Trimodal heading sizes make size-band derivation bail, so reconciliation bills get no account structure

[willhea]

What happens

derive_size_bands (src/deltatrack/parsers/pdf_anchors.py) returns None when a
document's sub-body heading sizes form more than one strong cluster. Account
detection is glyph-size based, so returning None means the bill gets no
account-level anchors at all — its deepest breadcrumb is TITLE/SEC.

Since #114 retired the For necessary expenses text-trigger fallback (ADR 0018),
there is no second path behind the size bands. Widening their coverage is now the
only way a bill in this state gets its account structure back.

Why it bails

Reconciliation bills print in three heading sizes rather than the two that
appropriations bills use. The clustering step splits the sub-body heading sizes,
finds a second cluster at >=30% of the dominant one, and bails rather than guess
which cluster is the account band:

dominant = max(clusters, key=len)
# More than one *strong* cluster (>=30% of the dominant) => trimodal => bail.
if any(c is not dominant and len(c) >= 0.3 * len(dominant) for c in clusters):
    return None

Bailing was the right call when it was written — a wrong band assignment mislabels
every heading in the document, and there was a fallback to land on. The fallback is
now gone, so the cost of bailing has gone up.

Affected bills

On the current 68-PDF corpus this is 119-hr-1 (engrossed-in-house,
placed-on-calendar-senate, engrossed-amendment-senate — ~1,000 pages each). The
other bills that reach the degraded path are unaffected by this issue: 11
enrolled/public-law files are declined outright as unnumbered layouts (#141, and
#261 tracks a pass for them), and 19 introduced/engrossed shells of 3-14 pages
contain no accounts to find.

Measured 2026-08-04 against develop. Note the PDF corpus is smaller than the XML
corpus, so treat the bill list as a floor rather than a census.

What "done" looks like

derive_size_bands returns usable bands for a trimodal document — picking the
account band on evidence rather than on cluster size alone — and 119-hr-1 gains
account anchors, without changing the anchors any currently-working bill produces.
The anchor goldens (test_pdf_anchor_golden.py) are the regression check for that
second half: they must not move.

Worth confirming before building: that reconciliation bills have an account level
worth recovering. 119-hr-1 is not an appropriations bill, and the one anchor the
retired trigger produced on it was a wrapped section-heading fragment, not an
account. It may be that the right band assignment yields sections, not accounts —
which would still be a coverage improvement, just a different one than the title
suggests.

Context

• Lever 1 of Keep appropriations-specific text triggers out of account-name & structure detection (reserve them for financial semantics) #114, split out so retiring the text trigger did not have to wait on it
• ADR 0018 (docs/decisions/0018-text-triggers-are-financial-only.md) — why there is
  no fallback behind the size path
• PDF diff: replace appropriations text anchors with glyph-size heading segmentation #89 introduced glyph-size detection; PDF + XML leveled heading tree (epic): agency level, section nesting, money rollup #54 is the leveled-heading-tree epic

[willhea]

Reconciliation bills have no account level to recover, so giving up here is the right result. Checked 2026-10-05 on develop (48082b2) with tests/corpus/119-hr-1, the only reconciliation bill committed:

• The XML has no appropriations headings. Both XML versions have 0 appropriations-major, -intermediate or -small elements. Their structure is title, subtitle, part, subpart and section.
• The three heading sizes aren't accounts. In the reported-in-House PDF, the sizes below body text are 10.0 pt (270 lines: section titles such as WORK REQUIREMENTS.), 11.2 pt (602: quoted-law catchlines such as ‘‘(A) IN GENERAL.—) and 12.0–12.6 pt (44: PART 1—… headings). The 10.0 pt group is 45% the size of the largest, above the 30% cutoff in derive_size_bands.
• Every band choice would invent accounts. Forcing the 11.2 pt band emits 308 false account and 45 false agency headings from quoted law. Forcing the 10.0 pt band emits 31 accounts; every one sampled is a quoted SEC. line. Today the PDF gets 11 titles, 355 sections and 936 subsections, and no account or agency headings.

Across the 39 corpus PDFs that do get size bands, no second size cluster reaches 3% of the main one.

Two related needs are tracked elsewhere:

• The 30% cutoff is fragile on bills that do have accounts. Per PDF report shows missing or wrong account headings on a supplemental bill and an authorizing law #738, 117-s-2938 (an authorizing law with an appropriations division, not yet in the corpus) sits at 29.4%. This is in PDF report shows missing or wrong account headings on a supplemental bill and an authorizing law #738 (a supplemental and an authorizing law lose account headings), under its second Done-when item.
• Subtitle and part headings. Show every heading the XML tags in its breadcrumb, and measure the ledger on each kind of bill that appropriates #739 (draft PR) records that reconciliation amounts are filed "true but shallower" because the PDF has no subtitle or part level. If that's wanted, file it as a new feature, not a fix to the size bands.

#551's sub-issue table, its "How it shows up" bullet on 119-hr-1 and its "Where to start" should drop #508.

---

Generated by Claude Code