figure-style

# Publication-Grade Figure Rules

*A checklist for correct, legible, internally-consistent scientific figures. This
skill does not impose a visual house style — frame, font, and palette are
parameters. Load it and call `apply_figure_style()` before any plot.*

## §0 Scope

§1–§3, §8, and §9 are **correctness** — they apply to every plot, in every
context, and have no aesthetic content. §4–§7 are **guidance** — defaults that
produce a clean result but that a deliberate alternative can override
(individual rules inside §4–§7 that state a factual/perceptual invariant — e.g.
§4.4 semantic-zero centring, §4.5 CVD, §6.9 leader anchoring — still bind). On
its own, this skill is the inner tier (make one plot good); `figure-composer`
and `paper-narrative` supply multi-panel and whole-paper context.

---

## §1 Data fidelity & self-consistency

**1.1 Excluded rows.** A row marked excluded or flagged in the source data is
either omitted entirely or drawn with a visually distinct open/hatched marker
and named in the key. It **never** enters a summary statistic plotted alongside
the included rows.

**1.2 Comparable conditions only.** Arms measured under non-comparable
conditions (different N, epoch budget, initialisation, protocol) are not plotted
as visual peers. Separate them with a facet break or a marker on the label, and
state the difference once in the caption.

**1.3 Self-consistency.** Every key, threshold, and title inside the figure must
be satisfied by every plotted row. Before saving, walk each categorical outcome
label back to the rule that defines it; if a row's value contradicts its label
or the title, the figure is wrong, not the data.

**1.4 Claim-titles must be true.** A sentence-title (§5.1) is tested against
every category on the axis before rendering. If any contradicts it, qualify the
title ("on 3 of 4 pairs") or downgrade it to a description.

**1.5 State n and what was held fixed.** Every panel that draws a summary mark
states `n` and the unit of replication, and every small-multiple that holds a
variable fixed states the fixed value — in the panel or, when §2 budget is
tight, in the caption.

**1.6 Reference structure is reference.** A tree, ordering, or topology drawn as
*context* (a scale bar, a category strip) uses an established reference, not
one inferred from the plotted data. Infer the structure only when the structure
*is* the result.

**1.7 One number per claim.** A quantitative claim (runtime, accuracy, count)
has exactly one canonical value across every panel, caption, and the abstract.
Define what it measures and use that value everywhere.

---

## §2 Label economy — floor and ceiling

The figure shows the pattern; the **caption** carries the context. Design for a
general scientific reader, not the author.

**2.1 Floor (non-removable).** Every distinct mark, series, glyph, or comparator
must be identifiable from the figure alone. The caption explains *why it
matters*, not *what it is*. A label is non-removable if deleting it leaves a
reader asking "what is that?"; it is removable only if the question becomes "why
is that there?". Comparator labels name the thing ("prior method", "no joint
training"), never a bare role word ("baseline", "previous"). Any term a general
scientist can't parse gets a one-word gloss.

**2.2 Ceiling.** Per panel: title + axis labels + tick labels + series identity
(labelled once per row of small multiples) + at most 2–3 result annotations.
Count the strings; >6 beyond axes/ticks means you're over. The ceiling counts
*narrative* annotations (callouts, value labels, brackets) — identity labels are
floor, not budget.

**2.3 Move to the caption:** n=, what's-held-fixed, abbreviation expansions,
non-comparable footnotes, exclusion rationale, methodological caveats.

**2.4 Titles are takeaways.** A reader seeing only the title knows what the
panel shows. "Robust to gene dropout" passes; "Fewer genes" fails. Test: read it
aloud cold — if the listener asks "fewer genes *what*?", rewrite. For a row of
small multiples that vary one thing, drop per-panel titles for one row-header.

**2.5 Value-on-mark only for the headline number** — the one a reader would
quote. Everything else is read off the axis.

**2.6 When in doubt, delete the label and re-read.** If the message survives, it
stays deleted.

---

## §3 Axes, scales, small multiples

**3.1 Axis padding.** Axis limits clear the data by ≥ one marker radius on every
side; markers and text never touch a spine. `ax.margins(0.04)` after plotting,
or extend the limit past any annotation.

**3.2 Axis breaks over wasted range.** When data occupy <40 % of an axis, break
the axis or start it at the data floor with a clear non-zero tick. Never draw a
reference line, threshold, or annotation inside a broken-axis gap — the gap has
no coordinate.

**3.3 Log axes get human-readable ticks** — `10²`, `10³`, or `1k / 10k / 100k`,
not raw exponents. **Never** draw filled bars on a log-scaled value axis (bar
length encodes ratio to an arbitrary floor); use points + median tick instead.

**3.4 Shared axes across small multiples.** A row or column of small multiples
shows tick labels once (leftmost / bottommost panel); interior panels keep ticks
but drop labels. When the panels share a y-axis and differ only by x-variable,
render them as abutting subplots (`wspace≤0.06`) with one row-header title.

**3.5 Fill the box.** A panel's data envelope occupies ≥75 % of its allotted
rectangle. If a panel's natural aspect leaves dead bands, reshape the grid
(rowspan, stacked complementary panels) — don't pad the panel.

**3.6 Direction of goodness.** When higher- or lower-is-better is not obvious
from the axis label, place a small upright cue ("higher = better") in the
margin — once per row of panels, never per panel, and never only in the caption.
A directional glyph embedded in rotated text rotates with it; set the cue
upright.

**3.7 Physical width.** A single-row figure at 300 dpi fits the venue's
double-column width. Ad