[MaxGhenis]

This is part 1 of milestone D4 of the income-and-resources design (#559: the D4 row of the milestone table). It adds pinned captures and parsers for aggregate tables. Part 2 adds the SSA and BLS sources that refuse plain requests, through Internet Archive captures. Thirteen independent review rounds asked for fixes. 8cb397f5, on 17bdcf00, fixes round 13's one medium finding, a regression against the parent. The round's low findings change no record on a pinned input, so they stand as the limits listed under Open items, and the claims they falsified now say less. The round-14 delta review approved 8cb397f5. Since then we merged master into the branch four times and recounted the tiers three times. Only the merge of 89d064b5 conflicted outside the tier files; it combines D4's birth-evidence exclusions and reachability checks with #573's SSI0 additions.

What it adds

• Captures in data/external/snapshots/income_resources_d4_v1/: 54 pinned sources. Each per-source manifest records time, size, SHA-256 and URL. A folder-level .gitattributes (* -text) keeps the 37 captures with CRLF line endings byte-exact.
• Registries: 874 entries, comprising 861 numeric tables and 13 references:
  • SSA Annual Statistical Supplement Tables 4.B1, 4.B2, 4.B10 and 4.B11;
  • 777 POMS SI 01415 annual state-supplement payment charts;
  • POMS RS 01505 and SL Section 218 coverage references;
  • 18 Federal Reserve Z.1 tables (June 2026 codes);
  • 52 DOL Private Pension Plan Bulletin tables (2023 edition);
  • ten IRS SOI IRA tables (tax year 2023).
• Readers (src/populace_dynamics/income_resources/aggregate_*.py):
  • every public reader checks the capture's size and SHA-256 before reading a cell;
  • they return typed cells, with table, row, column, unit and year.

