FreePDF 7 support and release policy

Runtime and package scope

Use a maintained, patched .NET 10 or .NET 11 runtime for new applications. The portable packages contain net8.0, net9.0, net10.0 and netstandard2.0 assets; .NET 11 applications use the net10.0 assets. CI executes package consumers on Linux and Windows for .NET 8/9/10/11 and on Windows for Framework 4.6.2/4.7.2 targets, and runs the regression tests of the net10.0 build on the .NET 11 runtime. The .NET 11 consumers install the same candidate packages that are published; the packages are built with the .NET 10 SDK. Framework execution uses the runner's maintained .NET Framework 4.8 runtime; this does not certify historical CLR installations. Prefer Framework 4.8/4.8.1 on an OS supported by Microsoft.

As checked on 2026-11-10, .NET 8 and 9 reached the end of Microsoft's support on 2026-11-10; .NET 10 (LTS) and .NET 11 (STS, released 2026-11-10) are supported through 2028-11-14. The compatibility assets can remain available after runtime support ends, but FreePDF cannot provide security servicing for Microsoft's runtime. Consult the current policy before deployment: https://dotnet.microsoft.com/en-us/platform/support/policy/dotnet-core

Tested OS profiles are Ubuntu 24.04 and Windows Server 2022 in CI, plus a dedicated Windows lab host. macOS, browser/WASM, mobile, Native AOT and trimming are not validated release profiles.

The GDI and WPF packages (io7.FreePDF.Gdi, io7.FreePDF.Layout.Gdi, io7.FreePDF.Wpf, io7.FreePDF.Layout.Wpf) contain net8.0-windows7.0, net9.0-windows7.0, net10.0-windows7.0 and net462 assets. CI executes their package consumers on Windows for net8.0-windows, net11.0-windows and Framework 4.7.2, including an image drawn from the platform's own bitmap type. Their drawing results follow GDI+ or WPF on the host and are not covered by the cross-platform pixel comparisons.

Supported feature claims

The release gates cover drawing and text with explicit font resolution and with installed TrueType fonts found without a font resolver (Core build), paragraph/table layout and PDF/RTF generation, page import/merge/stamping, attachments, AES-128/256 password encryption, and new detached CMS signatures using SHA-256/384/512. Original PDF/A-1b, 2b and 3b fixtures are independently validated. PDF/A conformance is a property of each output document; validate your own output, fonts and images before archival use. Since 7.8.0, PDF/A-1 CID fonts carry a CIDSet, and saving an opened file whose XMP claims PDF/A keeps its XMP in line with the document information and adds no page transparency group; veraPDF validated such files in the 7.8.0 release checks.

Merging and splitting with PdfDocument.AppendPages and ExtractPages carry form fields, bookmarks, links and GoTo actions; regression tests cover them, and a corpus of 2,943 veraPDF and producer files was appended and split without errors before the 7.3.0 release. Tags (the structure tree), page labels, optional content settings and XFA are not carried, and copied signatures cannot be valid in the new file.

Saving a document opened with PdfDocumentOpenMode.Append writes an incremental update and keeps existing signatures valid. Regression tests check that the original bytes stay unchanged, that only new and changed objects are written, and that a signature made before filling in a field or adding a second signature still verifies. The release checks also validated signed, filled and countersigned files with poppler's pdfsig and pyHanko (whose difference analysis rated the changes as signing and form filling), countersigned a file that pyHanko had signed, and saved each of 2,930 veraPDF and producer files unchanged (same bytes) and with one change (only the changed objects written) before the 7.4.0 release. Since 7.5.0, encrypted documents keep their encryption: regression tests cover RC4 and AES-128/256 files made by FreePDF, a form that pyHanko encrypted with AES-256 and signed was filled in and countersigned by FreePDF with both signatures valid in pyHanko and pdfsig, and the 13 encrypted corpus files were saved unchanged (same bytes) and with one encrypted change. Since 7.5.1, empty streams and strings in AES-encrypted documents are encrypted too (32 bytes), so pyHanko can read them. Since 7.5.2, documents with EncryptMetadata false whose metadata streams are encrypted (as pyHanko writes them) open and read correctly: regression tests cover RC4 and AES, and 400 pyHanko files were read. Since 7.6.1, FreePDF writes only the document's own metadata unencrypted when EncryptMetadata is false, as ISO 32000-1 specifies, and pyHanko reads the page metadata of such files. Since 7.7.0, pages can be added, removed and moved in Append mode also when the page tree is nested: regression tests check the objects each update writes and the tree's counts and parents, and nested files that pyHanko signed were edited with page sizes and rotation intact in poppler and MuPDF and every signature valid in pdfsig.

