# Scientific figures in the current host workflow

Use this method before creating or redesigning statistical figures, mechanisms,
conceptual frameworks, graphical abstracts and method architectures. It carries
the scientific reference-deconstruction and component-routing method inside the
single PaperSpine Skill. Ordinary host drawing and image tools can execute it;
historical Runner contracts, JSON attestations and a second Skill are unnecessary.

Apply the stages needed for the requested output. A reference-selection or design-
brief task ends with viewed sources and an executable brief; it does not require
creating a paper task, filling a whole-paper learning quota or generating final
figures. Label this stage accurately. Rendering, selection and final visual review
become necessary when the requested work reaches those outputs.

## Scientific figure methods, read at the relevant stage

Use this workflow before generating a candidate, including during revision.
Follow the same reference-to-figure method for a graphical abstract, conceptual
framework and mechanism schematic.

Before planning prose around existing figures, read the complete
[figure-story.md](figure-story.md), including the scientific content
inside its planning and integration examples. Inspect the real image, identify
the question, dominant claim, panel roles, evidence and uncertainty; preserve
the per-panel planning details, choose keep/redesign/create, and write Results
as question → evidence → bounded answer. Apply its execution wrappers through
the current method routing notes instead of dropping their scientific content.

Before reference-guided generation, read the scientific mapping bullets and
content-transfer limits in [figure-reference-mapping.md](figure-reference-mapping.md),
plus its Background Rule and Reuse of Point/Line Grammar. Actually open the
reference source and figure. Record which layout, palette, typography, marks,
panel hierarchy and annotations transfer to the present question. Map each
panel to the current data/units/transform and preserve domain differences.
Never transfer the reference paper's numbers, conclusions or patient details.
Without a lawful viewable reference, say the design is original; do not claim
reference-guided work. Prefer white/transparent scientific canvases unless the
user or venue calls for another justified presentation.

After selection and assembly, use figure-story.md section 6's scientific image/body checks:
inspect final pixels against source results, labels, captions, manuscript
references, panel roles and claim boundaries; inspect actual reference/current
views at readable size. Keep editable figure sources and a short mapping note.

The references' legacy compatibility sections describe Runner schemas, Master
attestations and PaperFigure handoffs only for explicit legacy maintenance.
Current work uses ordinary notes, sources, public choices and real-file review.

## Plan the paper's figures before narrowing the Results

Inspect all supplied result tables, figures and relevant methods, including useful
results outside the current draft. Before choosing chart types or a figure count,
map the paper's scientific questions to available evidence in the existing notes.
Distinguish measured results, method explanations, literature-supported concepts
and evidence not available under the saved scope. Decide which questions benefit
from a visual argument; then compare strong reference structures for those jobs
before settling the figure and panel plan. Do not first fix a few simple plots
and search only for published plots that resemble them.

Give figures complementary jobs, connecting setup, primary findings, comparisons,
explanation and robustness where supported. Account for useful supplied results
in the main figures, tables, text or supplement, or explain their omission. This
is coverage of the evidence, not a demand to visualize every result. Merge
redundant displays and retain informative comparisons, uncertainty and negative
results. The article type, venue space and supported questions determine figure
count; there is no numerical default or quota. A single figure can suffice when
it covers the needed visual argument. Extra panels, formats or repeated statistics
cannot compensate for a missing evidence layer. Before finalizing the count, scan
actual outputs for distinct evidence families: cohort/availability, primary effect,
comparison/discrimination, uncertainty or calibration/validation, robustness or
sensitivity/subgroup, and explanation. Assign each useful family to a figure, table,
prose or supplement, or state the scientific reason for omission in the existing
plan. Do not let the files already drawn define this inventory.

Preserve valid supplied analyses and rich figures without demanding raw arrays.
`materials_only` prohibits new analyses, experiments, training, evaluation and
reruns; it permits reorganizing and redrawing existing results. If a proposed
display needs results that are not supplied, record that specific gap and use a
supported design. Do not split the same summary measures into several figures
to inflate the count, synthesize
clinical results, or silently approve journal-level completeness despite a known
scientific gap. Synthetic demonstrations can teach composition, never supply data
for a real study.