Invariants

• Hash and size match before a reader reads any cell.
• Missing, suppressed and unavailable cells stay None, never zero.
• Valid separators and footnotes keep magnitude and sign. Malformed numbers refuse without echoing their content.
• HTML spans keep a rectangle, its hierarchy and its footnotes.
• Duplicate cell identities refuse.
• POMS cells keep state, claimant, living-arrangement code and category.
• The POMS reader reads a State's code definitions from every definition table in its section and merges them. A code that two tables define differently refuses. A jointly published code ("C/H", "A or B") names each single code, and a full-width note row defines no code.
• The POMS reader reads every chart's selector (State, claimant type, effective date and definition tables) right, or refuses with TableStructureError, on every form its module docstring says it reads. The docstring's Not modelled, Limits and leave-standing lists name the forms it does not read that the review probes found, among them an abbreviation without its period, text outside any block or in a chart's own caption, a long line of prose, an effective-date sentence of 80 characters or more whose second date takes another form, a claimant named in other words than its three claimant words, a date with a two-digit year in a form no sentence reads ("January 1, 31"), a date split across two blocks, a date split across a <br> in a block of 80 characters or more, a date in a form no list names ("July 1", "7/1", "July of 2030"), and a short line outside a heading that holds only years before the dates in force and no dating word. The tests pin the round-12 review's table rows and witnesses and many of the listed limits; neither the tests nor the lists show that no other form exists.
• Every block that ends a paragraph, apart from a table or rule, names its State, claimant or date from its own text, read whole and line by line. A line ends at a <br> and where a nested block starts or ends, and drops Unicode format (Cf) characters and the text of script and style, which HTML does not render, a footnote marker's text included. The reader reads each block's whole text and each line five ways:
  • joined, as HTML renders inline text;
  • bare, without footnote markers: a sup, or a link to an anchor on the page (href="#..."), whose text is one to three word characters or footnote signs. A longer sup, or a link elsewhere, is text, so no word of prose vanishes;
  • spaced, with a space at every element edge;
  • unmarked, for the date checks only: without any element either marker rule counts, the bare reading's markers and any sup or link whose href holds "fn".
  • unlinked, for the date checks only: the unmarked reading without any link whose text has a marker's size, wherever its href points.
• A State heading and a claimant come from the joined or bare reading, so Utah<sup>1</sup> heads Utah. Every State net reads all three readings, so no inline edge hides a State.
• The reader chooses the date sentence from the joined reading. It refuses when its bare, unmarked or unlinked reading states other dates or a range the joined one does not, so a marker by either rule, or a link of a marker's size whatever its href, inside a date (January 1<sup>2</sup>, 2031, January 1<a href="0501415010#note5">5</a>, 2031) changes no date silently; the page refuses. A line that is a dated sentence only in its bare, unmarked or unlinked reading refuses too, unless what those readings drop all leads it, as a footnote body's markers do (<sup>1</sup>Effective ...), which dates nothing. A block's whole text is its sentence when it is one; otherwise each line that is one dates the block, and lines that date differently remain named alternatives. A range is "effective from", a date, "to" or "through", and a date; it dates a chart from and to, in its own order, when its two dates are the only ones the sentence states, and other dates remain named alternatives. A "to" that no date follows ends no range, and a range that does not end after it starts refuses, one from a date to the same date included.
• A section's date stands until another sentence or State heading. A heading, or a line under 80 characters, that names a date no sentence of its block or section states, in a sentence's date form ("Payment levels as of January 1, 2031", "Effective: January 1, 2031") or one of these others ("Jan. 1, 2031", "January 2031", "1 March 2031", "2031-03-01", "July 1st, 2030", "1st of July", "July 1,2030", "July 1 , 2030", "07.01.2030", "2030/07/01", "7/2030", "as ofJuly 1, 2030"), leaves the date unread, and the next chart refuses unless a later sentence dates it; a date in one of those other forms does so even in the section's own year, apart from an ISO date that equals a date in force. So does a heading that holds a four-digit year from 1900 to 2099 outside the years of the dates in force, and a line under 80 characters that holds a later year, or an earlier one with a dating word ("payment", "level", "effective", "as of" or "rate"), as "Payment levels for 2031" or "Payment levels for 2029" in a section dated 2030. Each line names its own dates, without a revision range in title case, and a heading's whole text, or that of a block under 80 characters, adds the dates it names more often than its lines do, as one split across a <br>, though a revision range on a line still counts among the lines. Outside a heading, a footnote body dates no chart: a line, or a block's whole text, that opens with markers by either rule, every word they drop one to three digits or footnote signs or a single letter.
• Only a claimant line names a chart's claimant: a heading form ("a. Individual payment levels for Nevada", "Couple") or a caption ending with its claimant word after a State date line or effective-date sentence that ends with a date or a State ("Utah Effective 1/1/03 Individual"). A block with a claimant line names every claimant word, or its plural, that any reading of its whole text or lines holds, so a block names at most one claimant type, across a <br> as within a line. A heading, or a line under 80 characters, that holds a claimant word or plural but is no claimant line ("Couple payment levels:", "Eligible couple", "Couples") names a claimant the reader does not read: the next chart refuses unless a later claimant line names one, and one that holds both types refuses as naming two. Prose of 80 characters or more outside a heading names none, though the State guard reads it. Every chart needs its own claimant line.
• Every line passes the State guards, including the lines of a block whose whole text is a State date line, which do not head a section. A line that dates another State refuses, and so does an effective-date sentence, or a line that mentions a claimant, that names any State but its section's. The guards recognize a State date line by the names, D.C. aliases and postal suffixes a section heading takes. A block whose lines head two States refuses, as does a line shaped as another State's heading inside a whole-text date line, and any line shaped as a State heading whose postal code is not that State's.
• Every line of an h1–h6 heading may name only its section's State. A line under 80 characters, in a block of any kind, that opens as another State's caption refuses: after optional leading dashes, bullets, asterisks, quotation marks or parentheses, a list label ("1.", "2)", "(b)"), a heading prefix, "The" and "State of", that State's name, spelling, D.C. alias (read first, so "Washington, D.C." is the District), abbreviation or upper-case postal code, as "Utah (continued)", "(b) Utah", "The State of Utah", "N.J. notes" or "UT (continued)". Any other line under 80 characters that names another State, as "Supplements for Utah", refuses too. Prose of 80 characters or more may name any State.
• "Washington,DC" names the District and "Hawaiʻi" names Hawaii; postal codes, in parentheses or upper case standing as a word, and 14 abbreviations name States to the nets but never head a section. A lower-case word that spells a postal code, as "de" in "de facto", names none. Until the first State heading opens a section, no net that needs one runs.
• The shared HTML tree builds paragraphs, headings, lists, buttons, cells, rows, table sections, captions and colgroups as the WHATWG parsing rules do, end-tag scopes included, so selectors do not depend on omitted or stray end tags. It reads CR LF and a lone CR as one newline, drops a newline that comes right after a pre, listing or textarea start tag, and gives a later html start tag's attributes to the html element, the document root standing for one HTML implies. It does not model the adoption agency rules or HTML's reopening of a formatting element left open, foster parenting, leading content HTML moves into an implied head, a form start tag inside an open form, a self-closing slash on another non-void element, or foreign content (svg and math, which HTML parses by other rules); no pinned capture opens a form inside another. Since HTML ends an svg or math at its breakout start tags (h1–h6, p, div, li, table, b, span and others) and keeps other HTML names there as foreign elements, the tree lets only svg and MathML elements start inside svg or math, and refuses a page on which svg or math ends other than at its own end tag, the end of the page included. It does not model an HTML element built inside an integration point: in <svg><foreignObject><g></svg> HTML ignores the end tag and keeps the svg open, while the tree ends it. A CDATA section holds no text in the tree, though HTML reads one inside svg or math as text. The tree reads a page the same way under Python 3.13 and 3.14: it pins raw text to HTML's six raw-text elements (script, style, xmp, iframe, noembed and noframes; noscript stays markup, as with scripting off), and a plaintext start tag or an abruptly closed empty comment refuses, and so does a script whose text holds <!-- and then a script start tag, which HTML's script data escape states may run on past its next </script>, a self-closed raw text or RCDATA start tag (<script/>), whose slash HTML ignores outside svg and math, and a raw text or RCDATA element inside svg or math whose text holds <, which HTML reads as markup there. A later html start tag only adds the attributes the html element lacks, the tree ignores a noscript end tag with a special element open inside it, and a table part's start tag outside a table builds nothing, as in html5lib.
• No reader function calls itself; the tree builder, the document walk, the line, marker and cell readers and the chart skeleton keep their own stacks. A probe parses and iterates a page nested 300,000 deep under a recursion limit of 200, and a test reads a chart cell nested 50,000 deep in its own process.
• Every chart code has a definition, or poms_open_questions.json records the chart. We infer no definition.
• Payment-validity dates stay distinct from section-revision dates. Conflicting dates stay named options, within one sentence or across a block's lines.
• Narrative state coverage text stays cited text. It never becomes an observed coverage share.
• Totals reconcile within rounding, or keep a named published discrepancy without repair.
• The parsers contain no print or logging calls.

