Remove the sample documents, PDF and licence copies from tests/fixtures/. tests/fixtures/manifest.json lists each file's download URL (pinned to a commit for GitHub sources), size and SHA-256. tools/fetch_fixtures.py downloads them into the git-ignored tests/fixtures/files/ and verifies them. The fixture tests read from there and skip when it is absent.
5.3 KiB
How this converter is developed
The INDD format is undocumented. This project learns it only by examining INDD files and comparing them with IDML files exported from the same documents.
Rules
- No Adobe software. Contributors who analyse the format have not installed InDesign or accepted Adobe's terms of use.
- Only files made by others, for other purposes. Sample documents come from published templates, test suites and archives. We do not ask anyone to create documents for this project.
- No Adobe binaries, and no code under incompatible licences. We do not read, decompile or use InDesign binaries, the InDesign SDK, or code from other INDD readers whose licence is incompatible with this project's.
- Published source under compatible licences may be used, with attribution. The list is in "Permitted references" below.
- The IDML side follows Adobe's published IDML specification.
- No sample files are committed. The open-licensed test fixtures
are downloaded from their sources (
tests/fixtures/README.md). Other samples stay in the git-ignoredcorpus/directory.
Method
The format was learned by black-box observation of files. Nothing else was examined.
What was done
- Collected third-party samples. 357 INDD files made by other people
for other purposes: published templates, test suites and archives.
240 of them come with an IDML file exported from the same document.
No INDD file was made for this project, and we cannot make one: we
have no InDesign. The samples stay in the local, git-ignored
corpus/directory and are not redistributed. - Paired INDD objects with IDML elements. An IDML element's
Selfidentifier is the INDD object's UID in hexadecimal (Self="u1a2"is UID 0x1A2). This lets us put an INDD object's bytes next to the attributes InDesign exported for the same object. - Correlated fields across many files. For each candidate field we counted how often the INDD value equals the IDML value, over every pair in the corpus. A fact is accepted only when it holds across the samples. Values that are the same in every sample are not mapped, because they cannot be told apart.
- Recorded each fact with its evidence.
docs/format/states each fact, the files it was checked on, and the match count (for example "171 of 171"). The code relies only on facts recorded there. - Checked the converter against the reference IDMLs.
tools/compare.pyconverts every corpus pair and compares the result with the IDML that InDesign exported. Output is also validated against the IDML RelaxNG schemas.
What was not done
- InDesign was never installed and Adobe's terms of use were never accepted.
- No Adobe binary was disassembled, decompiled, traced or opened, including the InDesign application bundle and its plug-ins.
- The InDesign SDK and its documentation were not used.
- No code from other INDD readers was used.
- No one was asked to create INDD files for this project.
Points a reviewer should know
-
A few facts rest on a sample that is not public. Where no public sample shows a value, the owner allowed the use of an INDD file made by a third party for another purpose, before this project, together with its print PDF. It has no IDML and is not redistributable, so those facts cannot be re-checked by others. They were found by black-box observation of the PDF (rendered glyphs, text positions, page labels), and
docs/format/marks each one. -
Adobe XMP Toolkit source was read. The XMP Toolkit SDK is published by Adobe under the BSD 3-Clause licence. Its INDD handler (
InDesign_Handler.cpp) was used for the outer container: master pages, contiguous objects and the byte-order flag (docs/format/container.md). Every one of those facts was also checked against the corpus. If a reviewer needs the work to rely on no vendor source at all, the container section can be re-derived from the corpus alone and the reference removed. -
The analysis was done with an AI model. The model may have seen InDesign SDK material in its training data. The rules above forbid using such knowledge, and every recorded fact rests on corpus evidence rather than on recall. Absence of such knowledge cannot be proved, though.
-
There was no two-team separation. A traditional clean room keeps the people who study the format apart from the people who write the code. Here the same process did both. In its place, every fact in
docs/format/can be re-checked by anyone with similar sample files, and the converter is measured against InDesign's own IDML output.
Where format knowledge is recorded
Each fact about the INDD format is written in docs/format/ together with
the observation that supports it, before or alongside the code that relies
on it.
Permitted references
| Reference | Licence | Used for |
|---|---|---|
Adobe XMP Toolkit SDK, InDesign_Handler.cpp |
BSD 3-Clause (third_party/xmp-toolkit-sdk/LICENSE) |
Master pages and contiguous objects (docs/format/container.md) |
| Adobe IDML specification | Published specification | IDML output |
| IDML RelaxNG schemas (generated by InDesign, published in the metanorma/idml repository) | Published with the IDML specification | Validating converter output locally. Not committed. |