diff --git a/crates/engine/src/cmd/fileio/tests_aiimport.rs b/crates/engine/src/cmd/fileio/tests_aiimport.rs
index 3f0dce5c1..a51c1f590 100644
--- a/crates/engine/src/cmd/fileio/tests_aiimport.rs
+++ b/crates/engine/src/cmd/fileio/tests_aiimport.rs
@@ -535,6 +535,27 @@ fn an_ai_file_whose_pdf_part_is_only_text_still_opens_from_editing_data_that_rea
     check_sample(&s);
 }
 
+#[test]
+fn an_ai_file_whose_pdf_part_is_cropped_to_its_art_opens_from_its_layers() {
+    // A 400 × 200 pt artboard with two red bars across (100, 50)–(300, 150). The PDF part's page
+    // is that art's box, 200 × 100 pt (the editing data's bounding box), not the artboard.
+    let red = "0 1 1 0 1 0 0 Xa\n";
+    let art = ai::layer("Art", &format!("{red}{}f\n{red}{}f\n", ai::rect(100.0, 50.0, 180.0, 150.0), ai::rect(220.0, 50.0, 300.0, 150.0)));
+    let data = ai::editing_data_in([100.0, 50.0, 300.0, 150.0], [0.0, 0.0, 400.0, 200.0], &art);
+    let page = "1 0 0 rg 0 0 80 100 re f 120 0 80 100 re f";
+    let mut s = Session::new();
+    let r = open(&mut s, "cropped.ai", &ai::ai(&compressed(data.as_bytes()), page), json!({})).unwrap();
+    assert!(!r["warnings"].to_string().contains("weren't read"), "{r}");
+    let doc = &s.doc().unwrap().doc;
+    assert_eq!(names(&doc.layers), ["Art"]);
+    assert_eq!(doc.artboards.len(), 1);
+    assert_eq!(doc.artboards[0].rect, vectorcraft_geom::Rect::new(0.0, 0.0, 400.0, 200.0), "the file's artboard, not its page");
+    let bars = doc.layers[0].children().unwrap();
+    assert_eq!(bars.len(), 2);
+    let b = bars[0].geometric_bounds().unwrap();
+    assert!((b.x0 - 100.0).abs() < 0.01 && (b.y0 - 50.0).abs() < 0.01, "where the artboard has it: {b:?}");
+}
+
 #[test]
 fn editing_data_false_opens_only_what_the_page_prints() {
     let mut s = Session::new();
diff --git a/crates/eps/src/import/native/build.rs b/crates/eps/src/import/native/build.rs
index 24001e8f0..a9c0c9bf8 100644
--- a/crates/eps/src/import/native/build.rs
+++ b/crates/eps/src/import/native/build.rs
@@ -79,8 +79,18 @@ pub(super) fn build_from(
         let page_rect = visible.artboards.first().map(|a| a.rect).ok_or("it has no page")?;
         // The layers' art in the document's terms: a Large Canvas editing copy is written at 1/N
         // scale, which `to_doc` carries, so its art and the page reach the same size (#2031).
-        let rect = to_doc.transform_rect_bbox(art_box);
-        if (rect.width() - page_rect.width()).abs() > 0.5 || (rect.height() - page_rect.height()).abs() > 0.5 {
+        let fits = |r: &Rect| (r.width() - page_rect.width()).abs() <= 0.5 && (r.height() - page_rect.height()).abs() <= 0.5;
+        let mut rect = to_doc.transform_rect_bbox(art_box);
+        // A `.ai` whose PDF part's pages are cropped to its art (each page the box of all of it,
+        // the editing data's bounding box) rather than to its artboards: the layers are compared
+        // with the page on that box, and keep their artboards.