Open items

• POMS:
  • Fifteen Iowa (2006–2011) and Pennsylvania (2008–2010 couple) charts publish codes that no definition table in their section defines. We record them and leave their codes undefined.
  • The District of Columbia's 1999 tables carry two conflicting dates, so the year stays unknown.
  • Two Hawaii couple tables (2022, 2023) each have one malformed amount, which stays missing with named options.
  • Thirty selectors carry 33 published component-versus-total discrepancies. We flag them and correct none, and SSI2 adjudicates them before using them as conformance truth.
• HTML limits, documented in _HTMLTree, _cell_token and the POMS module:
  • The tree does not model HTML's adoption agency rules, its reopening of a formatting element left open (in <h3>w0<b><i>w1</b>w2<h4>w3, HTML's h3 reads "w0w1w2w3" and the tree's "w0w1w2"), foster parenting of content misplaced in a table, leading content HTML moves into an implied head, a self-closing slash on another non-void element, or foreign content. So it refuses a page on which an element other than svg and MathML ones starts inside svg or math, breakout tags and HTML inside svg's foreignObject or MathML's mi included, on which svg or math ends other than at its own end tag, or whose script, style or title there holds <, and it refuses any self-closed raw text tag. The stated limits below name what that rule misses. The 40 pinned HTML captures hold 159 svg icons of path, title and desc elements and no math, so no pinned page refuses.
  • A <br> or nested block in one cell joins its lines with a space, so two amounts stacked that way would read as one space-grouped number. No pinned cell joins two amounts by a <br> within one paragraph. POMS charts stack amounts in paragraphs, which the reader splits into subrows.
  • The POMS nets read text that HTML or CSS hides, other than script and style (the hidden attribute, for one), as text. They never read a caption in a table cell, a chart's own caption element or text outside any block. They over-refuse where an acronym spells a postal code, as "MA" or "IA" in Pennsylvania's section, on a short line that names another State anywhere, on a heading or short line that names two claimant types or a date no sentence states, on a heading that holds a year outside the dates in force, on a short line that holds a later year, or an earlier one with a dating word ("Separate programs for 2029", whose "Separate" holds "rate"), on a number shaped as a dotted date ("Section 2.1.10"), on a footnote marker of two letters or of a digit and a letter, on a heading that opens with a footnote marker, on a link of a marker's size inside a date, and on a range from a date to the same date. No pinned page refuses.
• Stated limits from round 13. These hold on invented pages only. They change no record on a pinned input, and the checks under the thirteenth review's entry show that. None covers a reader over PE-US sources.
  • Integration points: <svg><foreignObject><g></svg>, and the same shape in desc, title, mi, mo, mn, ms, mtext or annotation-xml, ends the svg in the tree while HTML keeps it open, so later blocks and cells read here and stay foreign in HTML.
  • CDATA: a CDATA section holds no text here. HTML reads one inside svg or math as text, as <svg><text><![CDATA[Couple]]></text></svg> reads "Couple", and in HTML content reads the text after its first >.
  • Split date beside a revision line: Effective Dates: 03/01/2031 - Present<br>Payment levels as of March 1,<br>2031 keeps the section's date.
  • Split date in a long block: a date split across a <br> in a block of 80 characters or more, outside a heading, names no date unless the block is an effective-date sentence, so a split date in a year in force keeps the date before it.
  • Dates with no year: "July 1", "July 1st" and "7/1" name no date: the date before them stands, and in a sentence the sentence's other dates stand alone.
  • Other date forms: "July of 2030", "01-Jul-2030", "Jul-01-2030", "July first, 2030", "2030 July 1", "July the 1st, 2030", "July 1, '30", "7 - 1 - 2030" and "1.VII.2030" name no date: the date before them stands unless the year net reads their year, and in a sentence the sentence's other dates stand alone.
  • Footnote bodies: a leading link or sup of one letter or digit makes a footnote body even where it opens a word or a year, as <a href="#top">A</a> new payment level applies from March 1, 2031 or <a href="#y">2</a>031 payment levels, so its date or year dates no chart.
  • Lower-case revision text: Effective dates: 03/01/2031 - Present reads as a sentence and dates the chart, since the label reads in title case only.
  • Title-case label in a long sentence: "Monthly payment levels effective January 1, 2031 (Effective Dates: 01/01/2031 - Present)" names no date, and a long line that holds the label and another State's name passes the effective-date sentence's State guard.
  • Second title-case range: a second "Effective Dates:" range in a section drops from its line and dates no chart.
  • Where the section's revision dates come from: the first title-case range anywhere on a page sets them, a script or a table cell included.
  • Dating words: "Amounts for 2029", "Benefit standard 2029", "COLA 2029", "2029 supplement amounts", "Supplement for 2029", "FBR 2029", "Chart for calendar year 2029" and "In 2029:" hold an earlier year and no dating word, so they keep the section's date.
  • Nested svg: parse_html's raw-text check walks every svg's whole subtree, so nested svg parses in quadratic time: 0.27 s at 4,000 levels, and the review estimates 40 s at 50,000.
  • Refusal message: aggregate_tables.py says "HTML element other than svg or MathML" for svg a and filter, which refuse because they sit outside _FOREIGN_VOCABULARY.
  • Mutants: Z1, Z2, Z3, Z5, Z6, Z7, Z8, Z9, Z10, Z12, Z14 and Z15 of the review's set survive its prescreen of the test module's pages, so no test pins their forms.
• Census Annual Survey of Public Pensions: the data portal refused a plain request (SOURCE_BLOCKED), so we infer no coverage share.
• Part 2:
  • SSA Supplement Tables 7.A–7.E and 3.C6;
  • the SSI Annual Statistical Report 2025;
  • SSI Recipients by State and County 2025;
  • BLS NCS benefits.

Review and fixes

The first review found that the POMS reader kept only the last definition table, so California 2004 lost codes A–F. The fix reads every definition table in the State's section. It also surfaced three more reader problems, now fixed:

• The reader did not recognize the 2009 and 2010 headings "New Jersey (NJ)", so it labelled four New Jersey charts Nevada. The fix renames their registry entries new_jersey.
• Iowa 1999–2005 defines codes jointly ("C/H", "D/I"); each joint row now defines both codes.
• The reader took a full-width note row in Massachusetts 1999 for a code; it now skips the row.
• A State named only in an "X Effective" date line would have kept the previous State's definitions. That path now refuses; no pinned capture uses it.

The fix adds these checks:

• definition codes match an independent raw-HTML locator on all 777 charts;
• registry selectors equal the structural reader;
• each chart's State matches the State its date sentence names (775 charts name one);
• a California 2004 regression;
• Hypothesis properties for definitions split across tables and for joint codes.

Each of the four bugs, reintroduced, fails at least two tests. The birth-evidence seal's exclusions list the five new source files, with a reachability check.

The second review found that a definition table placed between two charts of one State section would have reached only the later chart. The source does not say which earlier chart such a table serves, so the reader now refuses it. No pinned capture has that order, and all 777 charts describe as before. The fix adds:

• an invented regression of that ordering;
• a Hypothesis property over any order of definition tables and charts within a section;
• a stronger raw-HTML differential, which now locates State sections and definition tables itself and checks them against both the registry and the live reader.