For `mechanism_figure: auto` or `prefer`, record the suitability decision in the
existing figure plan before candidate generation. If the explanatory job is central
and adequately sourced, make a complete mechanism/conceptual/architecture candidate
in addition to necessary data figures; do not count a cohort flowchart, ROC panel or
statistical dashboard as fulfilling it. If omitted under `auto`, state the concrete
reason.

For an explanatory figure, assess literature support separately from this study's
ability to establish causality. A cross-sectional or observational study can still
use a literature-informed conceptual framework; lack of a tested mechanism is not
by itself a reason to omit it. When it helps answer a planned question or is
requested, use relevant public literature to explain the system and label observed associations, established
knowledge and hypotheses distinctly. Do not invent causal evidence or replace the
explanation with numerical summary cards. Omit it when it adds no useful explanation
or lacks adequate sources, and record that concrete reason in the figure plan.
Judge omissions by the proposed figure's requirements: absent follow-up can rule
out a survival plot, while absent causal validation does not rule out a sourced
conceptual framework. A redundant display or a text/table that communicates the
question better is also a reasonable no-figure decision.

## Read the reference before designing

Open the existing figure and its editable source when present. Decide keep,
redesign or create from its scientific job and visible quality. Preserve valid
user-provided research and reusable high-quality elements.

For reference-guided work, actually view the complete reference figure at page
size and inspect its relevant detail before producing candidate bytes. Start with
the same-field and target-venue learning sets. If these contain only statistical
figures, find a relevant, strong mechanism/framework/architecture figure for that
explanatory figure; a survival plot cannot serve as its mechanism-design exemplar.
Assess relevance, scientific communication and quality at the target venue's level
or above. Do not choose a weak diagram merely because its image is easy to obtain.
Keep the local reference, its paper/venue/year and exact page/figure locator.

Separate a **knowledge source** from a **visual design exemplar**. A relevant paper's
simple summary diagram may establish which processes to discuss while offering too
little structure to guide a rich mechanism. Name each source's role in the existing
figure notes. For the primary design exemplar, inspect how it explains the process
inside the system: spatial compartments or modules, meaningful internal structures,
connections across scales, and branches or feedback where scientifically relevant.
These are transferable design opportunities, not a quota of organs, molecules or
arrows. A high venue rank, related title or realistic organ icon cannot substitute
for them. When the topic-matched reference is visually weak, retain it for knowledge
and find a stronger design exemplar; do not lower the visual benchmark to match it.

For a new complex main figure or architecture, compare a shortlist of strong
figures before settling on a reference, using the learning sets and additional
sources where needed. Compare how the alternatives organize the needed evidence
or explanation, not how many examples were collected. This focused figure study
does not replace the saved paper-learning requirements.
Prefer references with substantial transferable scientific structure: meaningful
multi-panel integration, a dominant result, informative comparisons/uncertainty,
linked scales or modules, and readable detail at publication size. Select one
primary exemplar and complementary exemplars where they fill a specific design
need, explaining why they fit. Avoid treating an easy-to-draw forest plot or a
few generic boxes as sufficient when better-matched, richer exemplars are available.
Reuse an already strong, actually inspected reference when it still fits; no
repeated search or fixed scoring dossier is needed for a minor correction.

Match the primary exemplar to the figure's scientific job and information
structure, not merely a shared chart mark or topic. For example, advice to show
individual points does not by itself guide a linked joint-distribution, marginal
and rank-comparison composition. Assess each complex figure's relevant shortlist;
unrelated survival plots, graphical abstracts and rejected examples do not add
up to a strong mechanism shortlist. One exemplar may serve related panels when
its structure actually transfers. State the missing transferable layer and seek a
better-matched example when needed. Honest original design remains useful work,
but it does not complete a requested matched-reference comparison.

Resolve that locator from the article's actual figure label and attached caption,
then follow its XML/HTML graphic binding or locate the figure on the publisher
PDF page. Image filenames and download order can include graphical abstracts or
supplements and need not match figure numbers. Inspect the bound image and its
panel labels together with the caption before recording a visual lesson. If a
prior mapping is wrong, preserve the original and correct the locator and affected
comparison; do not silently relabel an unrelated image. Check the exemplar's own
panel/caption consistency as well, and reject any misleading encoding it contains.