+        if matches!(page, Page::Artboard)
+            && !fits(&rect)
+            && let Some(art) = bbox.map(|[a, b, c, d]| to_doc.transform_rect_bbox(Rect::new(a, b, c, d))).filter(fits)
+        {
+            rect = art;
+        }
+        if !fits(&rect) {
             // Both sizes in points as the fields show them, and the editing data's page to the
             // file's as a scale (1:10 is a tenth of its size) when its sides are in proportion and
             // the scale doesn't round to 1:1.
diff --git a/crates/eps/src/import/tests_native.rs b/crates/eps/src/import/tests_native.rs
index 4e1848f4f..56bcf3fbc 100644
--- a/crates/eps/src/import/tests_native.rs
+++ b/crates/eps/src/import/tests_native.rs
@@ -381,11 +381,18 @@ fn an_ai_file_that_cant_be_read_through_its_editing_data_keeps_its_pdf_part() {
     let (d, w, read) = crate::layered_ai(b"nothing", pdf.clone(), vec![], false, false, 1.0);
     assert_eq!(d.layers.len(), pdf.layers.len());
     assert!(w.is_empty() && !read);
-    // Editing data that isn't for this page.
-    let other = native(&layer("Jig", true, &square(10, 10))).replace("%AI3_Cropmarks: 0 0 100 100", "%AI3_Cropmarks: 0 0 50 50");
+    // Editing data that isn't for this page: neither its artboard nor its art's box is the page.
+    let other = native(&layer("Jig", true, &square(10, 10))).replace(" 0 0 100 100", " 0 0 50 50");
     let (d, w, _) = crate::layered_ai(&private(&other), pdf.clone(), vec![], false, false, 1.0);
     assert_eq!(names(&d), names(&pdf));
     assert!(w.iter().any(|w| w.contains("for a 50 × 50 pt page, the file's 100 × 100 pt page at 1:2")), "{w:?}");
+    // A page cropped to the art's box, not the artboard, is that art's page: the layers are read,
+    // with their artboard.
+    let cropped = native(&layer("Jig", true, &square(10, 10))).replace("%AI3_Cropmarks: 0 0 100 100", "%AI3_Cropmarks: 0 0 50 50");
+    let (d, w, read) = crate::layered_ai(&private(&cropped), pdf.clone(), vec![], false, false, 1.0);
+    assert!(read, "{w:?}");
+    assert_eq!(names(&d), ["Jig"]);
+    assert_eq!(d.artboards[0].rect.size(), vectorcraft_geom::Size::new(50.0, 50.0));
 }
 
 /// #2034: editing data for a page of another size gives both sizes in points, in an EPS's note and
diff --git a/crates/testkit/src/ai.rs b/crates/testkit/src/ai.rs
index b7af4b5a5..c7bbcbff1 100644
--- a/crates/testkit/src/ai.rs
+++ b/crates/testkit/src/ai.rs
@@ -5,9 +5,17 @@
 /// Editing data of a `width` × `height` page: header comments (the art's bounding box and the crop
 /// marks are the page), `body` (setup and layers), the trailer.
 pub fn editing_data(width: f64, height: f64, body: &str) -> String {
+    editing_data_in([0.0, 0.0, width, height], [0.0, 0.0, width, height], body)
+}
+
+/// [`editing_data`] whose art's bounding box (`x0 y0 x1 y1`, art space) is `art` and whose
+/// artboard (the crop marks) is `board`.
