diff --git a/package.json b/package.json
index 6fbd153..4e3826e 100644
--- a/package.json
+++ b/package.json
@@ -21,7 +21,7 @@
     "typecheck:web": "tsc --noEmit -p tsconfig.web.json --composite false",
     "typecheck:test": "tsc -p tsconfig.test.json",
     "typecheck": "npm run typecheck:node && npm run typecheck:web && npm run typecheck:test",
-    "test": "node --import ./test/ts-resolve.mjs --disable-warning=MODULE_TYPELESS_PACKAGE_JSON --test \"test/**/*.test.ts\"",
+    "test": "node --experimental-transform-types --disable-warning=ExperimentalWarning --import ./test/ts-resolve.mjs --disable-warning=MODULE_TYPELESS_PACKAGE_JSON --test \"test/**/*.test.ts\"",
     "start": "electron-vite preview",
     "dev": "electron-vite dev",
     "build": "npm run typecheck && electron-vite build",
@@ -36,6 +36,13 @@
     "db:push": "drizzle-kit push",
     "db:studio": "drizzle-kit studio"
   },
+  "//test": [
+    "`node --test` strips TypeScript types rather than compiling them, and strip-only",
+    "mode rejects the enums the production sources use (src/shared/utils/logger.ts is",
+    "on every loader's import path). `--experimental-transform-types` lets those",
+    "modules load unmodified, so parser tests can import the real loaders instead of",
+    "a test-only copy."
+  ],
   "//dependencies": [
     "ONLY packages that cannot be bundled belong here.",
     "electron-vite defaults build.externalizeDeps to true, so every `dependencies`",
diff --git a/test/fixtures/headings.docx b/test/fixtures/headings.docx
new file mode 100644
index 0000000..0d8bd68
Binary files /dev/null and b/test/fixtures/headings.docx differ
diff --git a/test/fixtures/multipage.pdf b/test/fixtures/multipage.pdf
new file mode 100644
index 0000000..f3d6431
Binary files /dev/null and b/test/fixtures/multipage.pdf differ
diff --git a/test/fixtures/sample.md b/test/fixtures/sample.md
new file mode 100644
index 0000000..d99f4e4
--- /dev/null
+++ b/test/fixtures/sample.md
@@ -0,0 +1,11 @@
+# Sample Markdown
+
+Intro paragraph for the fixture.
+
+## Section One
+
+Body of section one.
+
+## Section Two
+
+Body of section two.
diff --git a/test/parsers.test.ts b/test/parsers.test.ts
new file mode 100644
index 0000000..627c902
--- /dev/null
+++ b/test/parsers.test.ts
@@ -0,0 +1,258 @@
+import { test } from 'node:test'
+import assert from 'node:assert/strict'
+import { readFile } from 'node:fs/promises'
+import { join } from 'node:path'
+import { fileURLToPath } from 'node:url'
+import { FileParserService } from '../src/main/services/FileParserService.ts'
+import { PdfLoader } from '../src/main/services/loaders/PdfLoader.ts'
+import { DocxLoader } from '../src/main/services/loaders/DocxLoader.ts'
+import { MarkdownLoader } from '../src/main/services/loaders/MarkdownLoader.ts'
+import { WebLoader } from '../src/main/services/loaders/WebLoader.ts'
+import type { DocumentLoadResult, SectionInfo } from '../src/main/services/loaders/types.ts'
+
+/**
+ * Golden-file tests for the document loaders.
+ *
+ * `DocumentLoadResult.structure` is about to become load-bearing for provenance
+ * (#66/#67) and nothing under `test/` exercised a loader before this file. These
+ * tests parse the committed fixtures in `test/fixtures/` and assert the *shape*
+ * of the result — page count, page boundaries, heading outlines, offset
+ * arithmetic — rather than the extracted prose. A slightly different extraction
+ * still passes; a loader that stops emitting `structure`, miscounts a page
+ * boundary, or loses a heading level fails.