Since 7.5.0, an Append-mode save refuses changes to signed signature fields and to fields a signature locks (FieldMDP, /Lock), and changes a certification signature (DocMDP) does not permit. Regression tests cover each permission level against filling in fields, signing, the document information, comments, page content, pages and the catalog; changing, removing and moving signed and locked fields and the objects they refer to; and saving to a path that cannot be written. pyHanko's difference analysis was run on forms that pyHanko certified (P=1, 2, 3) or signed with and without a field lock, also with nested fields, and that FreePDF then changed: every change FreePDF saved for which pyHanko has a rule was accepted, and every change FreePDF refused was rejected by pyHanko too. pyHanko's default policy is stricter than the permissions in six places, which FreePDF follows the permissions in: added comments, new visible signature fields after a certification, any change to pages after an ordinary signature, new signature fields after a signature that locks all fields, new fields other than signature fields after an ordinary signature (checked with 7.9.0's PdfForm.Add… methods), and selecting a radio button after a certification that permits form filling (the group's value and the button's appearance state change; pyHanko rejects pypdf's update with the same two objects too).

Since 7.6.0, DigitalSignatureOptions.FieldName signs an existing, empty signature field. Regression tests cover a field that is its own widget, one with a separate widget and an invisible one; a field lock recorded as FieldMDP and then enforced; certification levels; fields locked by an earlier signature; and seed values. Forms that pyHanko made with an empty field (plain, with a field lock, with a required reason, certified with P=2) and FreePDF signed are VALID in pyHanko, the certification also in poppler's pdfsig; after pyHanko changed the field that the lock names, pyHanko judged FreePDF's signature INVALID, as it must. pyHanko's default policy reports earlier signatures as changed when the signed field's widget is a separate object, or the field has several widgets or shares an appearance; fields that are their own widget pass.

Since 7.9.0, PdfForm creates text, check box, radio button, combo box, list box and signature fields with appearances. Regression tests read created fields back with their types, widgets, rectangles and appearances, fill them after reopening, flatten, merge and sign them. Before the release, MuPDF and pypdf read the type, value and on states of every created field in new, filled and reopened-then-filled files, and MuPDF and poppler rendered them. Text and choice fields use Helvetica, which is not embedded, so PDF/A documents (new, or opened with an XMP PDF/A claim) refuse them. Fields are not tagged yet, so tagged and PDF/UA documents refuse all of them. DigitalSignatureOptions.Certification writes a certification signature (DocMDP), only as the document's first signature, and honours a signature field's seed value for the signature type (/MDP); regression tests check the signature reference, the catalog's /Perms and FreePDF's own enforcement, and pyHanko read levels 1 to 3 from FreePDF's files and judged filling text fields, check boxes and combo boxes after certification at levels 2 and 3 as permitted (VALID); radio buttons are one of the places listed above.

Factur-X / ZUGFeRD output (FacturXManager) is checked for the PDF side: a PDF/A-3b file validated by veraPDF, with the invoice XML embedded, listed in /AF and declared in the XMP extension schema. FreePDF only checks that the XML is a well-formed Cross Industry Invoice; it does not validate invoice content against EN 16931, a profile or a national CIUS such as XRechnung.

Since 7.9.0, QR codes (CodeQR), Data Matrix ECC 200 (CodeDataMatrix), Code 128 (Code128, also GS1-128), EAN-13, EAN-8 and UPC-A are encoded from their ISO/IEC standards and drawn as vector shapes, also by MigraDoc's Barcode element. Regression tests compare QR codewords with a published example, check the Reed-Solomon codewords of interleaved blocks and the format and version information, and compare Code 128 symbol values and EAN/UPC check digits with worked examples. Before the release, every QR version 1 to 40 at every error correction level and in each mode, at full capacity, and text with UTF-8, control characters and payment-code content, was decoded by an independent reader (zxing-cpp), also with one copy of the format or version information destroyed; drawn codes of each type were decoded from pages rendered by MuPDF and poppler. Data Matrix regression tests compare the codewords of the standard's encodation example and of ASCII, Upper Shift, ECI, FNC1 and Base256 encodation, and check finders, clocks and the Reed-Solomon blocks of 52 x 52 and 144 x 144; before the release, symbols of all 30 sizes at full capacity and with Latin-1, UTF-8, GS1 and Base256 content were decoded by zxing-cpp, and drawn symbols from MuPDF and poppler renders. Whether a printed code scans also depends on its size, contrast, quiet zone and print quality.