The third review found that the reader did not recognize a State heading in a paragraph left unclosed, such as <p>Nevada followed by the chart's heading and table. The shared HTML tree never closes a paragraph implicitly, so the paragraph's text ran on into the chart, and the previous State's selector and definitions carried into the next State's chart. The reader now reads each paragraph or heading from its own text, up to the first nested element whose start tag closes a paragraph in HTML (we checked the set against html5lib 1.1). No pinned capture nests a block in a paragraph or heading (none of 68,336), and all 777 registered selectors, definitions and parsed cells keep their hashes. The fix adds:

• the reviewer's minimized regression, with and without </p>;
• a Hypothesis property that selectors do not depend on omitted paragraph ends;
• a Hypothesis property that a block with no nested block reads its full text.

With the old reader, the regression and the omitted-ends property fail.

The fourth review found that this fix cut a heading's text at blocks that HTML keeps inside a heading, so <h3><div>Nevada</div></h3> again kept California's definitions. It also found that an omitted </p> in a stacked payment cell merged its amounts, that a page of thousands of unclosed paragraphs exhausted the stack, and that the closer set lacked dialog and search. The fix (401b3bc7):

• the shared HTML tree follows the WHATWG "in body" rules for optional end tags: it ends a paragraph in button scope at every start tag that closes one, ends a heading when another starts while it is the current node, and closes an open cell or row at a new one;
• the POMS reader reads a heading's text through nested blocks, reads a div as a block that can carry a State heading, refuses a claimant line that names a State other than its section's, and walks the document without recursion.

Against html5lib 1.1, the new tree agreed on 20,000 random pages of paragraphs, headings and block containers; the old tree disagreed on 6,639.

The fifth review found one medium and three low problems:

• Medium: reading a heading through nested blocks up to the next table meant a heading that holds a block after its words, such as <h3>Nevada<p>Invented note</p>, no longer matched the State pattern, so the previous State's selector carried on. Claimant headings failed the same way. The tree also ended a heading only at its own end tag, where HTML ends it at any h1–h6 end tag.
• Low: thead, tbody, tfoot, caption and colgroup start tags did not end an open cell or row, so a header cell took in the next section's text. This predates the round-4 fix.
• Low: the reader still collected heading text recursively.
• Low: the claimant-State guard refused safe lines: "Washington, D.C." in the District's section, and a claimant line before any State heading.

The fix (8613b313):

• the tree ends the open heading in scope at any h1–h6 end tag, clears the stack back to the table at a caption, colgroup or table-section start tag, and ends an open caption at a row or cell start tag. The fix leaves col out: all 4,045 col tags in the 28 pinned captures that have one sit in a colgroup;
• the POMS reader checks a heading's whole text, then each line split at its nested blocks, for the State, the claimant and the effective date. A whole-text "State Effective …" line stays a date line, and a heading whose lines name two States refuses;
• "Washington, D.C." reads as the District of Columbia, and the claimant-State guard waits for the first State heading;
• the readers read heading text and cell text without recursion; a cell nesting 1,500 spans had raised RecursionError too.

Nothing changes on the pinned captures: the tree comparison builds 44 identical trees, and all 874 registered entries hash identically. (Round 4's commit called these "43 HTML captures". The registry holds 41 HTML captures, and the comparison's byte filter also admits two xlsx workbooks and a zip archive.) Against html5lib 1.1, on 5,000 invented pages per mode, the new tree has no mismatch in headings (401b3bc7: 506) or in tables with optional cell, row, section and caption ends (401b3bc7: 1,230). Each of 27 mutants of the tree and the POMS reader fails at least one test, as does each of five mutants of the cell reader.

The fix adds, on invented HTML only:

• the reviewer's State and claimant heading cases: blocks after the heading's words, open headings and mismatched end tags;
• an effective line after heading words, a table inside an open heading, and a heading whose lines name two States;
• deep headings, paragraphs and cells;
• a claimant line before any State heading, and Washington, D.C.;
• table sections, captions and colgroups ending open cells and rows, and heading end tags ending any heading in scope but stopping at scope;
• a property that wrapping or ending heading text keeps the selectors;
• a differential property of the cell reader against the recursive one.

The sixth review (delta, at 697ea340) found one medium and three low problems:

• Medium: in a heading, a line from a nested section, list item or blockquote set the chart's date and claimant without the State guards that the same line passes in a paragraph, and the result changed with where </h3> sat.
• Low: _payment_rows deep-copied the chart, and ElementTree's copy recurses, so a cell nesting 50,000 spans crashed the interpreter. The recursion invariant above was therefore false.
• Low: a State heading with a <br>, a zero-width character or another State's postal code was misread, and so was a "Washington, D.C." heading.
• Low: mutants of the new tree rules survived the tests.

The fix (4aae2820):

• every line of a block passes both State guards, and the reader reads every block that ends a paragraph, apart from a table or rule, from its own text. A heading whose first line names a State other than its section's refuses;
• the reader reads the chart through a skeleton built without recursion, whose cells hold keys to the original cells. A table nested in a cell stays that cell's text, as before;
• lines end at a <br> and drop zero-width characters. A postal code after a State's name must be its own, and "Washington, D.C." and "Washington DC" name the District in headings;
• the tree builds lists, buttons, nested tables, captions, colgroups, cols and implied table sections and rows as HTML does, and closes end tags by HTML's scope rules.

Nothing changes on the pinned captures: every registered reader's output, cell values included, hashes as at 697ea340. On the 40 pinned HTML captures, the tree's tables, block texts and nesting now equal html5lib 1.1's on all 40 (697ea340: 40, 40 and 32). That comparison removes comments and noscript from both trees. On 5,000 generated pages per mode, the tree has no heading or table mismatch with html5lib (697ea340: 1,561 and 3,969). Each of 46 mutants, one per new or changed rule, fails at least one test. 51 of the 157 tests in test_aggregate_poms_invented.py fail against 697ea340's modules.