In the existing figure notes, briefly identify the reference's scientific job,
panel story, dominant element, reading order, data/relationship encoding, visual
hierarchy, annotation strategy, why its evidence or explanation is persuasive,
and which of those principles can transfer to this study or must be rejected. Map these
to the present question before drawing. Transfer useful relationships, composition
and visual grammar; do not import the source's results or scientific entities just
to fill the scene. A short design sketch and note suffice; do not create another
approval or receipt chain.

Turn that comparison into the actual drawing brief: name the useful feature in the
reference, the corresponding region or relationship in this figure, and how it will
be rendered. Carry this brief into the whole-image prompt and editable construction.
If all planned transfers are just colors, fonts or icons, the mechanism's scientific
composition has not yet been learned. Revisit the reference choice or its structural
deconstruction before spending time on component generation. A simpler final can be
right when the evidence warrants it; state the concrete reason and retain the useful
structure that the present scientific explanation does support.

An abstract, citation, caption, saved filename or reference count is not visual
inspection. If no suitable reference can actually be viewed, record original
design and the access gap honestly. Continue useful work, but do not report the
requested reference-guided figure complete or backfill a reference after drawing.

## Explore the complete explanatory composition before component production

First distinguish the scientific role: a biological mechanism explains processes;
a graphical abstract summarizes a paper; a workflow shows procedures; an
architecture shows the actual system. One can combine roles when the visual
narrative remains clear, but a comparison flowchart does not satisfy a requested
biological mechanism simply because it contains an organ icon.

For a new or substantially redesigned mechanism/conceptual figure, after viewing
strong references, use GPT Image 2 to explore the COMPLETE composition first.
The concept draft establishes the reading order, spatial relationships, dominant
structures, magnified details and visual language. Inspect it before investing
in isolated organs. Private records and patient details do not enter generation;
concept exploration can use a public reference and a generic scientific description.
Keep the concept image with the working sources, explicitly as a design prototype.
Before component production, distinguish useful structure to retain from scientific
errors to correct. In the existing brief, carry supported compartments, distinct
entities, measurement layers and cross-scale links into the final composition
where they explain the system. Correct an erroneous label or arrow locally; it
does not justify deleting the surrounding valid process. Each entity has a distinct
meaning/visual key, each relation has named source and destination and the right
direction, and any formula has its variable and unit convention. Do not leave an
observed error only in notes while presenting the uncorrected prototype as final.
Keep necessary biological limitations in the caption; they do not replace accurate
arrows, compartments or coherent internal structures.

Specify relationship meanings before choosing arrow styles. An observed association
can use an undirected connector; a literature-supported process needs a directed,
attributed relation; a measurement/readout link identifies what was measured; a
brace or formula identifies an index's components. Do not render index construction
or a measured association as a causal pathway. Adapt these encodings to the domain
and give a concise key when multiple relation types coexist.

Reconstruct in working passes, reusing the existing brief and source:

- **Structural skeleton:** place the supported compartments/modules and their
  internal objects; connect explicit endpoints before adding paragraphs or result
  cards. Mark each retained reference/concept feature at its destination in this
  source. A text label naming a process does not depict that process.
- **Evidence and semantics:** bind measured marks to supplied results; give
  entities distinct visual identities and separate process, measurement and
  mathematical links. Remove prototype inventions, repairing their local relation
  without discarding valid surrounding structure. A generated whole scene remains
  a prototype; recover bounded objects, not a wallpaper under opaque text boxes.
- **Construction and review:** add editable labels, keys and concise annotations,
  then compare the complete reference, concept and reconstruction. Check visible
  endpoints and occlusion in exported pixels, including drawing order: an arrow
  that exists in the source may be hidden beneath a compartment or image.
  Keep containers and their plots in a consistent coordinate frame, checking
  that exported axes, labels and annotations remain within the intended region.

