From styling to fidelity

The same synthetic SSP trajectories are shown four ways: Matplotlib defaults, appearance-only “IPCC-ish” styling, an explicit adapted profile, and the stricter ar6-report contract. The point is not that the fourth mini-panel is itself an official IPCC figure; it is that a fidelity claim needs semantic tokens, an explicit profile, audit gates, and disclosed requirements rather than visual resemblance alone.

The comparison is reproducible with examples/visual_comparison.py; PNG and SVG outputs plus the interpretation boundary live in examples/visual_comparison/.

See the reference reproductions first

These figures are regenerated from pinned official IPCC AR6 WGI source repositories, not synthetic demo data. Source commits, physical dimensions, and SHA256 checksums are recorded in the reference gallery and manifest.

Chapter 3 — Figure 3.2b Scatter + fitted relationship.	Chapter 10 — Figure 10.20b Mediterranean station map.
Chapter 2 — Figure 2.3 Paleo CO₂ proxies + uncertainty.	Chapter 6 — Figure 6.18 source Historical + scenario CH₄ emissions.

Why this is different

A lot of “IPCC-style” plotting stops at a diverging palette, a sans-serif font, and some hatching. This repository treats AR6 WGI as a visual system with evidence and failure conditions, not a theme.

• Evidence-backed rules — WGI guides, TSU review comments, official colormaps, chapter code, Atlas guidance, and published reference figures are kept distinct.
• Semantic colour — SSP/RCP colours and variable-specific map palettes carry meaning; they are not a decorative colour cycle.
• Explicit fidelity modes — strict/IPCC-faithful and adapted/IPCC-inspired are separate claims.
• Fail-closed strict mode — no silent fallback to viridis, RdBu, cmocean, or an arbitrary font while still claiming fidelity.
• Method ≠ style — median, 17–83%, 80% agreement, FDR, weighting, and projection remain analysis/reference choices unless evidence says otherwise.
• Auditable outputs — physical dimensions, typography, semantic colours, official-colormap use, provenance, and reference regressions can be checked.

The goal is not to make a plot look vaguely IPCC-ish. The goal is to make the fidelity claim inspectable.

Audit the fidelity claim

The audit layer is designed for both humans and CI. Existing code can keep using audit_figure(fig); new integrations can use the structured audit_figure_report(fig) API or the ar6plot CLI.

A figure script exposes a zero-argument factory returning a Matplotlib Figure:

def make_figure():
    ...
    return fig

Run the machine-checkable contract:

ar6plot audit examples/audit_demo.py --strict-dimensions

Example report:

AR6 fidelity audit — PASS
Profile: ar6-report

PASS  text.unit-convention         axis and annotation unit syntax passed
SKIP  typography.arial             strict font check not requested
PASS  delivery.width               figure width matches an IPCC delivery width
PASS  delivery.height              figure height is within the delivery maximum
PASS  scenario.color.0.0           scenario 'SSP1-2.6' uses the ar6-report semantic colour
PASS  scenario.color.0.1           scenario 'SSP2-4.5' uses the ar6-report semantic colour
PASS  scenario.color.0.2           scenario 'SSP5-8.5' uses the ar6-report semantic colour
SKIP  map.official-colormap        official map-colormap check not requested
SKIP  reference.manual-review      projection, panel geometry, annotation, and scientific method require reference-specific review

Summary: 6 passed, 0 failed, 3 skipped

For CI and other tooling:

ar6plot audit examples/audit_demo.py \
  --strict-dimensions \
  --format json \
  --output outputs/audit.json

Exit codes are deliberate: 0 = all requested machine checks pass, 1 = one or more checks fail, 2 = the audit could not run. Exact reproduction still requires the explicitly reported manual/reference-specific checks.

Quick start

Requires Python 3.11+.

python -m pip install ar6-sciplot

The PyPI distribution is named ar6-sciplot; the Python import remains ipcc_sciplot.

Run a minimal, copy-pasteable example:

import matplotlib.pyplot as plt
import numpy as np

from ipcc_sciplot import audit_figure, axis_label, publication_context, scenario_style

year = np.arange(2015, 2101)
warming = np.linspace(1.1, 2.7, year.size)
style = scenario_style("SSP2-4.5", profile="ar6-report")

with publication_context(width="double", strict_font=False):
    fig, ax = plt.subplots()
    ax.plot(year, warming, color=style.color, label="SSP2-4.5")
    ax.set_xlabel("Year")
    ax.set_ylabel(axis_label("Temperature change", "°C"))
    ax.legend()

issues = audit_figure(
    fig,
    profile="ar6-report",
    strict_font=False,
    strict_dimensions=True,
)

print("fidelity audit:", issues or "passed")
plt.show()

This first run deliberately allows a font substitution. A figure should only be called strictly IPCC-faithful when all strict requirements are satisfied, including Arial where required.