The fix adds, on invented HTML only:

• the reviewer's nested-line cases, for a section, list item and blockquote, with and without </h3>, and a property that selectors do not depend on heading ends or block kinds;
• the claimant guard on every claimant line, and a <br>-separated line that dates another State;
• headings with <br>, </br>, </p>, zero-width characters, postal codes and Washington, D.C.;
• headings and paragraphs that name another State, refused or admitted by the first-line rule;
• 12 table pages and 13 block pages whose expected trees are html5lib 1.1's;
• a chart cell nested 50,000 deep, read in its own process, and a table nested in a chart cell.

The seventh review (delta, GPT-6.1 Sol, at 4aae2820) found three medium and three low problems and one note:

• Medium: a block whose whole text is a State date line skipped its lines, so a later line could date or name another State.
• Medium: a claimant line that named another State passed whenever it also named the section's own.
• Medium: an effective-date sentence other than "State Effective ...", or one with a D.C. alias or postal suffix, carried another State's date into the section.
• Low: a space at each inline element's edge split words such as "Utah".
• Low: a later heading line shaped as a State heading, with another State's postal code, kept the section's selector.
• Low: three tree mutants survived the tests: list items around an address, dd and dt ending each other, and </html> after an html start tag.
• Note: the tree's WHATWG wording covered nested forms, which it does not model.

The fix (89b40e46):

• every line of a block passes the State guards, and a whole-text State date line's lines still head no section;
• claimant lines and effective-date sentences may name no State but the section's, and the guards recognize State date lines by the heading reader's names, aliases and postal suffixes;
• inline elements join without a separator within a line;
• any line shaped as a State heading with another State's postal code refuses;
• tests pin the three tree cases to html5lib 1.1's blocks, and the tree's docstring lists nested forms among what it does not model.

The pinned 2011 California section's date sentence ends charts<span class="poms-bold">:</span>, which the old line reader read as "charts :". Its two charts' parameter_effective_date_text in poms_registry.json now read "charts:". Every registered reader's output, cell values included, hashes as at 4aae2820; only that page's selector digest changes. The html5lib comparisons hold as before: 40 of 40 pinned captures, and no mismatch on 5,000 generated pages per mode. Against 4aae2820's reader, 18 of the 187 invented POMS tests fail, the same 18 in each of three Hypothesis seeds; the new tree cases and the later-claimant-line case pass there, since 4aae2820 already built those trees and refused that line. This round's 14 mutants (G1–G11 for the guards, O1–O3 the review's tree mutants) and round 7's 41, rerun on the final code, all fail tests. P11 (the claimant guard stops at its first claimant line) passed at first, because a whole text naming another State now refuses on its own; a block too long to read as one claimant line, with its own State's claimant line first, now catches it.

The fix adds, on invented HTML only: the review's four repro fragments and their variants, refused or admitted, including long blocks whose whole text is no effective-date or claimant line, one with the section's own claimant line before another State's; controls that read the section's own State, date and claimant; an effective-date line before any State heading; and five tree pages whose expected blocks are html5lib 1.1's.

The eighth review (delta, Opus 5.5, at 93824c2c) found two medium and four low problems and three notes, and CI failed on Python 3.14.7:

• CI: Python 3.14's HTMLParser reads xmp, iframe, noembed and noframes as raw text and turns the rest of the page into text after a plaintext start tag, so an invented paragraph-closer test failed there.
• Medium: joining inline text without a separator hid a State name that touches a word character across an inline tag, such as a footnote sup, from every State net; 4aae2820 had refused those pages.
• Medium: in a block whose whole text is a State date line, a line shaped as another State's heading kept the section's selector.
• Low: three mutants (R1, R3, R5) survived the tests.
• Low: claimant selection was silent and inconsistent: a block whose claimant lines named two types read one of them, and the claimant carried over after a chart.
• Low: two of round 7's own boundary variants still read wrong without refusing, because no net checked heading lines after the first for another State.
• Low: "Washington,DC" without a space did not name the District.
• Notes: Unicode format characters, "Hawaiʻi", a postal code alone in a date sentence and "N.J."; a mid-body html start tag; and the round-8 count of round 7's mutants.

The fix (dfb30420):

• pins the raw-text set to HTML's six elements, refuses plaintext and abruptly closed empty comments, and closes the parser at the end of the page, so the tree is the same on 3.13 and 3.14;
• reads each block and line joined, bare and spaced, and runs every State net over all three;
• refuses a line shaped as another State's heading inside a whole-text date line;
• refuses a block naming two claimant types, and resets the claimant after each chart;
• nets every line of an h1–h6 heading, reversing an earlier deliberate admission of a second heading line that names another State;
• matches "Washington,DC" and "Washington,D.C.", strips Cf characters, reads "Hawaiʻi", and reads postal codes and 14 abbreviations as naming States without heading a section;
• models a mid-body html start tag as html5lib does, and, found by the extended differential, a noscript end tag with a special element open inside it and a table part outside a table.

Every registered reader's output, cell values included, hashes as at 93824c2c on both Python versions, and all 777 pinned charts' selectors equal 93824c2c's. On 3.13, 92 of the 300 invented POMS tests fail against 93824c2c's reader, the same 92 in each of three Hypothesis seeds. mutants_r9.py's 60 mutants fail tests but one, H1, which nets only a heading's whole text and first line; the whole text holds every line's text, so it differs only for an alias split across lines. Of earlier rounds' mutants, the tests catch all 7 of mutants_r8.py's and all 31 of mutants_r7.py's that still apply; we ported the rest into mutants_r9.py, or they match a caught mutant. (Round 8's "round 7's 41" meant the 41 of mutants_r7.py's 46 that applied then.)