Editable scientific layers do not require every
organic object or texture to be manually vectorized. Preserve the useful detail
and visual quality of the chosen concept; a crude native substitute is not an
upgrade merely because it is editable. Keep suitable bounded assets replaceable
inside the native composition. Reuse valid source images; generate additional
bounded, text-free components only where native drawing is inadequate. Match their
viewpoint, palette, edge style and shading to the whole composition. The concept
image is not a statistical result or an editable final. Existing excellent figures
and small local corrections do not need needless whole-image regeneration.
If the image tool is unavailable, draw a complete native concept sketch and state
the limitation; do not label component-only work as a whole-image generation.

A purely geometric method/system architecture whose modules and connections can
be expressed faithfully with native vectors may start with a complete native
concept instead. Retain the same reference deconstruction and whole-composition
inspection. This does not replace GPT Image 2 with another model for organic
mechanism exploration, or waive the whole-concept step for a new mechanism.

For an iterative algorithm, read the actual methods/code for opportunities to
explain time and state: which prior prediction is checked against which later
observation, when the model updates and issues its next prediction, and how a
new sample/state feeds execution and the next measurement. If the method uses
continuous mixture weights, show the contributing distributions and weight meaning
without turning them into an exclusive switch. If probabilities become quotas,
candidates and accepted points, expose the consequential transformations. Distinguish
normal updates, fallback and terminal regimes where their rules differ, retaining
valid identity encodings. These are applicable design opportunities, not features
to invent in every architecture. A static method need not acquire a feedback loop.
Illustrative states must remain distinct from recorded trajectories; missing
outputs cannot be filled by synthetic points or an unauthorized rerun.

For other architectures, expose the actual input representation, transformations,
branch/fusion rules, repeat counts and output decision where they carry the method.
For example, a real attention block may need token construction, attention,
residual/normalization and output heads; these are source-dependent opportunities,
not a template to impose on other models. Link a supporting result panel to the
specific module it examines only when that evidence already exists under scope.

Compare the chosen reference, the whole concept and the reconstructed final at the
same intended viewing scale. Check the planned transfers in the actual pixels:
which internal detail, connected process or reading hierarchy survived construction,
which was lost, and whether that loss has a scientific reason. Restore useful detail
lost merely because native drawing was easier with boxes. Fix scientific errors in
the concept while preserving its good composition; editability and scientific caution
are not reasons to replace a coherent explanation with an organ-and-box flowchart.
Keep this brief gap-to-change note with the existing sources, not in the manuscript.

Make the first review about the whole scientific explanation: connected structures,
meaningful interior detail, explicit relationship endpoints and consistent style.
Avoid independent pastel cards that sever a continuous process, unlabeled decorative
cells, a realistic organ pasted into unrelated clip art, or a large disclaimer/AUC
block replacing mechanism content. Preserve concise necessary boundaries in the
legend; use a small in-figure key to distinguish observed, literature-supported
and hypothetical relationships.

## Build the scientific composition, then choose tools by component

Define the figure's one-sentence message and the role of each region. Preserve
useful scientific complexity: expose the actual module internals, data structure,
comparison layers and connections between methods and results. Make the dominant
panel larger when that helps; use shared legends, aligned small multiples, nested
scales and routing space to organize supporting detail. Solve crowding by changing
layout or removing redundant wording, not by discarding valid evidence or shrinking
all labels until they technically fit. Ground data
marks in current results, architecture in the actual methods/code, and mechanism
entities and arrows in cited knowledge or explicit hypotheses. Preserve the
difference between what this study measured and a literature-based interpretation.
Lay out the actual system and the relationships the reader should understand;
generic boxes plus a decorative organ do not replace this design work.

Route individual components rather than classifying the whole figure as either
"data" or "image generation":

| Component | Producing method |
|---|---|
| Measured values, axes, intervals, distributions | Deterministic plotting from existing authorized results; retain data and script |
| Labels, arrows, boundaries, legends, regular geometry and overlays | Editable native vectors/text using the appropriate plotting, SVG or document tool |
| Existing microscopy, structure or other authoritative scientific image | Preserve the verified source asset and its provenance; do not synthesize a replacement observation |
| A difficult-to-draw anatomical/organic object or texture needed by the planned explanation | Generate only that bounded, text-free component with GPT Image 2, then place it in the native composition |