+ *
+ * Fixtures:
+ *   multipage.pdf  three pages, one text run each (generated for this issue)
+ *   sample.pdf     the existing single-page smoke fixture
+ *   headings.docx  minimal OOXML with Heading1/Heading2 styles (generated)
+ *   sample.docx    the existing fixture, deliberately heading-less
+ *   sample.md      ATX headings: one h1 and two h2
+ *   sample.html    the existing Readability smoke fixture
+ */
+
+const FIXTURES = fileURLToPath(new URL('./fixtures/', import.meta.url))
+const fixture = (name: string): string => join(FIXTURES, name)
+
+/** The committed outline each fixture is expected to produce. */
+const DOCX_OUTLINE = [
+  { level: 1, title: 'Overview' },
+  { level: 2, title: 'Details' },
+  { level: 1, title: 'Summary' }
+]
+const MARKDOWN_OUTLINE = [
+  { level: 1, title: 'Sample Markdown' },
+  { level: 2, title: 'Section One' },
+  { level: 2, title: 'Section Two' }
+]
+
+/** `structure` is optional on the result, so every structural test has to ask for it. */
+function requirePages(result: DocumentLoadResult) {
+  assert.ok(result.structure, 'the loader must populate `structure`')
+  assert.equal(result.structure.type, 'pages')
+  assert.ok(result.structure.pages, 'a `pages` structure must carry `pages`')
+  return result.structure.pages
+}
+
+function requireSections(result: DocumentLoadResult) {
+  assert.ok(result.structure, 'the loader must populate `structure`')
+  assert.equal(result.structure.type, 'sections')
+  assert.ok(result.structure.sections, 'a `sections` structure must carry `sections`')
+  return result.structure.sections
+}
+
+function flattenSections(sections: SectionInfo[]): SectionInfo[] {
+  const flat: SectionInfo[] = []
+  const walk = (nodes: SectionInfo[]): void => {
+    for (const node of nodes) {
+      flat.push(node)
+      if (node.children?.length) walk(node.children)
+    }
+  }
+  walk(sections)
+  return flat
+}
+
+function outline(sections: SectionInfo[]): { level: number; title: string }[] {
+  return flattenSections(sections).map(({ level, title }) => ({ level, title }))
+}
+
+/**
+ * Siblings must not overlap, children must stay inside their parent, and no
+ * range may be inverted. This is the invariant #66/#67 are most likely to break
+ * while they rework how offsets are computed.
+ */
+function assertSectionOffsets(nodes: SectionInfo[], label: string, parent?: SectionInfo): void {
+  nodes.forEach((node, index) => {
+    assert.ok(node.endOffset >= node.startOffset, `${label}: ${node.title} has an inverted range`)
+    if (parent) {
+      assert.ok(
+        node.startOffset >= parent.startOffset && node.endOffset <= parent.endOffset,
+        `${label}: ${node.title} escapes its parent ${parent.title}`
+      )
+    }
+    if (index > 0) {
+      assert.ok(
+        node.startOffset >= nodes[index - 1].endOffset,
+        `${label}: ${node.title} overlaps ${nodes[index - 1].title}`
+      )
+    }
+    if (node.children?.length) assertSectionOffsets(node.children, label, node)
+  })
+}
+
+// --- PDF ---------------------------------------------------------------------
+
+test('PDF: pages are numbered in order and their offsets slice back to the page text', async () => {
+  const result = await new PdfLoader().loadFromPath(fixture('multipage.pdf'))
+  assert.equal(result.mimeType, 'application/pdf')
+  assert.equal(result.metadata?.pageCount, 3)
+
+  const pages = requirePages(result)
+  assert.equal(pages.length, 3)
+  assert.deepEqual(
+    pages.map((page) => page.pageNumber),
+    [1, 2, 3]
+  )
+
+  for (const page of pages) {
+    assert.equal(
+      result.content.slice(page.startOffset, page.endOffset),
+      page.content,
+      `page ${page.pageNumber} offsets do not slice back to its text`
+    )
+  }
+
+  for (let i = 1; i < pages.length; i++) {
+    assert.ok(pages[i].startOffset >= pages[i - 1].endOffset, 'page boundaries overlap')
+    assert.ok(pages[i].startOffset >= pages[i - 1].startOffset, 'page offsets are not monotonic')
+  }
+  assert.ok(
+    pages[pages.length - 1].endOffset <= result.content.length,
+    'the last page boundary runs past the document'
+  )
+})
+
+test('PDF: the existing single-page fixture keeps its text and page box', async () => {
+  const result = await new PdfLoader().loadFromPath(fixture('sample.pdf'))
+  assert.equal(result.metadata?.pageCount, 1)
+
+  const pages = requirePages(result)
+  assert.equal(pages.length, 1)
+  assert.match(pages[0].content, /KnowNote Import Test/)
+  assert.equal(result.content.slice(pages[0].startOffset, pages[0].endOffset), pages[0].content)
+  assert.equal(pages[0].metadata?.width, 612)
+  assert.equal(pages[0].metadata?.height, 792)
+})
+
+// --- DOCX --------------------------------------------------------------------
+
