FOUR SCIENTIFIC FIGURES: VECTOR, PRINT PNG, AND REPRODUCIBLE SOURCE
Prepared 6 October 2026

Contents
--------
Figure_1.pdf/.svg/.png: continuation-specific paired audit schematic
Figure_2.pdf/.svg/.png: DMPC continuation gaps, discordance and target-matched scores
Figure_3.pdf/.svg/.png: rolling-deployment shortfall frequency and severity
Figure_4.pdf/.svg/.png: availability, scoring-threshold and obligation-level checks
build_figures.py: deterministic plotting script
requirements.txt: exact plotting package versions used
PROVENANCE.json: manuscript/archive/input hashes and source lineage
plotted_values.json: exact numerical values shown by the plotting script
MANIFEST_SHA256.csv: package file integrity hashes
source_data: supplied in the data/ folder
qa/: PDF render previews, text-bound checks, and validation summary

Use
---
Use PDF when the journal accepts vector artwork. SVG is editable vector artwork.
PNG copies are suitable for Word and raster upload fields. Each PNG is exactly
2100 pixels wide with 300 dpi metadata, corresponding to 7 inches / 177.8 mm.
Heights: Figure 1 5.20 in; Figure 2 6.05 in; Figure 3 3.00 in; Figure 4 3.50 in.
Do not increase figure count: these files replace Figures 1-4.

Regenerate from an extracted copy
-------------------------------
python build_figures.py
or: python build_figures.py --output-dir /path/to/new/output

The script reads only bundled CSV/JSON data. It performs no experiment, model
fit, controller rerun, bootstrap resampling, web retrieval, or image generation.
The input hashes, primary DMPC gaps, four discordance pairs, and agreement of the
full-precision obligation results with the archived three-decimal table are
checked before output. Outputs use embedded TrueType fonts in PDF and editable
DejaVu text in SVG. For exact byte reproduction, use the package versions in
requirements.txt and the original runtime recorded in PROVENANCE.json.

Scientific source and handling
------------------------------
The active four figure panels were checked against
Scientific_Reports_Final_Submission.docx, not an older Round9 article. In that
DOCX, the active image relationships point to image5.png through image8.png;
older unused image1.png through image4.png remain in the DOCX archive.

Eleven bundled data files are byte-identical to their entries in
Supplementary_Data_1_Core_Reproducibility_Final.zip. The twelfth file contains
the independent 6 October obligation-level regeneration at full precision.
Every one of its 12 rows agrees with the original frozen obligation-level CSV
at the CSV's three-decimal precision. File hashes and archive-relative source
paths are in PROVENANCE.json.

Figure 1 fixes the original overlapping boxes/labels and removes the ambiguous
alternating replan/hold depiction. Every deployed interval replans, at the
10-minute cadence. The same pre-action state, exact first command, realized wind
path and obligation are preserved. B and P are the two named reference
continuations; D is rolling deployment. Y_c abbreviates the article's Y_t(c)
for one paired origin. The factor 100 expresses mean paired gaps in percentage
points. Target-matched Brier scoring is visually separated from event gaps.

Figure 2 retains the four-panel comparison, the asymmetric 95% seven-day block
percentile ranges (2,000 resamples), exact discordance counts, and four Brier
scores. Blue identifies FULL and orange identifies the origin-fixed plan.
Filled versus hatched bars distinguish estimates and realized targets without
introducing an extra color. Cohort sample sizes and the conditional nature of
uncertainty remain as described in the article and its legend.

Figure 3 uses unrounded frozen operational endpoints and aligned scales across
both cohort panels. It does not assert an overall controller ranking.

Figure 4 preserves all three latest panels, including full DMPC obligation-level
reruns in panel c. Color identifies the reference; solid filled circles identify
UK development, and dashed open squares identify the Australian stress test.
Shaded percentile ranges are hatched for Australia. Lines join tested settings
and do not imply experimental evidence for untested intermediate settings.

Quality checks
--------------
All four PNG exports and all four independently rendered PDF pages were visually
inspected. Text is inside the figure canvas; no overlap or clipping was found.
All PDFs contain vector graphics with embedded fonts and no raster images.
All SVGs contain vector artwork rather than embedded raster figure images.
PNG size/dpi and the 7-inch PDF page width were verified. Reproduction in a
separate output directory is checked in qa/validation.json.

This package replaces presentation assets only. It does not change the frozen
scientific outputs or establish new calibration, controller-superiority,
field-reliability, causal-mechanism, or safety claims.