Choose component generation because of a concrete drawing need after reference
deconstruction, not just because the figure is called a mechanism. Use the planned
viewpoint, anatomy, palette and region dimensions to guide the component. Retain
the actual prompt and returned asset in this task's paper sources immediately;
copy generated files out of tool/session caches and record their task-relative
paths and roles with the producing source. A later host must be able to find
the original assets from the task alone, without recovering another chat.
Reference retained assets from producing scripts using task-relative paths,
not a previous host's working directory or generated-image cache. Before
handoff, check that those dependencies exist in the task and that the prompt
belongs to the actual generated asset. Preserve an adequate component when
revising labels or relations; native editability does not require replacing
fine organic artwork with a crude vector substitute.
Public reference crops may guide generation
when lawfully usable; private materials and patient information stay local.

Whole-figure concept generation above is design exploration. Do not generate a
finished data plot or call a complete labeled raster an editable component. Assemble the figure with native labels,
connectors and measured marks. A programmatically assembled figure can still fail
the scientific composition or visual-learning requirement; this tool choice alone
is not a quality pass. Preserve source assets and the actual SVG/plotting/document
source, then export the required PDF/SVG/PNG at the intended physical size.

## Design data figures from the supported findings

Before choosing chart types, identify the primary finding and supporting jobs:
actual distributions and sample sizes, relationships between components,
comparators, meaningful heterogeneity, uncertainty and robustness where available.
Existing distributions or authorized individual results can support a more
informative view than summary bars. Availability of raw data alone is not
authorization to calculate new results in `materials_only`. Unknown clinical follow-up
does not prevent improving an honest distribution or uncertainty display.

Inspect a reused drawing script before running it: it may also clean/deduplicate
records, fit models or recompute study statistics. In `materials_only`, reuse
frozen results or editable figure layers and separate pure drawing from those
analysis steps. Geometry, label placement and unit-preserving axis transforms
of existing results remain ordinary figure work; no new proof form is needed.

Choose a grammar for the comparison it makes inspectable. For example, paired
unit-level points against an identity line reveal consistency that an average
can hide; a complete ranked landscape shows the spread beyond selected winners;
an aligned local output and source-grounded module explain where a method acts.
Use these structures only when the corresponding retained results exist; they
are design opportunities, not instructions to compute missing analyses.

Use one dominant panel and supporting panels where it makes the argument clearer.
Share legends and stable color meanings, align comparable axes, label the result
that matters, and remove redundant numeric inventories across tables and figures.
Propose alternatives with a meaningful difference in scientific communication,
not just rotated plots or reordered boxes. Keep panels readable at manuscript size.
Do not add irrelevant endpoints, unsupported significance, survival curves or
model-performance claims just to increase complexity.

Apply the reference shortlist to these scientific jobs before fixing plot types.
A shared mark (points, bars or forest intervals) alone is not a structural match.
Carry useful nested detail or linked scales into the supported evidence; keep a
simple plot when it communicates the question fully. Complexity is justified by
the information it reveals, not by the reference's prestige.

## Let the user compare and decide per figure

Apply the panel/assembly choice sequence to actual requested user decisions.
Reuse adopted selections and explicit delegated-choice authority. Do not create
another choice for each spacing, typo, color or export correction already covered
by the request. A changed scientific combination or a pending revise/reject
decision still follows the relevant public choice; never call an Agent repair
a new user confirmation.

Use the existing public decision and feedback interfaces. Split a complex figure
into selectable subfigure steps when the user is choosing panels; each step has
one scientific panel, its relevant reference and alternatives. Follow these with
a whole-figure assembly check. Do not split a single inseparable pathway merely
to create more steps. Provide a descriptive figure name and
question, a current version when available, genuine alternative designs of the
same evidence, and the relevant original-reference/current comparison. Use exact
task-relative `files` paths accepted by the current schema; do not invent fields.
Keep source components and alternate export formats out of the main choice set.
With the current public 1.1 schema, put explicit presentation in payload.figure:
figure_id and label are stable for the scientific figure; panel_id/panel_label
identify a subfigure. Omit panel_id for the integrated whole-figure check.
reference_files and current_files contain exact discovered task-relative paths;
comparison_files contains the rendered, readable side-by-side reference/candidate
comparison. Supply all three lists explicitly, using an empty current_files list
for a new figure. If no suitable reference is available, leave reference_files and
comparison_files empty and explain the concrete original-design/access situation
in reference_note; this does not fulfill requested reference-guided work.
Declare panel_ids on every new request: all independently selectable panel IDs for
a composite figure, or [] for a whole inseparable figure. Start each selectable
panel with the complete plan, preserve the plan across its choices, then request
the integrated figure after all panel choices are adopted/kept. The product can
check this declared plan; it cannot discover undeclared scientific panels from pixels.
allowed_decisions[].files contain the candidate being chosen; use explicit action
keep, adopt, revise or reject, especially with custom option IDs. Use
replaces_decision_id for a revised version of the same selectable unit.
Read the actual schema first. Metadata is presentation over the same decisions,
not a new state store. Do not infer this identity from shared material sources,
filenames or candidate letters; old records with no identity remain independent.