Tagged PDF/UA-1 output from the layout engine (PdfDocumentRenderer.TaggedPdf) covers headings (outline levels 1 to 6), paragraphs, lists, tables with heading rows, images and charts with alternative text, hyperlinks and text frames; headers, footers, borders and shading are artifacts. A multi-page fixture with all of these is validated by veraPDF's PDF/UA-1 profile in CI. veraPDF checks what a machine can check; reading order, alternative text quality and heading structure need human review (the Matterhorn Protocol lists these checks). A table split across pages is one Table element, with heading rows repeated on later pages as artifacts. PDF/UA-2 is not covered.

The signing checks verify cryptographic integrity and tamper detection using synthetic self-signed certificates and, since 7.10.0, a test PKI with an in-process OCSP responder, CRLs and an offline time-stamp authority. They do not cover real certificate authorities' and time-stamp authorities' services, EU trusted lists or a legal signature status. No FIPS certification is claimed. MD5 exists only for legacy PDF-format operations and checksums, never as the recommended signature digest.

Since 7.10.0, PdfSharpDefaultSigner writes PAdES B-B and B-T signatures (ETSI.CAdES.detached) on request. Regression tests check the CMS attributes (content-type, message-digest, signing-certificate-v2 with the certificate's hash, issuer and serial number, no signing-time) and the signature time-stamp against an offline test authority. pyHanko validated FreePDF's B-B and B-T files, SHA-256 and SHA-512, also countersigned in Append mode and countersigned after a certification for form filling (the update that adds the ESIC extension is rated "form filling"), as valid with valid time-stamps; poppler's pdfsig judged the same signatures valid but does not report time-stamps. These checks cover CMS validity and the attributes above, not full ETSI profile conformance: no conformance checker was run. Signature time-stamps need .NET 8 or later. Since 7.10.0, FreePDF also writes validation data into the document security store and adds document time-stamps (B-LT, B-LTA), in Append mode and also under a "no changes" certification. Regression tests check the DSS entries, that the signed bytes stay as they were, the document time-stamp token against the covered bytes, and the permission rules at P=1. pyHanko 0.38.1 rated the DSS and document time-stamp updates as LTA updates (also after a P=1 certification), judged the document time-stamps valid and trusted, and its AdES LTA validation (EN 319 102-1 5.6.3) passed offline with the test certificates as trust anchors. AddValidationDataAsync was tested against a test PKI (root, intermediate, signer, time-stamp authority) whose OCSP responder and CRLs answer in-process: chains, OCSP responses, the CRL fallback, revoked and unreachable cases. With only that PKI's root trusted, revocation data required and online fetching off, pyHanko judged the B-LT file VALID from the DSS alone, and its AdES LTA validation of the B-LTA file passed now and when simulated 30 days later, after the signer's 2-day certificate had expired; the B-T file without a DSS found no chain. Real CAs' responders, delegated OCSP responders and partitioned or delta CRLs were not tested. poppler's pdfsig does not recognise document time-stamps.