The round-9 build listed a prose line holding a claimant word, such as "A recipient or couple qualifies if ...", as able to set the next chart's claimant, and counted 70 such line forms on the pinned captures. That overstated the exposure. Those 596 lines are the block lines that hold a claimant word and then "or" under the old pattern. The 592 in table cells, which the claimant reader never reads, each hold two claimant words, one of each type; the other 4 are one h4 heading form with one claimant word and the source's "or" for "for", a claimant line under the new rules too. Of the 777 pinned charts, 594 take their claimant from h1–h6 lines, 99 from short whole-text paragraphs and 84 from State date captions; none takes it from prose. Round 10 makes prose name no claimant (below), and all 777 charts read as before.

The ninth review (delta, Opus 5.5, at 7adafa66) found three medium and two low problems and seven notes:

• Medium A: prose set a chart's claimant: any line holding a claimant word was a claimant line, and a line naming two types read as its first. A test pinned "Individual or couple for Washington, D.C." as individual.
• Medium B: the reader read the date sentence joined only, so a footnote marker inside it changed the date (January 1<sup>2</sup>, 2031 read January 12) or hid the sentence, silently.
• Medium C: the bare reading dropped any sup, and any link whose href held "fn", so link or sup text vanished and prose read as another State's heading. Round 8 introduced this.
• Low D: five of the reviewer's six mutants survived the tests: O1, O2, O3 and O6, and O4, which the reviewer judged equivalent.
• Low E: the nets checked h1–h6 lines for another State, but not paragraph captions, though 183 of the 777 pinned claimant setters are paragraphs: <p>Utah (continued)</p><p>Couple</p> charted Nevada.
• Notes: acronyms over-refuse in Pennsylvania's section; mutant H1 is equivalent but for a D.C. alias split across lines; the reader read script and style text; a census log printed amounts; the open claimant paragraph overstated exposure; the html5lib generator lacked comments, inline elements, the implied-html attribute case, table parts outside a table, RCDATA and upper-case raw-text tags; one mutant slice started above the load limit.

The fix (8a772f5a):

• only a heading form or a date caption is a claimant line; prose names no claimant but still passes the State guard, and every claimant word of a claimant line counts, so a block naming two types refuses within a line as across lines. The Washington, D.C. test now states that "Individual or couple for Washington, D.C." names no claimant, so its chart refuses, and that single-type lines read;
• a date sentence whose bare reading states other dates, or another range reading, refuses, and so does a line that is a dated sentence only when bare, unless its markers lead it. The pinned 2023 District of Columbia footnote body between two charts still dates nothing;
• the bare reading drops only a sup or # link holding one to three word characters or footnote signs; _cell_token keeps its wider rule;
• a short line in any block that opens as another State's caption refuses. The net reads the D.C. alias first, codes and abbreviations included, and runs after the other nets, so their messages stand;
• every reading drops script and style text;
• the html5lib generator gains the six gaps. It found three tree gaps on pages with no misnested formatting element, now modelled and tested: a later html start tag's attributes on a page with an implied html element, the newline HTML drops after pre, listing and textarea, and CR LF read as one newline. The tree's limits now also name HTML's reopening of formatting elements, which nests a heading inside one left open.

Every registered reader's output, cell values included, hashes as at round 8 on both Python versions, and all 777 pinned charts' selectors equal 7adafa66's; no pinned page refuses. On the 40 pinned HTML captures the tree's tables, blocks and nesting equal html5lib 1.1's. On 3.13, 94 of the 434 invented POMS tests fail against 7adafa66's reader, the same 94 in each of three Hypothesis seeds. All 41 of mutants_r10.py's mutants fail tests: at least one per new rule or guard, and ports of the round-9 and round-7 mutants whose snippets this round removed. A mutant of the marker pattern's three-character bound alone survived, because the marker walk capped text at three characters too; both now read one constant, and the tests catch its mutant. The review's O4 is not equivalent: a footnote body opening a State date line, followed by a line shaped as that State's heading, refuses on the head and reads under O4, and a test now pins it. Of the older sets, the tests catch the mutants whose code still exists: 5 of 5 from own_mutants_r9.py, 53 of 54 from mutants_r9.py (H1 survives, as in round 9), 7 of 7 from mutants_r8.py and 30 of 30 from mutants_r7.py.

The tenth review (delta, Opus 5.5, at 8a772f5a) found two medium and four low problems and six notes:

• Medium 1 (regression): a block with a claimant line counted claimant words in its claimant lines only, so <h4>Individual<br>or couple</h4> read as individual and <h4>Couple<br>Eligible individual</h4> as couple.
• Medium 2: a heading the claimant pattern did not admit, such as "Couple payment levels for Nevada:", "Eligible couple" or "Couples", left the claimant line before it standing, so the chart read the earlier type silently.
• Low 1 (regression): a footnote marker holding a script or style joined a date (January 1<sup>5<script>x</script></sup>, 2031 read January 15).
• Low 2: the caption net missed list labels and leading punctuation ("1. Utah", "(b) Utah", "- Utah", "(Utah)", "The State of Utah").
• Low 3: a script holding <!--<script> ended at its first </script> in the tree, though HTML runs it on.
• Low 4: three of the reviewer's mutants (M21, M4, M16) and round 9's H1 survived.
• Notes: a footnote link by a full or relative URL changed a date; a long block took its first of two date sentences; the Limits and Not modelled lists missed several forms; two sentences of this PR's review text were wrong (corrected above); the in-cell newline rule; new date refusals.

The fix (85958483):

• a block with a claimant line counts the claimant words of every view, its whole text included, and plurals count;
• a heading or short line that holds a claimant word but is no claimant line names a claimant the reader does not read, so the chart refuses unless a later claimant line names one; one with both types refuses as naming two. The Washington, D.C. test's "Individual or couple for Washington, D.C." now refuses as naming two claimants;
• the same shape, swept for the other selectors: a short line naming another State anywhere refuses, and a heading or short line naming a date no sentence states, in a sentence's date form or the four other forms the docstring lists, leaves the date unread so the next chart refuses (the sweep's table is in the evidence folder); round 11 found common forms this missed, which round 12 nets (below);
• a marker's text drops script and style; the caption net takes list labels, leading punctuation and "The"; the tree refuses a script holding <!-- and then a script start tag;
• the date checks read an unmarked reading, which drops any element either marker rule counts, so a footnote link whose href holds "fn", by a full or relative URL, changes no date silently; round 11 found that a link to a footnote whose href holds no "fn" still did, which round 12 fixes (below);
• lines of a long block that date differently remain named alternatives, and a range dates a chart only with two dates;
• the module's Limits and Not modelled list the caption nets' over-refusals, a chart's own caption, text outside any block, the script escape states and the new date refusals.

