Provenance

Source

One source: the published directory, United Nations Trade Data Elements Directory (UNTDED) 2005 — ECE/TRADE/362, published in parallel as ISO 7372:2005. The working copy’s pages 20–132 carry a text layer; the front matter (pp. 9–19) exists only as page images in the original publication and was transcribed from renders of it.

Extraction

• The element table (pp. 28–132) is extracted from the text layer with layout preserved (pdftotext -layout), routed by column bounding boxes into the nine printed columns.
• Wrapped cells are joined by explicit rules (the printed hyphenation of narrow columns, e.g. “Document/m essage”, is reassembled; prose spacing is never glued).
• The result is a YAML single source of truth: 1504 elements, one file per tag-range category, every entry carrying its PDF page.

Verification

Every claim on this site is backed by a check that runs in the dataset’s test suite:

• Regeneration — the YAML must be byte-identical to a fresh extraction from the source PDF; any hand-drift fails the build.
• Character conservation — the exact multiset of non-space characters is conserved from the text layer through routing, joining and normalization, in two independently checked stages.
• UID reconciliation — every printed four-digit tag became exactly one element.
• Independent re-reads — sampled rows are re-read from page images via OCR (tesseract) and compared against the extracted cells.
• Bridge zones — parsed UNLK zones stay within printable ranges with ordered spans.

The review queue

Where the source itself is ambiguous (a wrapped digit, a change tag printed outside the legend), the entry is recorded with its provenance and flagged in a review queue — never silently guessed. Two change tags (cd, cr-class) are printed outside the section 4.1 legend and are marked as such wherever they appear.

Corrections

Suspect a cell? The contributing page describes how to raise a correction request; every correction is re-verified against the source before it lands.