Italic first pass

Experimental

Glyphs MCP can generate guarded Raw, Cursivy, and Balanced first-pass layers from a Roman master. The result is a construction draft for the designer, not a finished or authentic italic design.

Design goal

An italic commonly works as a companion to the Roman, providing contrast and emphasis inside text. The Glyphs MCP workflow has the narrower goal of helping the designer begin that companion efficiently: it establishes a controlled slant, preserves reusable structure, and reports geometry that needs review.

It does not decide the italic's voice. Character construction, optical correction, rhythm, spacing, kerning, alternates, interpolation, and proofing remain design work. In particular, mechanically transformed a, e, f, g, k, v, w, x, y, punctuation, brackets, quotes, and symbols usually need deliberate redrawing.

Modes

• raw applies a deterministic affine shear around the selected origin.
• cursivy uses Glyphs' callable Transformations filter when available. On hosts where that public Python entry point is unavailable, it reports and uses a conservative pure-Python straight-stem fallback.
• balanced is a reproducible pure-Python pipeline. It builds Raw, applies the conservative straight-stem correction at partial strength, interpolates between those compatible coordinates, and then applies the requested final compensation to confidently detected vertical or diagonal stem pairs. It never invokes the Glyphs Transformations filter.

Balanced is the recommended experimental deterministic option. Calls that omit slant_mode continue to use Cursivy for backward compatibility.

Balanced defaults to curve_strength=0.75 and stem_compensation=1.0. Both accept values from 0 through 1. curve_strength=0 is exactly Raw before final compensation, while curve_strength=1 uses the complete deterministic partial-correction candidate. The controls are independent. At stem_compensation=1, accepted stems converge to their Roman perpendicular width, so the final correction can make the intermediate curve_strength difference visually subtle.

Symbols that may stay upright

Unicode identity can identify glyphs that deserve review, but it cannot decide their italic treatment. The workflow therefore reports the tiers below as advisories. It does not silently exclude glyphs, change protected_glyphs, preserve Roman outlines, or apply a zero-angle workaround.

The evidence sample used the 20 highest-ranked families with Regular and Italic styles in the Google Fonts family metadata on July 30, 2026: Roboto, Open Sans, Google Sans, Inter, Montserrat, Poppins, Lato, Roboto Condensed, Arimo, Roboto Mono, Noto Sans, DM Sans, Raleway, Nunito, Nunito Sans, Playfair Display, Rubik, Ubuntu, Kanit, and Merriweather. We compared the filled-outline slant of each Roman and Italic glyph with fontTools StatisticsPen, treated an absolute slant delta below 0.035 as upright, and visually reviewed the results.

Review tier	Unicode values	Evidence and recommended handling
Empirical upright candidates	U+00A4 ¤, U+00A9 ©, U+00AE ®, U+00B0 °, U+212E ℮	¤, ©, ®, and ° stayed upright in 12/20 families; ℮ did so in 11/18. Compare with the Roman before deciding.
Empirical candidates with limited coverage	U+2190 ←, U+2191 ↑, U+2192 →, U+2193 ↓, U+25A0 ■	Each stayed upright in 7/8 supporting families. Treat the result as strong but smaller-sample evidence.
Traditionally upright math operators	U+002B +, U+003C–U+003E <=>, U+00AC ¬, U+00B1 ±, U+00D7 ×, U+00F7 ÷, U+220F ∏, U+2211 ∑, U+2212 −, U+221A √, U+221E ∞, U+222B ∫, U+2248 ≈, U+2260 ≠, U+2264–U+2265 ≤≥	Microsoft's character-design guidance describes math signs as traditionally upright, but popular text-font italics often slant them. Only 4/20 reviewed families kept + upright. Decide from the family's purpose and math support.
Design-dependent forms	U+0023 #, U+002A *, U+0040 @, U+007C |, U+007E ~, U+2022 •, U+2122 ™	Inspect each design. Only 7/20 reviewed families kept @ statistically upright, and some kept its enclosure upright while italicizing the internal letter.

Sources and interpretation

• Microsoft's Latin-1 math character-design standards state the traditional upright convention and give design, alignment, width, and spacing guidance for +, −, =, ≠, <, >, ≤, ≥, ±, ×, ~, ¬, and ≈. This is design guidance, not a requirement that every text italic preserve every operator unchanged.
• The Microsoft OpenType MATH table specification distinguishes slanted characters from straight operators or delimiters when explaining italic correction. That model supports separate treatment, but it describes mathematical layout rather than general-purpose text italics.
• Unicode Technical Report #25, Unicode Support for Mathematics separates operators, identifiers, delimiters, arrows, and geometric shapes. It explicitly calls for an upright U+220A ∊ and notes that integral slant varies by typographic tradition, so Unicode does not justify a blanket upright rule for every mathematical symbol.
• Google Fonts family metadata supplied the popularity order and available Regular/Italic styles for the dated corpus. The Google Fonts static-font guidance documents its Italic style and negative exported italicAngle convention. Popularity and font binaries can change after the snapshot date.
• fontTools StatisticsPen supplied the area-based glyph slant measurement. Its documentation warns that self-intersecting outlines can affect statistics; visual review remains part of the method.

