PDF report shows a false "Moved" card when only a heading's line wrap changed #648

Re-scoped 2026-10-07 (comment).

• Cause 2 below is fixed by Decide a move's kind in the differ, by one rule on both pipelines #820: no card whose label isn't SEC. n reads "Renumbered" any more.
• Cause 1 remains: 18 wrap fragments, such as NAVY AND MARINE CORPS → AND MARINE CORPS, are now relocated, so they show a false "Moved" card instead.
• Done when, first item, now reads: no moved card at all for a heading whose printed text is the same in both versions.
• Still blocked on Recover PDF heading structure for the bill ledger, measured by agreement with the XML twin #734. The original text follows unchanged.

What's wrong

In a PDF comparison, the report can say a heading was "Renumbered" when the heading is printed identically in both versions and only its line break moved.

Some background, for anyone new to this:

• PDF pipeline. DeltaTrack can compare two printed PDFs of a bill. The PDF reader infers headings and sections from the print.
• Anchors. An anchor is the label the PDF reader records where a section or heading starts, taken from the printed heading text.
• Change cards and moved cards. The report lists each change as a card. When the two versions' text is matched but sits in a different place, the card is "moved". The card then reads either Moved: <old breadcrumb> → <new breadcrumb> or Renumbered: <old label> → <new label>.
• Canonical diff. This is the JSON both pipelines produce (schema/canonical-diff.md). There, move.kind is renumbered when "the section's anchor identifier changed", and relocated otherwise.

Example (checked 2026-10-05 on develop, d3935ee)

118-hr-4366 (FY2024 appropriations), Senate calendar print (v3) → Senate amendment (v4). Both print FAMILY HOUSING OPERATION AND MAINTENANCE, NAVY AND MARINE CORPS, wrapped differently:

• v3, p. 10: FAMILY HOUSING OPERATION AND MAINTENANCE, / NAVY AND MARINE CORPS
• v4, p. 15: FAMILY HOUSING OPERATION AND MAINTENANCE, NAVY / AND MARINE CORPS

The report says Renumbered: NAVY AND MARINE CORPS → AND MARINE CORPS. The same pair also reports Renumbered: ARMY → FAMILY HOUSING OPERATION AND MAINTENANCE, ARMY.

Reproduce (about 5 seconds):

uv run ./diff_pdf.py tests/corpus/118-hr-4366/3_placed-on-calendar-senate.pdf \
  tests/corpus/118-hr-4366/4_engrossed-amendment-senate.pdf --format json -o out.json

Across the 17 PDF pairs the product accepts, develop produces:

• 161 moved cards, 153 of them "Renumbered".
• 130 of those are SEC. n → SEC. m. Not checked one by one; most are likely genuine.
• 23 are heading labels. Most are fragments of a wrapped heading: FACILITIES → STATE EXTENDED CARE FACILITIES, POWER ADMINISTRATION → ADMINISTRATION, DEBT → FEDERAL DEBT. Three were checked against the printed page lines, and all three are wrap-only.
• Of those 23, 3 look like genuine heading edits: ENERGY SECURITY → ENERGY SECRUITY, RELATED AGENCY → RELATED AGENCIES and ADMINISTRATIVE PROVISION → PROVISIONS.

Why it matters

A staffer is told that a provision was renamed or renumbered when it wasn't. That is a false statement in the part of the report meant to say what changed, and it looks the same as a real one. The pairs where it happens are re-typeset amendments, which are exactly the comparisons staff run.

Cause

There are two places:

1. The PDF reader takes a heading from one printed line. A wrapped heading becomes a fragment that depends on where the line broke. This is the same root as In PDFs, a long account name that wraps onto a second line shows up as a fake agency plus a fragment account #524 (a wrapped account heading read as an agency plus an account) and Reading a bill from PDF, headings printed across more than one line get the wrong name or parent #732 (headings spanning more than one printed line are misread). Recover PDF heading structure for the bill ledger, measured by agreement with the XML twin #734 (the open PR that rewrites PDF heading detection) addresses it.
2. _pdf_move labels any text difference "renumbered". It is in src/deltatrack/formatters/canonical.py:262-274, and _move_info_html (lines 521-536) prints it. It compares raw anchor text: any difference becomes "renumbered", including a heading edit, which is not a renumbering. Recover PDF heading structure for the bill ledger, measured by agreement with the XML twin #734 doesn't touch it.