+pub fn editing_data_in(art: [f64; 4], board: [f64; 4], body: &str) -> String {
+    let [a0, a1, a2, a3] = art;
+    let [b0, b1, b2, b3] = board;
     format!(
-        "{PS_HEADER}\n%%Creator: VectorCraft tests\n%%BoundingBox: 0 0 {width} {height}\n%%HiResBoundingBox: 0 0 {width} {height}\n\
-         %AI5_FileFormat 14.0\n%AI3_Cropmarks: 0 0 {width} {height}\n%%EndComments\n\
+        "{PS_HEADER}\n%%Creator: VectorCraft tests\n%%BoundingBox: {a0} {a1} {a2} {a3}\n%%HiResBoundingBox: {a0} {a1} {a2} {a3}\n\
+         %AI5_FileFormat 14.0\n%AI3_Cropmarks: {b0} {b1} {b2} {b3}\n%%EndComments\n\
          %%BeginProlog\n%%EndProlog\n%%BeginSetup\n%%EndSetup\n{body}\n%%PageTrailer\ngsave annotatepage grestore showpage\n%%Trailer\n%%EOF\n"
     )
 }
diff --git a/docs/file-format-parity.md b/docs/file-format-parity.md
index 945b64db4..6f1b96872 100644
--- a/docs/file-format-parity.md
+++ b/docs/file-format-parity.md
@@ -1,6 +1,6 @@
 # File format parity
 
-> **Last reviewed:** 2026-10-10 · **Last updated:** 2026-10-29 · **Change:** minor (`.ai` and Illustrator EPS import: the editing data is read up to an eighth of installed memory, 512 MiB to 2 GiB, #3510, #4628) · **Target:** Adobe Illustrator 2026 (30.x)
+> **Last reviewed:** 2026-10-10 · **Last updated:** 2026-10-30 · **Change:** minor (`.ai` import: a PDF part cropped to the art no longer hides the layers) · **Target:** Adobe Illustrator 2026 (30.x)
 
 Every format Illustrator opens, places, saves or exports, against VectorCraft's format table (`FORMATS`,
 `OPEN_EXTS`, `PLACE_EXTS` in `crates/engine/src/cmd/fileio/formats.rs`). Illustrator's list comes from its public
@@ -84,7 +84,7 @@ private editing data. VectorCraft:
 
 ### Illustrator's private editing data
 
-the app's own editing copy of a file (an EPS after its `%%EOF`, a `.ai`'s `AIPrivateData` streams, a PDF's `AIPDFPrivateData` streams when it was saved with the editing data (#868), Zstandard or zlib, the EPS's in ASCII85) is not documented as a whole. VectorCraft reads its structure: layers and sublayers (names, order, visibility, lock, printing, preview, dimming, colour), groups as they were nested, compound paths (painted by their last path, or the last painted one, with the paths of groups inside them kept as their paths: #1410), clipping groups and layer clipping, object names, hidden and locked objects, fills and strokes (grey, CMYK, RGB, named and spot colours, spot colours defined in Lab with their Lab values (#1032), global process colours as global swatches at their tint), linear and radial gradients, pattern fills with their patterns as pattern swatches (#1025), opacity, blend modes, isolation and knockout, embedded images where their matrix puts them, rotated ones too (#2269), with their alpha channel (CMYK images keep their ink amounts, as they do when a `.ai` opens from its PDF part, and an alpha channel that hides part of a CMYK image becomes a grayscale opacity mask), placed (linked) files as linked images in their place, each showing the low-resolution preview the file keeps of it until its file is found or relinked in the Links panel (#1264), guides, and every artboard where it is. A gradient with stops in the middle runs its whole ramp across its shape, and a gradient object does not pass its fill to the next path (#1390). An object whose art style is one fill and one stroke at most (solid colours, its own gradient, a variable-width profile) with an Outer Glow or Drop Shadow comes in as itself with them; one with more, other effects, a brush or dashes aligned to corners comes in as its drawn look (a group named after it). Type keeps its art style as a live appearance whatever its stack, so it can be retyped: each fill and stroke of the style's stack (solid colours, gradients and pattern fills, placed as the drawn look places them) with its own opacity, blend mode and effects, the effects on the whole object, and where the characters paint among them; the effects read are Warp, Offset Path, Transform (with copies), Drop Shadow, Outer Glow, Feather, Gaussian Blur and Bas Relief (its settings at their defaults), and a style with others keeps the type's drawn look. The page's drawing of such a style isn't held to look like VectorCraft's where its drawn look lay. Gradients take their matrix (`Xm`, radials and skewed ones too), focal point and the order of stops sharing a ramp point; gradient meshes, a