Before a mechanical pass, the agent resolves scoped glyphs through their Unicode values and shows every match. The designer then chooses whether to include it, explicitly exclude it from that pass, or defer it for manual drawing. An upright symbol remains a separate drawing decision; the italic master keeps its actual italicAngle.

Broad-Latin evidence

The benchmark covers 543 encoded glyphs in both pinned Inter and Noto Sans sources plus the 391 values shared by pinned IBM Plex Sans Regular and Italic UFOs:

Family	Topology	Compensated pairs	Raw error	Partial error	Balanced error	Component-risk glyphs
Inter v4.1	543/543	135	2.7812	1.8078	<0.00000000000004	23
Noto Sans	543/543	102	1.2912	0.8393	<0.00000000000002	2
IBM Plex Sans	391/391	102	1.2607	0.8194	<0.00000000000002	0

Every generated mode preserved topology, review left all source layers unchanged, anchor error was zero, and no unsafe compensation was applied. Balanced exceeded the required 50% source-width improvement over both deterministic alternatives in each family. Recursive component analysis found 23 Inter glyphs and 2 Noto Sans glyphs whose transforms do not commute with the requested shear. Balanced blocked every one before application, so the safety gates pass with zero unsafe component applications. In Inter, this includes d and q, which are built from horizontally reflected components; Plex draws those forms directly and does not show the reversal.

This is evidence for a safer mechanical first pass—not evidence that Balanced resembles a designed italic. Mean Balanced-versus-official silhouette differences were 7.237% for Inter, 29.219% for Noto Sans, and 19.798% for IBM Plex Sans because the official italics include deliberate redrawing, recomposition, and spacing.

Read the complete three-family report. Click the combined story sheet for the full 4800 × 6424, 144-DPI source.

The family-specific audit sheets remain available for closer inspection:

Safety and diagnostics

preview_italic_first_pass_candidate runs the complete transformation on detached layer copies and opens an ephemeral candidate session. It reports topology preservation, the effective backend and strengths, origin pivot, anchor shifts, component placement, component-master conflicts, candidate bounds and metrics, and detected/compensated/skipped stem pairs.

Balanced mode:

• Never changes path count, node order, node types, smooth flags, or curve handles during stem compensation.
• Moves only both endpoints of accepted straight sides.
• Preserves live components and adjusts only component x placement using x += tan(angle) * y.
• Shears copied anchors using x′ = x + tan(angle) * (y - pivotY).
• Recursively inspects component chains and reports each matrix, determinant, shear commutator, and chain.
• Permits translation and commuting linear transforms, including uniform scale.
• Blocks reflection, rotation, non-uniform scale, cycles, unreadable transforms, and explicit component-master mismatches when safety cannot be established within 1e-6.
• Applies a reviewed batch all-or-nothing. It never silently excludes glyphs; the designer can explicitly rerun with skip_glyphs.

Safety is based on actual outlines and component construction, not glyph names or Unicode categories. Protected and design-sensitive forms still receive manual-review warnings without being blocked merely because of their identity.

stem_policy=copy_from_source remains review-only. It reports source stem availability but does not silently edit Font Info; use set_master_stem_metrics for that separate confirmed mutation.

Recommended workflow

1. Verify the source Roman and target italic masters and the target italicAngle.
2. For path-only or transform-safe constructions, compare Balanced with Raw and Cursivy; inspect blocked glyphs, components, anchors, topology, stems, bounds, and metrics.
3. Call preview_italic_first_pass_candidate for the explicit glyph/master scope, then inspect the difference-only Candidate Reporter.
4. Call review_outline_candidate_session; if the batch is blocked, decide whether to rebuild those components or explicitly rerun with skip_glyphs.
5. Call accept_outline_candidate_session with dry_run=true, show the exact reviewed changes, and stop for approval.
6. Apply only the reviewed session and one-time token with confirm=true.
7. Redraw, space, kern, and proof the result as an italic design.

Materialize the session only when the designer wants to edit a normal Glyphs layer before re-review. Direct italic review/apply commands were removed in 1.8; the session fingerprint and review token are the required batch safety boundary.

The tool never saves the font.