Where #734 leaves this (checked 2026-10-05, #734 head e3d3926)

The same census gives 144 moved cards and 137 "Renumbered": the same 130 SEC. cards and 7 heading labels.

• Gone: every wrap-only fragment above, including NAVY AND MARINE CORPS. The 2026-09-28 comment on this issue said that case was not fixed; it predates Recover PDF heading structure for the bill ledger, measured by agreement with the XML twin #734's last commits.
• Kept, correctly: the 4 genuine renumberings in 115-hr-5895 (SEC. 103 → SEC. 102, 104 → 103, 105 → 104, 107 → 105), and the 3 heading edits above.
• Still doubtful, and not about wrapping:
  • ELECTRICITY DELIVERY → NUCLEAR ENERGY and SEC. 5 → PROHIBITION ON USE OF FUNDS, both also on develop. They look like two different sections paired together.
  • a new crossed pair in 115-hr-5895: DEPARTMENT OF DEFENSE FAMILY HOUSING IMPROVEMENT FUND → MILITARY UNACCOMPANIED HOUSING IMPROVEMENT FUND, and the reverse.

Done when

• On 115-hr-5895 and 118-hr-4366, no "Renumbered" card appears for a heading whose printed text is the same in both versions. (Holds since Decide a move's kind in the differ, by one rule on both pipelines #820. Now: no moved card at all for such a heading.)
• The genuine SEC. renumberings in 115-hr-5895 are still reported.
• The genuine heading edits (SECRUITY, AGENCY → AGENCIES) are still reported as changes.
• A test pins these cases. The suite passes today with the defect present, so a passing suite is not proof.
• The PDF canonical baseline (tests/data/pdf_canonical_baseline.json) is regenerated on purpose, and its diff is reviewed row by row.

Scope

• In scope: false moved cards caused by line wrapping. (The _pdf_move half, whether its raw-text comparison is still needed, was settled by Decide a move's kind in the differ, by one rule on both pipelines #820, which removed it.)
• Out of scope:
  • Assembling wrapped headings itself: Reading a bill from PDF, headings printed across more than one line get the wrong name or parent #732 and In PDFs, a long account name that wraps onto a second line shows up as a fake agency plus a fragment account #524, which Recover PDF heading structure for the bill ledger, measured by agreement with the XML twin #734 would close.
  • What moved, renumbered and relocated should mean, under ADR 0006 (the canonical diff contract).
  • The crossed pairs and other correspondence errors. These may need their own issue.

Where to start

• Blocked on Recover PDF heading structure for the bill ledger, measured by agreement with the XML twin #734. It is open, not a draft, with changes requested on 2026-09-30. The author pushed fixes the same day, and it awaits re-review. It conflicts with develop. Re-scope this issue after it merges.
• Useful now:
  • Reproduce the example above.
  • Read docs/research/pdf-matching-convergence/moved-semantics.md.
  • Run the census on develop and on Recover PDF heading structure for the bill ledger, measured by agreement with the XML twin #734's branch: uv run python docs/research/pdf-matching-convergence/probes/pdf_move_user_facing.py. It takes about 80 s and writes docs/research/pdf-matching-convergence/results/move-user-facing.json.
• Code and tests: src/deltatrack/move_kind.py (since Decide a move's kind in the differ, by one rule on both pipelines #820), and _move_info_html in src/deltatrack/formatters/canonical.py, and tests/test_formatters_adapters_pdf.py.

Unverified

• Whether the 130 SEC. renumberings are all genuine.
• Whether wrapping also affects the table of contents or the Full bill view. Only change cards were measured.
• Bills outside the committed corpus.

History

The first measurement, at d5a1ad0, found 156 of 165 moved cards reading "Renumbered". The 2026-09-28 comment on this issue measured #734 at 156 → 143 against develop at c636448. The 2026-10-05 counts above were taken before #820; the 2026-10-07 comment has the counts after it.