Every registered reader's output, cell values included, hashes as at round 8 on both Python versions, and all 777 pinned charts' selectors equal 8a772f5a's; no pinned page refuses. On the 40 pinned HTML captures the tree's tables, blocks and nesting equal html5lib 1.1's. On 3.13, 94 of the 548 invented POMS tests fail against 8a772f5a's reader, the same 94 in each of three Hypothesis seeds, over 19 test functions. Of mutants_r11.py's 49 mutants (at least one per new rule or guard, and ports of the earlier sets' mutants whose code this round rewrote), 40 failed tests at 85958483, and nine passed all six modules, one of them (R21) equivalent. 858fb546 adds 15 test cases on invented pages, one to three per survivor; at that commit 48 of the 49 fail tests, each of the eight first failing its own witness. R21 walks script children whose text no reading uses, so no test can tell it apart. Every mutant of my_mutants_r10.py, mutants_r10.py and own_mutants_r9.py whose code still exists fails tests (8, 30 and 3), M4, M16, M20 and M21 among them; the review's H1 differs only where the reader over-refuses, so no test can kill it without pinning that over-refusal. Rerun at 858fb546, the three oldest sets give: of mutants_r9.py's 60, 50 fail tests, 9 change code that no longer exists and H1 passes all six modules; of mutants_r8.py's 14, 7 fail tests and 7 change code that no longer exists; of mutants_r7.py's 46, 30 fail tests and 16 change code that no longer exists; the W2a and W2b ports fail tests too.

The eleventh review (delta, Opus 5.5, at 858fb546) approved it with six low findings and seven notes:

• Low 1: the unread-date net exempted every date that a footnote body or revision range named, across its whole block, and took any leading sup or link whose href holds "fn" for a footnote marker, so short lines such as Payment levels as of March 1, 2031<br><sup>1</sup>Revised March 1, 2031 kept the section's earlier date silently. That made the Invariants' "never reads a wrong one silently there" false.
• Low 2: a script or style inside svg or math is raw text in the tree but markup in HTML, and Not modelled did not say so.
• Low 3: the range pattern read a prose "to" as a range's end, so "effective from January 1, 2031 apply to facilities licensed after July 1, 2030" dated a chart from 2031 to 2030.
• Low 4: the net missed common date forms ("January 1st, 2031", "Payment levels for 2031", "01.01.2031"), while round 10's fix list and commit message claimed that it read every form.
• Low 5: five claims overstated: the HTML-limits sentence on hidden text and captions, round 10's claims that the net read every date form and that a footnote link changed no date whatever its URL, the Invariants' "never reads a wrong one silently there", and a stale count of mutants_r7.py's results (all corrected above); the tree's docstring said HTML's escape states run a script on where they may not.
• Low 6: eleven of the reviewer's fifteen mutants survived all six modules.
• Notes: H1's random part never met a differing page; <script/> hides what follows in HTML but not in the tree; a short line that repeats the claimant in force refuses; a claimant named in other words keeps the earlier type; a range to the same date drops its end and a reversed range reads literally; padding can push a short line past 80 characters; process.

The fix (8efc4b65, with tests in 31ae38ef and docstring wording in 5bc9cb6f, and witness cases in 32cc85d0):

• the net names each line's dates on its own, without the revision range's text, skips footnote-body lines, and adds the whole text's dates only where no line names them; a footnote body needs every word its markers drop to have a marker's size, and a block whose whole text opens as one dates no chart;
• the tree refuses a page whose svg or math holds a raw text or RCDATA element with < in its text, and a page with any self-closed raw text or RCDATA start tag; a census finds neither in the 40 pinned HTML captures, and the tree's docstring lists foreign content under Not modelled;
• a range needs a date right after "from" and right after "to" or "through", gives its two dates in its own order, and refuses when it does not end after it starts, a range from a date to the same date included;
• a heading that holds a year outside the years of the dates in force, or a short line that holds a later one, leaves the date unread. Counting every year in short lines refused one pinned page, whose short line on a State-administered program names the year before its chart's, so short lines outside a heading count later years only; the docstring lists the forms, that limit and "January 1, 31";
• the date checks read a fifth, unlinked reading, so a link of a marker's size inside a date refuses whatever its href holds;
• the tree's docstring says the escape states may run a script on, and names the forms its check refuses beyond them (a "-->" between, a vertical tab or Unicode space, a long s); the PR's claims above now say what holds;
• tests pin each surviving mutant's witness, both sides of each 80-character bound included, and the docstring's limits name Note 4's claimant vocabulary.

All 777 pinned charts' selectors equal 858fb546's on 3.13 and 3.14 (28 files the same, 0 differ, 0 refused, 777 charts, on each), and every registered reader's output, cell values included, hashes as at round 8 (874 specs, byte-identical on 3.13 and 3.14). 61 of the 667 invented POMS tests at 5bc9cb6f fail against 858fb546's readers, over 15 test functions, none of them one 858fb546 defines. Of mutants_r12.py's 45 mutants (29 for the new rules, 15 from the review's set ported, with X4 split into X4 and X4b for the whole text's bound), 44 fail tests at 5bc9cb6f; S24, which let any four digits count as a year, passed all six modules until 32cc85d0 added witness cases. Rerun on 5bc9cb6f's readers, which 32cc85d0 leaves unchanged, the older sets give: of mutants_r11.py, 34 fail tests, 14 change code that no longer exists and R21, the equivalent one, passes; of mutants_r10.py, 28 fail tests and 13 change code that no longer exists; of my_mutants_r10.py, 7 and 3; of mutants_r9.py, 49 and 9, H1 passes for the reason round 11 checked, and V1 now stops the module loading, since _TEXT_CONTENT reads the set it deletes; of mutants_r8.py, 7 and 7; of mutants_r7.py, 30 and 16; and W2a and W2b fail tests. Of round 9's own set, on the same readers, 3 fail tests and 3 change code that no longer exists.