clipping path's own fill and the artboards' names are read too. `editingData: "always"` (`convert --editing-data`) opens a `.ai` or an Illustrator EPS from its editing data even where it draws differently from its page, which can be older, with a note. Type is made from the file's text document, whole and editable, shown or not: point type (its frame's matrix, which newer files leave out when it is neither turned nor scaled, read as none), area type (a story in several frames as threaded type) and type on a path, with its characters, fonts, size, leading, tracking, scaling, baseline shift, fill and stroke (grey, RGB or CMYK), alignment, indents and paragraph spacing; the page's drawing of it (often in pieces) is left out. Where the text document can't say (a frame of another kind, or a document that doesn't agree with the page), type that shows comes from the page into its text object's place (no group around it), and hidden type is left out with a note. Kerning, underline and the other character and paragraph options aren't read yet. A `.ai` saved without PDF compatibility (a placeholder page) opens from its editing data alone, its point type, area type and type on a path made from the text document. A `.ai` whose PDF part shows only text isn't taken for that placeholder when its page states an art, trim or bleed box: it opens from its editing data alone where that reads, else from its PDF part (#1387). Text objects with an appearance keep their own text and frame, type of hidden layers stays hidden, and type on later artboards lands on its text objects (#1391); area type whose frame is too short for a line of the fallback font shows its first line (the frame is made taller, with a warning in the open result) (#1392). Symbols, pattern fills whose pattern it can't read, placed files the data keeps no preview of and anything else it doesn't read make the file come in as its page (on a layer that doesn't show, they are left out of that layer), with a warning saying what. The layers' art is also compared with the page (by lightness, so colours written as their CMYK equivalents compare alike, and by whether a pixel is inked at all); the page is used when they differ by more than 5%, where one draws an object the other doesn't (away from type, whose lines may be laid out afresh), and when an EPS whose page is its art's box has art printed outside it (#735). An EPS whose page is its art's box keeps that page when it comes in through its layers, not the editing data's artboard, and a text object counts as drawn by the page by its colours, not only its lightness, so a red stroke stacked over a dark red one stays (#1102). `textAs: "outlines"` opens a `.ai` whose type shows from its PDF part, which has the outlines; `editingData: false` opens any of them as its page or PDF part only, for print pipelines (#735). A file whose import left something out (hidden text, art or layers) is never written over by an export or Save As to its own path unless `acknowledgeLoss: true` (`convert --replace-lossy`), so the original keeps what the document lacks. Where a `.ai` can't be read that way (or its layers are turned off in the import options), we read its PDF-compatible part, so Illustrator-only live objects arrive as appearance, and plain groups arrive ungrouped: that part draws each object on its own and marks only layers, clipping groups and groups with opacity, blending or a mask (those come back). Art outside the artboards isn't in that part at all (Illustrator's PDF part leaves it out), so it doesn't open; the import says so in a note (#472). Its pages still go where the editing data puts the artboards (a grid stays a grid), with the art on each, when it has as many artboards as pages, of the same sizes (#1068). Its pages are laid out so that art one of them draws past its artboard (notes beside a design, which the part doesn't clip) doesn't lie over the page before it in the row, where it was taken for that page's art: the check against the editing data then found art its layers don't have there and opened the file as its PDF part, without the art outside the artboards. That part also has no paragraphs: area type is written line by line, and comes back as area type only where its font is installed and lays the lines out with the same breaks, else as a point type object per line (#508). Art off the page that a PDF does draw opens on the pasteboard.