For map/climate workflows:

python -m pip install "ar6-sciplot[climate]"
git clone https://github.com/IPCC-WG1/colormaps.git
export IPCC_WG1_COLORMAPS_DIR=/path/to/colormaps

Strict maps load the official WGI colormap assets directly and intentionally have no generic palette fallback.

Choose the fidelity contract

Two style profiles are explicit and are never silently mixed:

ar6-report       → final-report-era AR6 semantics
wgi-guide-2022   → June-2022 WGI guide update

Execution model

SOURCE → PROFILE → FIGURE CONTRACT → RENDER → AUDIT → REFERENCE

What is encoded

delivery
  ├── 90 / 180 mm print widths
  ├── ≤ 250 mm height
  ├── 9 / 11 pt WGI typography
  ├── 0.5 pt axis grammar
  └── 350 ppi raster master

semantics
  ├── SSP / RCP colours
  ├── WGI generic line colours
  ├── variable-specific official colormaps
  └── report-era vs 2022 profiles

uncertainty
  ├── model agreement
  ├── insufficient data
  └── statistical significance

qa
  ├── text + unit conventions
  ├── semantic colour audit
  ├── physical-size audit
  ├── official-colormap audit
  └── source-data regression gallery

Repository map

ipcc-wg1-scientific-plotting-skill/
├── SKILL.md                         # skill contract + routing
├── scripts/ipcc_sciplot/
│   ├── tokens.py                    # semantic WGI colours + design tokens
│   ├── colormaps.py                 # official WGI asset loader
│   ├── style.py                     # delivery geometry + typography
│   ├── archetypes.py                # recurring figure families
│   ├── maps.py                      # map context + uncertainty layers
│   ├── fidelity.py                  # machine-checkable fidelity audit
│   ├── uncertainty.py               # statistical helpers, not style defaults
│   └── provenance.py                # reproducibility metadata
├── references/
│   ├── SOURCES.md                   # source hierarchy
│   ├── evidence_matrix.md           # rule → evidence → scope → confidence
│   ├── ipcc_visual_grammar.md        # canonical visual grammar
│   ├── figure_archetypes.md          # figure-family rules
│   ├── fidelity_checklist.md         # strict gate
│   └── statistical_rules.md          # method/style separation
├── examples/
│   ├── quickstart.py
│   └── ipcc_reference/              # pinned source-data regressions
├── templates/
│   └── figure_recipe.yaml
└── tests/

Verification

Run the local contract:

ruff check .
pytest -q
python examples/quickstart.py
python scripts/check_figure.py   outputs/quickstart.pdf   --metadata outputs/quickstart.provenance.json

Rebuild the pinned IPCC reference gallery:

python -m pip install cartopy
python examples/ipcc_reference/generate_all.py
git diff --exit-code -- examples/ipcc_reference/outputs

CI covers Python 3.11 and 3.12. The reference workflow regenerates the four official-source cases and fails if the committed gallery drifts.

Evidence model

The source hierarchy is deliberately explicit:

published reference figure
        ↓
WGI visual guidance + TSU review evidence
        ↓
official IPCC-WG1 colour assets
        ↓
chapter / final-figure implementation code
        ↓
Atlas uncertainty guidance
        ↓
general scientific-visualisation practice

Start with:

• source corpus
• evidence matrix
• visual grammar
• figure archetypes
• fidelity checklist

The boundary

This repository does not claim that one Matplotlib theme can represent all of AR6 WGI.

It also does not turn scientific-method choices into fake visual defaults.

17–83% range       ≠ IPCC style
80% agreement      ≠ IPCC style
median             ≠ IPCC style
equal-model weight ≠ IPCC style
Robinson           ≠ universal IPCC projection

Those choices belong to the analysis or the published reference figure.

License and third-party material

Original project code and documentation are licensed under the MIT License.

That license does not relicense IPCC figures, source data, colour assets, chapter code, fonts, or other third-party material referenced or fetched by the reproducibility workflows. See THIRD_PARTY_NOTICES.md and the source corpus before redistributing upstream material.

This is an independent project. It is not an official IPCC product and does not imply IPCC endorsement.