Offer keep, adopt a candidate and request changes as applicable. A keep or change
option should still link the image it concerns, so selecting it does not remove
the user's visual context. State each option's scientific and visual difference.
A comparison PDF contains both sides simultaneously at readable scale; a link to
a reference in another tab is not equivalent. Reference and comparison images are
never selectable as this study's result. A saved revise/reject choice does not
adopt a panel or unlock the integrated selection. When a user asks for changes,
apply their specific comments
before presenting the revised figure; do not count that resolution as adoption.
Use a short scientific title and question for each choice. Put longer methods,
limits and design reasoning in its description or notes, without hiding a fact
needed to make the choice. Label a prototype, component sheet, current figure and
revised candidate by their actual roles; include only genuine candidates in the
choice. A saved choice with requested changes is still work to resolve.
Preserve decisions for unaffected figures and ensure the manuscript uses the
chosen revision. Do not infer a final scientific choice from file sorting or names.

## Preserve and complete a figure revision

Before changing an existing figure or its producing script, retain its current
exports, editable sources, necessary source assets, caption and exact result inputs
in a named revision within the same paper. Reuse an existing version mechanism;
otherwise `scripts/preserve_figure_revision.py --paper-root <paper> --revision <id>
--file <paper-relative-file> ...` preserves an explicit file set without overwriting
an earlier revision. It copies only named files, never discovers private inputs or
changes task/choice state. Keep the entire input set stable during preservation;
coordinate with its writer or wait for an active export to finish. The helper
checks for changed inputs but does not lock other processes. Preserve relative
dependency paths and include necessary referenced assets explicitly. Do not put
the archive among active candidate files.
Do not claim pre-existing overwritten bytes can be recovered from a new snapshot.

Correct the producing source and export a new candidate path. Read computed
annotations from the retained results where practical rather than duplicating
mutable statistics as literals. Regenerating an unchanged figure is not another
improvement round. Keep one concise issue-to-change map in the existing notes,
including any comment still unresolved and the exact current output. Use the same
public decision identity and replacement field for a revised selectable unit.
Only invalidate choices whose actual content or meaning changed.

After each substantive revision, open the new output, not a cached preview of its
predecessor. For a local change, also compare the retained old figure with the
new one, checking both the requested edit and unaffected regions, marks, scales
and labels. This old/new regression comparison complements reference/current
learning; neither replaces the other. Complete the figure, its comparison and the affected manuscript
placement before moving on to unrelated reference searches or format polishing.
A component atlas is an intermediate asset, not a revised complete mechanism.
Reuse a valid concept and existing components for local corrections; repeating
whole-image generation is unnecessary when native reconstruction fixes the issue.

## Compare what was actually produced

Render the final figure at its actual manuscript column/page size. Check scientific
encoding before optional polish: when countable marks illustrate a stated fraction,
verify counts, denominator, preserved totals and category meanings against the source.
Generate counts and labels from the same parameters where practical. "Schematic, not
measured" permits illustrative positions, not a misleading numerical relationship;
use consistent illustrative counts or a clearly non-counted encoding. Do not turn
this into a point quota for unrelated conceptual drawings.