+the app's own editing copy of a file (an EPS after its `%%EOF`, a `.ai`'s `AIPrivateData` streams, a PDF's `AIPDFPrivateData` streams when it was saved with the editing data (#868), Zstandard or zlib, the EPS's in ASCII85) is not documented as a whole. VectorCraft reads its structure: layers and sublayers (names, order, visibility, lock, printing, preview, dimming, colour), groups as they were nested, compound paths (painted by their last path, or the last painted one, with the paths of groups inside them kept as their paths: #1410), clipping groups and layer clipping, object names, hidden and locked objects, fills and strokes (grey, CMYK, RGB, named and spot colours, spot colours defined in Lab with their Lab values (#1032), global process colours as global swatches at their tint), linear and radial gradients, pattern fills with their patterns as pattern swatches (#1025), opacity, blend modes, isolation and knockout, embedded images where their matrix puts them, rotated ones too (#2269), with their alpha channel (CMYK images keep their ink amounts, as they do when a `.ai` opens from its PDF part, and an alpha channel that hides part of a CMYK image becomes a grayscale opacity mask), placed (linked) files as linked images in their place, each showing the low-resolution preview the file keeps of it until its file is found or relinked in the Links panel (#1264), guides, and every artboard where it is. A gradient with stops in the middle runs its whole ramp across its shape, and a gradient object does not pass its fill to the next path (#1390). An object whose art style is one fill and one stroke at most (solid colours, its own gradient, a variable-width profile) with an Outer Glow or Drop Shadow comes in as itself with them; one with more, other effects, a brush or dashes aligned to corners comes in as its drawn look (a group named after it). Type keeps its art style as a live appearance whatever its stack, so it can be retyped: each fill and stroke of the style's stack (solid colours, gradients and pattern fills, placed as the drawn look places them) with its own opacity, blend mode and effects, the effects on the whole object, and where the characters paint among them; the effects read are Warp, Offset Path, Transform (with copies), Drop Shadow, Outer Glow, Feather, Gaussian Blur and Bas Relief (its settings at their defaults), and a style with others keeps the type's drawn look. The page's drawing of such a style isn't held to look like VectorCraft's where its drawn look lay. Gradients take their matrix (`Xm`, radials and skewed ones too), focal point and the order of stops sharing a ramp point; gradient meshes, a clipping path's own fill and the artboards' names are read too. `editingData: "always"` (`convert --editing-data`) opens a `.ai` or an Illustrator EPS from its editing data even where it draws differently from its page, which can be older, with a note. Type is made from the file's text document, whole and editable, shown or not: point type (its frame's matrix, which newer files leave out when it is neither turned nor scaled, read as none), area type (a story in several frames as threaded type) and type on a path, with its characters, fonts, size, leading, tracking, scaling, baseline shift, fill and stroke (grey, RGB or CMYK), alignment, indents and paragraph spacing; the page's drawing of it (often in pieces) is left out. Where the text document can't say (a frame of another kind, or a document that doesn't agree with the page), type that shows comes from the page into its text object's place (no group around it), and hidden type is left out with a note. Kerning, underline and the other character and paragraph options aren't read yet. A `.ai` saved without PDF compatibility (a placeholder page) opens from its editing data alone, its point type, area type and type on a path made from the text document. A `.ai` whose PDF part shows only text isn't taken for that placeholder when its page states an art, trim or bleed box: it opens from its editing data alone where that reads, else from its PDF part (#1387). Text objects with an appearance keep their own text and frame, type of hidden layers stays hidden, and type on later artboards lands on its text objects (#1391); area type whose frame is too short for a line of the fallback font shows its first line (the frame is made taller, with a warning in the open result) (#1392). Symbols, pattern fills whose pattern it can't read, placed files the data keeps no preview of and anything else it doesn't read make the file come in as its page (on a layer that doesn't show, they are left out of that layer), with a warning saying