Since 7.10.0, PdfSignatureVerifier checks signatures. Regression tests cover valid, untrusted and tampered signatures, signature time-stamps, B-LTA in the test PKI with revocation from the DSS, an expired signer validated at its time-stamp, a revoked signer, missing revocation data, online revocation, countersignatures, and later updates allowed at P=2 and forbidden at P=1. On a corpus of such files pyHanko 0.38.1 gave the same verdicts, with two differences: after a P=1 certification, an update that adds document information is accepted by pyHanko (its policy allows such updates) and is a forbidden change for FreePDF, which follows ISO 32000-2 ("no changes" but validation data and document time-stamps); and a signer revoked before the signing time is "indeterminate" in pyHanko's AdES validation and "Invalid" in FreePDF. Tests also cover attacks found in review: a time-stamp from an untrusted authority (its time is not used, so it cannot rescue an expired signer), a forged delegated OCSP responder (the responder certificate must be signed by the CA, carry id-kp-OCSPSigning and id-pkix-ocsp-nocheck, and be valid when the response was produced), a later trailer that points at another catalog, changed page content or a page-covering comment after an approval signature, a visible "document time-stamp" field after a "no changes" certification, a later signature field without an intact signature, a signature dictionary copied into a later update, a signing-certificate attribute for another certificate, a signature removed from the form, "revoked" in the DSS next to an older "good", a CRL issued after the certificate expired, a decompression bomb in the DSS, a malformed signature dictionary, and a certificate not meant for signatures, a later field that points at a signed dictionary, a time-stamp authority revoked later for compromise, and a "document time-stamp" widget that is NoView but prints. Four review rounds (two by Claude agents, two by Codex) found these. Files signed and countersigned by pyHanko 0.38.1 verify as Valid. The verifier is conservative: SHA-1 and MD5 digests, unverifiable time-stamps, data after the signature in /Contents, delta and partitioned CRLs and unknown critical CRL extensions make a result Indeterminate or are not used, and a PAdES level is reported only for valid signatures. A visible empty signature field added or changed after a signature makes that signature Indeterminate until the field is signed. The appearance streams that filling in a form field writes (which ISO 32000 permits after a signature) are not inspected for what they draw. Since 7.10.1 an object that a page draws, or that decides how every page shows (the catalog's optional content properties and output intents), gets no exemption for later changes, whatever other role it also has, and an exemption for a role (an appearance, an annotation, the document information or metadata, validation data) holds only for an object the document reaches through that role alone. Actions and JavaScript are not judged for what they would show or hide when run. The check of later changes has two known limits. First, it assumes that an ordinary producer made the signed revision. A file that was built before signing so that one object has two roles can pass a later change to that object. Two cases are known. One is an array that is a page's /MediaBox and also its /Annots. The other is a bookmark that a form widget also lists as an appearance. Second, the check does not compare the values that FreePDF sets when it opens a file. These are the /Parent, /Prev and /Last of a bookmark, and the /SigFlags of a form that is an object of its own. Since 7.10.1 a page keeps the /Parent of the file in Append mode, so the check compares it. A viewer may take what a page inherits from the node that /Parent names. In a test, PDFium 159.0.8149.0 did, and poppler 25.03.0 and MuPDF 1.25.1 did not. Where no change after the signature is acceptable, require CoversWholeFile, or compare the revisions yourself. Since 7.10.1 a file with more than MaxRevisions revisions or MaxSignatures signatures (64 each by default) throws instead of being checked, earlier revisions are read one at a time and only where a revision ends, and DSS inflation counts against its budget, which bounds the cost of a hostile file. Outline items more than 200 levels deep are not read, so a deeply nested outline cannot overflow the stack. Since 7.10.1 the reader also refuses arrays and dictionaries nested more than 256 levels and form fields nested more than 100, does not follow a chain of indirect stream lengths, and bounds every walk up through a field's /Parent; a signature check throws when an earlier revision is refused that way. XMP metadata over 16 MiB is not read. Known effects of these limits: revisions are counted by the %%EOF markers in the file, so a file that embeds many other PDFs can exceed MaxRevisions with one revision; field names in results are cut at 1,000 characters; in a signed file with a form field name over 4,096 characters, updates after a signature are not compared, so that signature is not reported Valid, and FreePDF does not save such a file in Append mode; and bookmarks more than 200 levels deep are left out when pages are merged or extracted. Since 7.10.1 the cross-reference and object streams the reader decodes while opening a file decode to at most 64 MiB each, and to at most 256 MiB or 64 times the file's size together, counting every filter stage's output, also output a predictor then drops or that damaged data cuts short; a file beyond that throws PdfReaderException. Other streams (page content, images, fonts) are decoded only when they are used, without such a limit. It is a technical check: EU trusted lists, qualified status, signature policies and ETSI validation reports are not covered, and RSASSA-PSS signatures on certificates, OCSP responses and CRLs are not verified. Since 7.10.1 an OCSP response or CRL that FreePDF cannot read, parse or verify, and that could revoke a certificate, keeps that certificate from being shown as not revoked, so the signature is at best Indeterminate. With PdfRevocationCheck.Online, a current answer fetched from the issuer's responder or CRL can still show a certificate not revoked next to DSS data FreePDF could not read (too large, or in a form it does not decode), but not next to data it read and could not use. Since 7.10.2, for a signature with a time-stamp, a CA certificate's revocation data counts while its validity window covers the validation time and reaches the signing time; the signer's own certificate needs data issued after the time-stamp. So a revocation of a CA certificate that its issuer records after its latest CRL is not seen until the next CRL. With PdfRevocationCheck.Online, the CRL is asked when an OCSP answer does not settle the status, and a time-stamp authority's chain is asked about when the DSS's data for it has run out.

