Chart Prompts

Restyle my chart in the "Scaling Law" style described below. Keep my data, chart type, title meaning and units; change only the presentation. Use the charting library this project already uses (plain SVG if there is none). If the library can't express a detail, draw it in a custom layer instead of dropping it. Measurements assume a 960 x 600 chart; scale them proportionally.

Mood: the headline figure of a scaling-laws paper, set at night. Deep midnight blue, log-log axes with decade gridlines, families of runs colored along a magma ramp by model size, a bright white frontier and a dashed fit reaching into compute nobody has tried yet.

Colors (use these exact values):
- Background: #0B1026 (deep midnight blue). Untested region: #141C3D. Equation box: #131A3A with a 1px #3A4A85 border.
- Sequential magma ramp by size, smallest to largest: #7B2382, #B5367A, #DE4968, #F7705C, #FEA873, #FCE5A6. Use more or fewer steps of the same ramp if you have more or fewer sizes; never use it for unordered categories.
- Frontier, fit and prediction marker: #FFFFFF. Decade gridlines: #1C2750. Minor log ticks: #2F3F78 (dim blue). Axis lines: #3A4A85.
- Main text: #E6E9F5. Secondary text: #8F98BD.

Typography (Google Fonts):
- Lexend for the title (600, 27px, -0.3px letter-spacing, stating the finding, with exponents as real superscripts, e.g. "C" followed by a 17px raised "−0.160"), the subtitle (300, 15.5px, #8F98BD) and the equation in the box (400, 18px, white, superscript exponent at 12px).
- Lexend Deca for tick labels (14px; powers of ten written "10" with a 10px superscript exponent), axis titles (13.5px, #8F98BD), small caps labels (10.5px uppercase, 1.4px letter-spacing, #8F98BD), legend labels (12px) and annotations (12.5px white).

Layout:
- Title at y 52 and subtitle at y 80, left at x 96. Plot area from x 96 to 904 and y 118 to 486.
- Equation box (312 x 58, 10px radius) in the empty upper-right corner at x 574, y 132: a small-caps label "POWER-LAW FIT TO FRONTIER (E FIXED)" and the fitted equation, e.g. "L(C) = 1.70 + 1029.9 · C^−0.160", with the numbers computed from the data.
- Size legend directly below it: a small-caps "MODEL SIZE (PARAMETERS)" label and a row of 5px ramp-colored dots, each followed by its size label.
- Footer: source at the bottom-left and a note about the runs and the floor at the bottom-right, both 12px #8F98BD.

Axes and grid:
- Log scale on both axes. Vertical 1px #1C2750 gridlines at every decade of x; horizontal gridlines at a few round y values (2.0, 2.5, 3.0, 4.0, 5.0) with one-decimal labels.
- Minor log ticks at 2 to 9 times each decade: 5px #2F3F78 ticks rising from the x-axis; a few 5px minor ticks on the y-axis too.
- 1px #3A4A85 axis lines on the left and bottom only. Rotated y title "Eval loss L".
- Shade the compute beyond the largest tested run in #141C3D and label it "UNTESTED" in small caps.

Marks:
- Line chart: each family (one model size) is a thin 1.3px line at 90% opacity through its points, with 3.4px ramp-colored dots outlined 1px in the background color. The lower envelope across all families (the compute frontier) is a 2.6px pure white line with round joins, labeled "compute frontier" with a 1px white leader. The fitted power law is a 1.2px white dashed line (5 on, 4 off) at 55% opacity over the tested range and continues as a brighter 1.8px dashed segment (7 on, 5 off) into the untested region, ending in a 5px open white circle labeled "predicted 1.85" above it.
- Bar chart: vertical bars on a log y-axis, one per compute budget, filled with the ramp color of the model size that won that budget, 60% of the band, 2px top radius; the fitted power law drawn as the dashed white line across the bar tops, extending dashed into the untested region.

Avoid: a categorical palette, linear axes, solid fit lines in untested territory, heavy gridlines, glows, and a legend box with a border.

When you're done, render the chart and check that the ramp gets lighter as models get larger, every decade has a gridline and minor ticks, the equation's numbers match the fit you actually computed, and the extrapolated segment is visibly different from the fitted range.