptx.yaml \
+#       --out    template_output.pptx
+#
+# Output is a fully-editable PowerPoint slide (13.33 × 10 in) with:
+#   - left-side colored phase column (one merged label per consecutive same-stage spine block)
+#   - center white-fill spine of cohort-stage boxes
+#   - right-side white-fill exclusion boxes, connected by arrows
+#
+# Stage box text colour auto-picks BLACK or WHITE from fill luminance.
+# All text frames have word_wrap = True so prose never overflows.
+
+slide_size: [13.33, 10.0]
+title: "Figure 1. STROBE participant flow diagram"
+
+stages:
+  - {name: Identification, color: "#cfe1f5"}
+  - {name: Screening,      color: "#cfe1f5"}
+  - {name: Inclusion,      color: "#cfe1f5"}
+  - {name: Analysis,       color: "#cfe1f5"}
+
+spine:
+  - id: source
+    stage: Identification
+    text: "Source population\n(Database XYZ, 2010–2023)\ntotal records = 120,000"
+  - id: unique
+    stage: Screening
+    text: "Unique subjects after deduplication\nN = 85,000"
+  - id: eligible
+    stage: Inclusion
+    text: "Meeting eligibility criteria\nN = 80,000"
+  - id: cohort
+    stage: Analysis
+    text: "Analytic cohort\nN = 77,500"
+
+exclusions:
+  - after: unique
+    text: "Excluded (n = 5,000):\n- missing age/sex (n = 1,200)\n- outside age range (n = 3,800)"
+  - after: eligible
+    text: "Excluded (n = 2,500):\n- prior event at baseline (n = 1,500)\n- follow-up ≤ 0 days (n = 1,000)"