what. The layers' art is also compared with the page (by lightness, so colours written as their CMYK equivalents compare alike, and by whether a pixel is inked at all); the page is used when they differ by more than 5%, where one draws an object the other doesn't (away from type, whose lines may be laid out afresh), and when an EPS whose page is its art's box has art printed outside it (#735). An EPS whose page is its art's box keeps that page when it comes in through its layers, not the editing data's artboard, and a text object counts as drawn by the page by its colours, not only its lightness, so a red stroke stacked over a dark red one stays (#1102). `textAs: "outlines"` opens a `.ai` whose type shows from its PDF part, which has the outlines; `editingData: false` opens any of them as its page or PDF part only, for print pipelines (#735). A file whose import left something out (hidden text, art or layers) is never written over by an export or Save As to its own path unless `acknowledgeLoss: true` (`convert --replace-lossy`), so the original keeps what the document lacks. Where a `.ai` can't be read that way (or its layers are turned off in the import options), we read its PDF-compatible part, so Illustrator-only live objects arrive as appearance, and plain groups arrive ungrouped: that part draws each object on its own and marks only layers, clipping groups and groups with opacity, blending or a mask (those come back). Art outside the artboards isn't in that part at all (Illustrator's PDF part leaves it out), so it doesn't open; the import says so in a note (#472). Its pages still go where the editing data puts the artboards (a grid stays a grid), with the art on each, when it has as many artboards as pages, of the same sizes (#1068). A `.ai` whose PDF part's pages are cropped to its art (each page the box of all the art, which is the editing data's bounding box) instead of its artboards opens from its layers, with its own artboards, where they draw as that page does. Its pages are laid out so that art one of them draws past its artboard (notes beside a design, which the part doesn't clip) doesn't lie over the page before it in the row, where it was taken for that page's art: the check against the editing data then found art its layers don't have there and opened the file as its PDF part, without the art outside the artboards. That part also has no paragraphs: area type is written line by line, and comes back as area type only where its font is installed and lays the lines out with the same breaks, else as a point type object per line (#508). Art off the page that a PDF does draw opens on the pasteboard.
 
 ## Format by format
 
@@ -131,6 +131,7 @@ the app's own editing copy of a file (an EPS after its `%%EOF`, a `.ai`'s `AIPri
 
 | Date | Change | Summary |
 |---|---|---|
+| 2026-10-30 | minor | `.ai` import: a file whose PDF part's pages are cropped to its art rather than its artboards no longer comes in as that PDF part (each page a copy of all the art) but from its layers, with its artboards, where they draw as the page does; own-fixture test in `tests_aiimport`; scores unchanged |
 | 2026-10-29 | minor | `.ai` and Illustrator EPS import: the editing data is read up to an eighth of installed memory, from 512 MiB to 2 GiB on 64-bit builds (it was 512 MiB on every machine), so a layered file whose embedded rasters unpack past 512 MiB keeps its layers, groups and text on a machine with 8 GiB or more instead of opening as its page or PDF part (#3510, #4628). Measured: 1.1 GB of editing data peaked 1.1 GB above the start; data past the limit is still refused (a 2 MB file unpacking to 2.2 GB stops in under a second). 32-bit and web builds keep 256 MiB. Own-fixture tests `import::ai::tests::editing_data_past_512_mib_is_read_where_memory_allows`, `the_editing_data_limit_follows_installed_memory`; scores and hours unchanged |
 | 2026-10-29 | minor | Affinity import: attached Recolor Adjustment (RcRA) layers import as color overlays with HSL lightness (#4420); scores unchanged. |
 | 2026-10-30 | minor | EMF export: a pattern fill whose tiles stay apart (grid and brick tilings, art inside its tile) is written as one shape for each part of the tile, clipped once to what it fills, instead of every tile in a clip of its own; the picture is the same. A rotated two-colour stripe fill goes from 473 shapes and 246 clips to 2 shapes and 1 clip (122 KB to 61 KB; unrotated 99 KB to 33 KB). Tiles that overlap, hex tilings and tiles with strokes, images or nested clips are written as before (#3911). Tests in `crates/metafile/src/tests.rs` and `scene::merge_tests`; scores and hours unchanged |