Skip to content

PDF Tools v0.9.5

Choose a tag to compare

View all tags
@silverstein silverstein released this 05 Aug 13:52
v0.9.5
fcd481a
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

PDF Tools v0.9.5: Research-Paper Headings Across Layouts

PDF Tools v0.9.5 makes deterministic Markdown substantially better for
born-digital research papers with numbered sections, small-caps headings,
two-column layouts, wrapped titles, and lettered appendices.

What improved

The Markdown renderer now recognizes research-paper headings only when the PDF
supplies conservative structural evidence: an established body margin, section
spacing, bounded geometry, and matching font, size, or exact small-caps
relationships. A wrapped heading is joined only when its lines share the same
font and height, sit tightly together, and align within a bounded column or
page title region.

The exact public-paper validation set now recovers:

  • 22 of 22 numbered headings in Attention Is All You Need;
  • 17 of 17 numbered headings in Adam: A Method for Stochastic Optimization;
  • 31 of 31 content headings in BERT: Pre-training of Deep Bidirectional
    Transformers for Language Understanding    , including its two-line title,
    centered three-line appendix title, two body columns, and A/B/C appendices.

The same guard removes two sideways arXiv labels, eleven chart-caption or prose
false headings in Adam, and three diagram-label false headings in Attention.
Real paper titles and the exact structural title CONTENTS remain headings.

Validation

  • All three public PDFs are bound to exact SHA-256 identities in live tests.

  • The BERT trial produces zero replacement characters and zero explicit gaps.

  • Source and packaged Markdown runtimes and schemas remain byte-identical.

  • The independently reviewed implementation passed the focused renderer,
    worker, contract, malformed-input, extraction-integrity, and live-paper banks.

  • Three complete 55-page Shannon runs remain deterministic and content-identical
    to v0.9.4. Only the renderer's truthful heading-limitations disclosure
    changes. The v0.9.5 Markdown SHA-256 is
    4cdc3a17728bb07100752ad841745a5ecf49a0aade0cd39af69bca899217e203.

  • Complete repository gate: 2,007 tests passed; 85 were intentionally skipped;
    zero failed.

  • Native macOS safety partition: 62 tests passed; 9 were intentionally skipped.

  • The online release-only golden set passed all four real-document checks.

  • Packed MCPB smoke on macOS ARM64 passed with all 41 tools, all 14 prompts,
    canonical resources, PDF mutation, source-bound comparison, and native
    rendering.

  • Transactional share-package validation passed with 41 tools, 14 prompts, 112
    SBOM components, and native rendering.

  • Two isolated MCPB builds and two ZIP builds were respectively byte-identical.

  • Both archives passed integrity checks and contained no checkout, evidence,
    temporary-directory, or other local machine paths.

Honest limits

This release does not claim universal heading inference. Ambiguous, very short,
unsupported, or image-only heading styles remain plain text. It does not add
OCR or general mathematical reconstruction. Shannon's eleven ambiguous alpha
placements continue to abstain rather than guess.

Artifacts

fa65d2c0dc07811b4d5d71626a64036a4315ef69505f0af70abc1aafdc362568  pdf-toolkit-mcp.mcpb
b0e3e83bd17c54c7d2769533773dea842420210d0c7800265a0c6dad7ac90cc0  pdf-toolkit-mcp.zip
  • MCPB: 73,692,107 bytes, 2,996 files.
  • ZIP: 1,169,176 bytes, 28 files.

Install

Claude Desktop users can download and open pdf-toolkit-mcp.mcpb from the
release assets. Cursor and other local stdio MCP hosts can use
pdf-toolkit-mcp.zip or the checked-out server/index.js.

Assets 4
Loading