+test('DOCX: heading levels and their nesting survive the round trip', async () => {
+  const result = await new DocxLoader().loadFromPath(fixture('headings.docx'))
+  assert.equal(result.title, 'Overview')
+  assert.equal(
+    result.mimeType,
+    'application/vnd.openxmlformats-officedocument.wordprocessingml.document'
+  )
+
+  const sections = requireSections(result)
+  assert.deepEqual(outline(sections), DOCX_OUTLINE)
+  assert.equal(sections.length, 2, 'Overview and Summary are the roots')
+  assert.equal(sections[0].children?.length, 1, 'Details nests under Overview')
+  assertSectionOffsets(sections, 'DOCX')
+
+  for (const section of flattenSections(sections)) {
+    assert.ok(
+      result.content.slice(section.startOffset).startsWith(section.title),
+      `DOCX: ${section.title} startOffset does not point at its title`
+    )
+  }
+})
+
+test('DOCX: a heading-less document still reports an (empty) section structure', async () => {
+  const result = await new DocxLoader().loadFromPath(fixture('sample.docx'))
+  assert.match(result.content, /KnowNote Import Test/)
+  assert.deepEqual(requireSections(result), [])
+})
+
+// --- Markdown ----------------------------------------------------------------
+
+test('Markdown: ATX headings become a nested outline with consistent offsets', async () => {
+  const result = await new MarkdownLoader().loadFromPath(fixture('sample.md'))
+  assert.equal(result.title, 'Sample Markdown')
+  assert.equal(result.mimeType, 'text/markdown')
+
+  const sections = requireSections(result)
+  assert.deepEqual(outline(sections), MARKDOWN_OUTLINE)
+  assertSectionOffsets(sections, 'Markdown')
+
+  // Heading offsets are positions in the raw fixture text while `content` is
+  // trimmed, so the root section can end a newline past `result.content`; the
+  // assertions stay relative to the headings themselves.
+  for (const section of flattenSections(sections)) {
+    const marker = `${'#'.repeat(section.level)} ${section.title}`
+    assert.ok(
+      result.content.slice(section.startOffset).startsWith(marker),
+      `Markdown: ${section.title} startOffset does not point at its ATX marker`
+    )
+  }
+})
+
+// --- HTML --------------------------------------------------------------------
+
+test('HTML: Readability extracts the title and body and still carries a structure', async () => {
+  const html = await readFile(fixture('sample.html'))
+  const result = await new WebLoader().loadFromBuffer(html)
+  assert.equal(result.title, 'Sample Document')
+  assert.equal(result.mimeType, 'text/html')
+  assert.equal(result.metadata?.excerpt, 'Fixture for the packaged-app smoke test')
+  assert.match(result.content, /KnowNote Import Test/)
+  assert.match(result.content, /Hello World/)
+
+  // Readability hoists the document's lone <h1> into `title` and drops it from
+  // the body, so this fixture has no heading left to become a section. The test
+  // still guards the presence of the structure object itself.
+  assert.deepEqual(requireSections(result), [])
+})
+
+// --- structure opt-out -------------------------------------------------------
+
+test('preserveStructure: false drops the structure without dropping the text', async () => {
+  const buffer = await readFile(fixture('multipage.pdf'))
+  const result = await new PdfLoader().loadFromBuffer(buffer, { preserveStructure: false })
+  assert.equal(result.structure, undefined)
+  assert.match(result.content, /Page One Text/)
+})
+
+// --- dispatch ----------------------------------------------------------------
+
+test('FileParserService routes each extension and MIME type to the matching loader', async () => {
+  const parser = new FileParserService()
+
+  const pdf = await parser.parseFile(fixture('sample.pdf'))
+  assert.equal(pdf.mimeType, 'application/pdf')
+  assert.equal(pdf.structure?.type, 'pages')
+
+  const docx = await parser.parseFile(fixture('sample.docx'))
+  assert.equal(
+    docx.mimeType,
+    'application/vnd.openxmlformats-officedocument.wordprocessingml.document'
+  )
+  assert.equal(docx.structure?.type, 'sections')
+
+  const markdown = await parser.parseBuffer(await readFile(fixture('sample.md')), 'md')
+  assert.equal(markdown.mimeType, 'text/markdown')
+  assert.equal(markdown.structure?.type, 'sections')
+
+  for (const supported of ['pdf', 'docx', 'md', 'html', 'htm', 'pptx']) {
+    assert.ok(parser.isSupported(supported), `${supported} should be routable`)
+  }
+  assert.equal(parser.isSupported('exe'), false)
+
+  assert.equal(parser.getFileTypeFromMime('application/pdf'), 'pdf')
+  assert.equal(parser.getFileTypeFromMime('text/markdown'), 'md')
+  assert.equal(parser.getMimeType('md'), 'text/markdown')
+  assert.equal(
+    parser.getMimeType('docx'),
+    'application/vnd.openxmlformats-officedocument.wordprocessingml.document'
+  )
+})