Complex scientific plates¶

The three-paper reproduction audit exercised 43 axes, thousands of prediction trajectories and 240,000 response-map samples. The audit was first assembled in the 4.0.0.dev15 development snapshot.

Dense fields: choose the representation¶

import inklet as i

values = [[0.1, 0.4, 0.8], [0.2, 0.6, 1.0]]
colors = i.ramp(['#f5f5dc', '#4b94df', '#f28ca7', '#ffd700'])
heat = i.linear((0, 1))
p = i.panel(60, 40)
p.matrix(values, ramp=colors, scale=heat, vector='seamless')
p.colorbar(ticks=[0, .5, 1])

Seamless mode keeps foreground cell boundaries and the outer extent unchanged. Colors are not quantized. The underpaint only covers antialiasing cracks; it is not interpolation. Every compound path contains at most 512 cells. The field must use opaque colors with zero overlap. Whole-field transparency can be applied to a containing group. At extreme zoom the discrete source cells remain visible.

For the six 200 × 200 paper maps, seamless output was approximately 13 MB and 16 seconds, compared with 47 MB and 28 seconds for the earlier overlapping-cell workaround. PDF size increased from about 1.1 MB to 5.2 MB. These are local measurements of complete figure generation, not universal performance budgets.

Linear interpolation requires explicit raster=True, uniform spacing in the displayed coordinate system, and samples=2 through 16 (default 4). It interpolates values before applying even a nonlinear color scale. Missing samples remain missing throughout the interpolation footprint; values beyond the outer sample centers use constant extension. It does not modify the input array.

Place data rectangles precisely¶

p = i.panel(60, 40).axes(x='time / s', y='response')
p.line([(0, 0), (1, 1)])
placed = p.placed('15mm', '20mm')
area = i.plot_area(placed)  # (15, 20) to (75, 60), up to floating-point rounding

Panel.placed(x, y) uses the data rectangle's top-left, independently of axis labels or legends. Furniture is not scaled. For new grids, the existing facets API owns the shared axes and avoids hand-positioning altogether:

panels = [i.panel(35, 25).line([(0, n / 4), (1, 1)]) for n in range(4)]
grid = i.facets(panels, cols=2, share_x=True, share_y=True,
                x_label='time / s', y_label='response', gap=5, row_gap=8)

Supply panels without pre-existing axes when asking facets to create them. Use axes=False when each panel already owns its furniture.

A document() built from separately added cells, which is how letters and per-cell alignment are set, takes the same single names from doc.label(x=, y=); see one name for a grid.

Keys and labels that follow the plot¶

p = i.panel(60, 40, x=i.log((1, 100)), y=i.log((1, 100)))
p.guide((1, 1), (100, 100), label='k_{S} = k_{T}', at=.3, offset=2,
        label_style={'size': 3, 'font_style': 'italic'},
        stroke='black', stroke_dash=(1, 1))
p.matrix(values, ramp=colors, scale=heat, vector='seamless')
p.colorbar(corner='sw', length=12, thickness=1.5, pad=2, plate=True,
           title='Response', ticks=[0, 1], tick_font_size=2)

key = i.legend([('A', 'red'), ('B', 'blue'), ('C', 'green'), ('D', 'black')],
               columns=2, order='column', col_gap=6, row_gap=2)

Guide angles use displayed coordinates, so logarithmic scales and non-square panels produce the correct slope. at is the fraction along the displayed segment. offset is signed perpendicular clearance in millimetres. The guide label stays attached when the panel dimensions change; collision avoidance is not implied. Draw the matrix first if the guide line must sit above it.

Inset colorbars measure their full furniture, support an optional plate/title, and reject a key that does not fit. corner='auto' searches clear space using the same bounded placement mechanism as legends. Legend order can fill across rows or down columns, with independently measured row and column gaps.

Review provenance, geometry and appearance separately¶

review = i.review_figure(placed, page=i.Rect(0, 0, 90, 75),
                         rules=['OFF_CANVAS', 'TINY_TEXT'])
# Requires the render extra for PNG generation:
# review.save_bundle('out/review', reference='reference.png',
#     sources=[{'path': 'measurements.csv', 'sha256': expected_digest,
#               'url': source_url, 'license': license_url}],
#     panels={'a': placed}, reference_regions={'a': (0.1, 0.2, 0.8, 0.9)},
#     caption='Figure reproduction')

The bundle includes original-art SVG/PDF/PNG, a separate diagnostic overlay, source hashes, panel data/ink rectangles and an HTML reference comparison with a large-view toggle and panel selector. Explicit reference_regions register normalized image rectangles for corresponding reference crops. Without a registration the reference stays whole. Expected hashes are checked before writing. Reference images are copied only into the review bundle, never into scientific exports. The report does not certify scientific correctness or pixel identity.

The reproducible paper recipes are in examples/paper_recreations in the source checkout. Run recreate.py --review-bundle after fetching their pinned source data. The public gallery continues to use original Inklet compositions.