Text extraction (7.15.0) is covered by regression tests on hand-made fonts and CMaps and on FreePDF's own output, and the corpus checks it against the intended text of FreePDF, ReportLab and Chromium files. It reads the font's ToUnicode CMap or its encoding; text extraction lists what it does not read. PdfTextExtractionOptions bounds each page, but there is no time limit: extract text from untrusted PDFs in a separate process with memory and time limits.

Filling or flattening form fields with text the form's fonts cannot encode in WinAnsi, incremental updates of documents encrypted with a public-key or other non-standard security handler, PDF/UA for documents drawn without the layout engine, arbitrary PDF/A conversion, form fields created in PDF/UA documents, text and choice fields created in PDF/A documents, and complex script shaping are not promised in 7.10. Placeholder inherited types are not evidence that those features are complete.

Accessibility checking

7.11.0 adds an offline accessibility preflight in the Core, GDI and WPF library assemblies. The separate io7.FreePDF.Accessibility.Tool requires .NET 10. Candidate-package CI installs and runs the Linux-built tool package that publication promotes on Linux, and a package built from the same revision on Windows. The tool has not been tested on macOS or with .NET 11 roll-forward.

The 7.12.0 CLI can run a separately installed veraPDF validator with its PDF/UA-1 profile. The package checks exercise veraPDF 1.30.3 on Linux and Windows. Use freepdf-a11y --requirements to see which Matterhorn conditions need human review.

In 7.13.0 the native scanner checks tagged content, painting and fonts for visible text inside local Form XObjects, including nested forms. If the scanner cannot resolve a named marked-content property, it reports an incomplete scan.

In 7.14.0 the scanner reads content one operator at a time and has a time limit, MaxDuration (default 2 minutes; freepdf-a11y --timeout). Tests scan nine hostile inputs within a 256 MiB .NET heap. The limits are still not a sandbox: scan untrusted PDFs in a separate process with memory and time limits.

In 7.14.1 MaxDuration and the cancellation token also stop the reader while it opens the PDF. Opening keeps about 500 bytes per object in memory, which no limit bounds.

Scanner coverage is partial PDF/UA-1. A completed scan means its implemented checks finished; it does not establish ADA, WCAG or full PDF/UA compliance. Human review and independent validation remain required. Unsupported content and inspection limits produce an incomplete result. See accessibility-preflight.md for the exact check list and limits, and accessibility-roadmap.md for next work.

Maintenance and security

io7 maintains FreePDF independently. Updates need not match PDFsharp's cadence. Reproduce issues on the latest FreePDF patch and a supported runtime. Routine bug/security fixes preserve the supported public API within 7.x; incompatible changes require a major version, with an explicit exception only when a security fix cannot safely preserve behaviour. Release notes must explain any exception. Maintenance is best effort; this policy does not promise an SLA.

Keep documents and passwords out of public issues. Submit an original minimal reproducer where possible; follow SECURITY.md for sensitive reports. Process untrusted documents with application-level time and memory limits. The bounded mutation corpus is a regression gate, not proof against every malformed PDF. Since 7.8.0, a file whose cross-reference information cannot be read is rebuilt from its objects, as viewers do; a rebuilt file may lack objects its damage hid, and Append mode refuses it.

Promotion to stable

Every candidate requires green regression, API, package-install, advisory, accessibility-corpus, independent-validator and comparison-lab checks. The exact candidate packages, source revision, tool versions, baseline hashes and evidence must be retained. Known visual changes need an explanation and regression coverage; a green pixel match alone is insufficient. Review the migration guide and third-party audit.

The plain stable package rehearsal installs from an isolated local source and verifies library assembly versions and the accessibility tool's behavior. Rehearsal artifacts are unpublished and must not be mistaken for a released package. A maintainer reviews the final candidate and explicitly authorizes public publication; the checks contain no NuGet publishing step. The separate manual Publish verified packages workflow requires successful checks at its exact source revision and only publishes when its explicit publish input is enabled. See releasing.md for setup and verification.