Trace each scientifically distinct branch through its visible destination and any
required return path. Compare the same relation in source, figure and caption:
scoring vs admission vs fallback, thresholds, time indices, normal vs exceptional
scope, and which population a percentage describes must not change. Simplification
may omit detail, but must not replace it with a different rule. Labels belong to the
object they describe; lines must not cross meaningful text. Inspect page
view and close view: scientific hierarchy, entity identity, meaningful detail,
relationship topology and arrow endpoints, reading order, palette, text, spacing,
line weights and raster resolution. Prefer white or transparent scientific canvases
unless the user or venue calls for another justified presentation.
Do not remove useful complexity merely to make a diagram easy to draw, or add
unsupported molecular detail merely to make it look sophisticated.

Check the exported physical width/height and smallest labels against the intended
venue placement. For a raster region, calculate effective resolution from its
pixels divided by its printed width in inches; zooming a small source does not add
detail. Use native text/vectors or obtain a better source when needed. A low-resolution
reference may guide topology without being suitable as a final embedded asset.

Place the relevant original reference and current figure side by side with clear
attribution and readable scale. The mechanism comparison must contain an actual
mechanism reference; statistical comparisons belong with statistical figures.
Label external published references with source and figure locator, and this
study's earlier version as "Previous study figure". External-reference/current
comparison tests design learning; old/new comparison tests the requested repair
and regressions. Never merge these roles under "Reference / existing figure" or
use an old/new sheet as evidence of published-reference learning.
Keep each complete reference and current figure legible in the comparison.
For a comparison of existing local exports, `scripts/figure_comparison.py`
can place PDF/SVG/raster views side by side and export the same reference view.
Run from this task's paper directory and invoke the helper by its absolute
installed-Skill path:

```bash
python "<installed-skill>/scripts/figure_comparison.py" --output figures/comparison.pdf \
  --view "Published exemplar: author, year, Fig. X" sources/reference-paper.pdf 2 \
  --view "Previous study figure" figures/current.pdf 1 \
  --view "Candidate" figures/candidate.svg 1 \
  --clip 1 40 30 560 340 --export-view 1 figures/reference-view.pdf
```

Page numbers and view indices start at 1. Replace the example crop with explicitly
inspected coordinates: x0 y0 x1 y1 in points (1/72 inch), from the displayed page's
top left after rotation/CropBox, x right and y down. Omit --clip for a whole page.
The exported reference-view.pdf can populate reference_files; comparison.pdf can
populate comparison_files. Both reuse the same labelled view, preserving vectors,
proportions, source page and crop coordinates. Invalid inputs or conflicting output
paths fail before writing; existing files are never overwritten. Inspect the output
and retain original sources: cropping is a view, not redaction or automatic figure
selection. The helper does not run analyses or judge quality; it remains optional.
Size the panels to their actual aspect ratios and label density; do not shrink
several figures into small fixed rectangles inside a largely empty page. Prefer
native vector PDF/SVG content when available. For dense figures add an enlarged
detail view alongside the complete context, clearly identify the crop and retain
the original; a crop alone must not conceal a contradictory panel or caption.
Open the rendered comparison and confirm labels can actually be read at the
intended viewing size. Preserve a direct original/current file view for zooming.
Explain specific retained/adapted features and unresolved differences in the same
notes. Correct weak component integration, missing relationships or generic
flowchart substitution before treating reference learning as fulfilled.

First state the remaining gap against the chosen exemplar: missing evidence
layers, lost structure, weak hierarchy, ambiguous connections or insufficient
readability. If useful complexity was lost, revise the composition before minor
cosmetic polish. Check both the number of distinct supported visual arguments in
the paper and the depth inside each figure; neither panel count nor absence of
clipping proves that the requested figure quality was reached.

Before acceptance, the reviewer names the largest remaining difference from the
strong reference and determines whether it is justified by evidence or is a
repairable design weakness. Correct the latter in the producing source and
inspect the updated pixels. A user's preferred option remains a preference, not
automatic expert approval. Do not substitute scores or a form for this comparison.

The independent reviewer opens both sides and the manuscript page, checks actual
sources and captions, and assesses the requested scientific and visual quality.
If the reference was never viewed, or the output did not apply its useful design
principles, record that concrete failure even if numbers, citations, image files
and render checks pass. Repair the affected figure and review it again; do not
carry a statistical figure's approval over to a mechanism or graphical abstract.