Three-family Broad-Latin deterministic italicification benchmark

Purpose

This clean-room benchmark evaluates the experimental Glyphs MCP italic first-pass workflow in Inter, Noto Sans, and IBM Plex Sans. It compares three deterministic candidates generated from the same Roman outlines:

Raw is a pure shear.
Partial compensation is the legacy cursivy result key in this headless benchmark. It applies the repository's pure-Python stem correction at strength 0.35; it does not invoke the Glyphs Cursivy filter.
Balanced uses curve interpolation strength 0.75 followed by the same deterministic correction engine at full strength.

The generated outlines are design-assistance evidence, not finished italics. The official italics are qualitative references containing deliberate drawing, rhythm, spacing, component, and character-form decisions that a mechanical first pass should not attempt to reproduce automatically.

Fixed coverage

The committed manifest fixes 543 Unicode values and the Inter/Noto Sans source glyph names. IBM Plex Sans is resolved by Unicode from its pinned Regular and Italic UFOs; 391 of the 543 values are present in both. The JSON evidence records all 152 unavailable Plex entries rather than silently dropping them.

Group	Fixed manifest	Inter	Noto Sans	IBM Plex Sans
Basic Latin	95	95	95	95
Latin-1	91	91	91	91
Latin Extended through `U+024F`	259	259	259	158
General Punctuation	48	48	48	16
Currency Symbols	23	23	23	14
Letterlike Symbols	10	10	10	3
Number Forms through `U+218F`	17	17	17	14
Benchmarked total	543	543	543	391

Two historical Inter/Noto Sans source-name differences are mapped by Unicode: Tcommaaccent/Tcedilla at U+0162 and Iota/Iota-latin at U+0196.

Result: deterministic Balanced promoted

Balanced is the deterministic geometry winner in every family. It restores the measured source width of accepted conservative stem pairs and has the lowest equal-family mean error. It passes every safety gate and is promoted as the recommended experimental mode. The omitted slant_mode default remains Cursivy for backward compatibility.

Family	Topology	Source unchanged	Compensated pairs	Raw error	Partial error	Balanced error	Anchor error
Inter v4.1	543/543	543/543	135	2.7812	1.8078	<0.00000000000004	0
Noto Sans	543/543	543/543	102	1.2912	0.8393	<0.00000000000002	0
IBM Plex Sans	391/391	391/391	102	1.2607	0.8194	<0.00000000000002	0
Equal-family mean	—	—	339	1.7777	1.1555	<0.00000000000002	—

Balanced reduced mean source-width error by more than 99.99% versus both alternatives in each family. This near-zero result applies only to confidently accepted straight-side pairs. It does not measure optical completion or similarity to a deliberately drawn italic.

The review also found component transforms for which slanting a component locally and then applying its component matrix (L × S) differs from shearing the complete Roman construction (S × L). Deterministic Balanced now blocks those glyphs before generation:

Family	Affected glyphs	Reflected cases	Other non-commuting cases	Blocked	Safety gate
Inter v4.1	23	17	6 vertically scaled marks	23	Pass
Noto Sans	2	2	0	2	Pass
IBM Plex Sans	0	0	0	0	Pass

This explains the earlier d/q observation. Inter constructs d from a horizontally reflected b and q from a reflected p; local slant followed by reflection reverses their apparent direction. IBM Plex Sans draws d, p, and q directly, so those glyphs lean in the expected direction. The problem is construction-dependent, not a negative angle being supplied to selected glyphs.

The affected Inter rows are d, q, dcaron, dcroat, Ohungarumlaut, ohungarumlaut, Uhungarumlaut, uhungarumlaut, dtopbar, Tonetwo, dzcaron, Udieresismacron, udieresismacron, Adieresismacron, adieresismacron, Adotmacron, adotmacron, dz, dcurl, quotereversed, paragraphreversed, reversedsemicolon, and dong. Noto Sans is affected at paragraphreversed and reversedsemicolon.

This construction issue can also affect any method that slants component sources before applying a non-commuting transform. Balanced does not silently omit the glyphs: the reviewed batch is blocked, and the designer may explicitly rerun with skip_glyphs after deciding how to rebuild or draw those forms.

Raster review and mode separation

The 543 Inter glyphs, 543 Noto Sans glyphs, and 391 Plex glyphs were rendered at 2× scale in pages of at most 64 glyphs. The full paginated evidence and per-glyph JSON remain in the ignored .cache/italic-benchmark/results-broad-deterministic directory.

The combined 144-DPI story sheet is a deterministic 4800 × 6424 PNG. It shows capacity examples, forms that need deliberate italic redrawing, blocked component constructions, Balanced-versus-Raw overlays, and Balanced-versus-official overlays with magnified insets. Click it for the full-resolution image.

The committed 144-DPI audit sheets prioritize component-direction risks, accepted-pair glyphs, representative categories, and the largest qualitative official-italic differences. Click an image for the full 4300 × 23800 contact sheet or 4380 × 23844 difference sheet.

Inter v4.1

Noto Sans

IBM Plex Sans

Dark pixels overlap, blue pixels occur only in the candidate, and coral pixels occur only in the reference. Thin blue and coral lines show the respective advance widths.

Family	Partial vs Raw	Balanced vs Raw	Balanced vs Partial	Balanced vs official
Inter v4.1	0.123% (64/543)	0.269% (86/543)	0.151% (59/543)	7.237%
Noto Sans	0.121% (46/543)	0.205% (51/543)	0.088% (40/543)	29.219%
IBM Plex Sans	0.106% (27/391)	0.213% (50/391)	0.114% (36/391)	19.798%

Values are mean silhouette differences; parentheses give the count of glyphs with at least one different raster pixel. The median generated-mode difference is zero in all families because the conservative correction engine changes only accepted stem pairs. This is why Raw, partial compensation, and Balanced look identical for most glyphs, including many curved forms. With stem_compensation=1.0, the final correction restores accepted stems to their Roman perpendicular width, so it can also neutralize part of the visible intermediate difference controlled by curve_strength.

The deterministic Partial column is the pure-Python correction engine at strength 0.35. It is not the live Glyphs Cursivy filter. Real Cursivy remains host-specific and can vary with the installed Glyphs and Transformations filter version.

Official-italic differences remain qualitative and are not an optimization target. They show where a designer deliberately redrew forms beyond a mechanical emphasis companion.

Reproduction

Keep the pinned source checkouts beneath the ignored .cache/italic-benchmark directory, then run:

python3.12 scripts/benchmark_italic_sans_broad.py \
  --render-scale 2 \
  --page-size 64 \
  --audit-limit 64

The script exits zero when every family passes. Its JSON records coverage, topology, source immutability, anchors, components and transform risks, blocked constructions, bounds, advance widths, detected and skipped stem pairs, width measurements, all four raster comparisons, failures, and runtime.

Sources are pinned at:

IBM Plex's upstream project and all three pinned benchmark sources use the SIL Open Font License 1.1. No font source or binary is committed.

Clean-room provenance

The benchmark exercises the repository's original interpolation and straight-stem correction engine. It does not call, inspect, decompile, or copy Italify or another proprietary implementation. The same conservative detector, correction limits, and promotion rules apply without family-specific optical tuning.

Purpose​

Fixed coverage​

Result: deterministic Balanced promoted​

Raster review and mode separation​

Inter v4.1​

Noto Sans​

IBM Plex Sans​

Reproduction​

Clean-room provenance​