The twelfth review (delta, at 32cc85d0) approved it with five low findings and six notes:

• Low 1: dates in the section's own year, in forms neither date pattern read ("July 1st, 2030", "07.01.2030", "2030/07/01", "July 1,2030", "7/2030", a month glued to the word before it, and such a second date in a short date sentence), kept the section's date silently. The year net counts only years outside the dates in force, so the Invariants' "never reads a wrong one silently there" failed on forms the docstring did not list.
• Low 2: the tree built HTML block and cell names inside svg or math (<svg><section>, <svg><td> and others) as HTML, where HTML keeps them foreign, and ended an h5 at an end tag inside MathML's mi, which HTML ignores there, so a date, State or claimant dropped out of its block, and a chart row's amounts moved one column on.
• Low 3: footnote-body detection took any word of one to three characters that a leading sup or footnote link held for a marker, so <sup>As of</sup> March 1, 2031 and <a href='#new'>New</a> payment levels as of March 1, 2031 dated no chart.
• Low 4: six of the reviewer's twelve mutants (Q1, Q4, Q5, Q8, Q9 and Q11) passed the invented module; Q9's witness, <script src='x.js'/>, is the common real form.
• Low 5: five claims overstated: the Invariant above, the over-refusal sentence on years in short lines, the tree's Not-modelled sentence on which start tags end an svg, "28 files" for the 874-spec fingerprint, and two commits for one rerun (all corrected above).
• Notes: the breakout pages, all classified (the Tests section gives the counts); an earlier year in a short line left a wrong date silent; three line-by-line cases (a split date that a footnote line repeats, a comma after a <br>, a heading that opens with a marker); a lower-case "Effective dates:" line dated a chart; the new over-refusals; process.

The fix (1b319dca, with tests in be276733 and 179496ec and docstring wording in 17bdcf00):

• the unread-date net names six more forms, the section's own year included: an ordinal day, a comma with no space after it or a space before it, a dotted date, a date that opens with its year, a month and year in numbers, and a month glued to the word before it;
• a short line's earlier year counts when the line holds a dating word ("payment", "level", "effective", "as of" or "rate"); the pinned line on a State-administered program holds none and still dates no chart;
• inside svg or math the tree lets only svg and MathML elements start, and svg or math must end at its own end tag, the end of the page included, or the page refuses; the Not-modelled item names HTML's breakout start tags as html5lib 1.1 builds them;
• a footnote body's dropped words must be one to three digits or footnote signs, or a single letter, and a heading is never a footnote body;
• a block's whole text adds the dates it names more often than its lines do, so a footnote line that repeats a split date no longer covers it;
• the revision range, and the date sentence's exclusion of it, read "Effective Dates:" in any case;
• tests pin the six surviving mutants' witnesses, with new witnesses for Q8 and for the caught Q7, whose old witnesses now refuse with or without them, and the docstrings list the new over-refusals and two more forms left standing: a date split across two blocks, and a long date sentence whose second date takes another form.

All 777 pinned charts' selectors equal 32cc85d0's on 3.13 and 3.14 (28 files the same, 0 differ, 0 refused, on each), and every registered reader's output, cell values included, hashes as at round 8 (874 specs, byte-identical on 3.13 and 3.14). All 40 pinned HTML captures parse; their 159 svg hold only path, title and desc elements, and none holds math. 72 of the 774 invented POMS tests at 17bdcf00 fail against 32cc85d0's readers, over 12 test functions, none of them one 32cc85d0 defines. Of mutants_r13.py's 50 mutants (38 for the new rules and 12 from the review's set, ported), 49 fail tests at 17bdcf00, and Q2, the equivalent round 12 found, passes. In process, 49 of the 50 change the outcome of a page their tests build, all but Q2. Rerun at 17bdcf00, the older sets give: of mutants_r12.py, 31 fail tests and 14 change code that no longer exists; of mutants_r11.py, 33 and 15, and R21, the equivalent one, passes; of mutants_r10.py, 28 and 13; of my_mutants_r10.py, 7 and 3; of mutants_r9.py, 49 and 9, H1 passes for the reason round 11 checked, and V1 stops the module loading; of mutants_r8.py, 7 and 7; of mutants_r7.py, 30 and 16; and W2a and W2b fail tests. Of round 9's own set, 3 fail tests and 3 change code that no longer exists. Of the round-12 review's 12, 8 fail tests, 3 change code that no longer exists and Q2 passes, as above. be276733's message says two removed test cases each held a second dating word; "Effective in 2029" held one.

The thirteenth review (delta, at 17bdcf00) returned REVISE with one medium finding, six low findings and seven notes:

• Medium 1 (regression): 1b319dca answered round 12's Note 4 by reading "Effective Dates:" in any case. A long sentence that held "effective dates:" then stopped dating its chart and stopped meeting the State guards, so a page that named Utah in Nevada's section read as Nevada, where the parent refused. A range in another case, in script text or a table cell, set the section's revision dates.
• Low 1: inside foreignObject, html5lib ignores </svg>, so later blocks and cells stay foreign while the tree reads them. In a browser the same holds for desc, title and the MathML token elements.
• Low 2: the tree drops CDATA inside svg or math, which HTML reads as text.
• Low 3: the split-date rule counted a revision line's date on the lines' side only.
• Low 4: date forms that still read wrong, a split date in a long block, and two claims that overreached.
• Low 5: prose that still makes a footnote body.
• Low 6: twelve of the reviewer's seventeen mutants survived a prescreen.
• Notes: dating words; nested svg's quadratic parse time, which the builder's report had missed; a refusal message's wording; annotation-xml's second HTML encoding; the ISO form of the date in force; the equivalences it checked; process.

Under the project's review exit rule, a regression against the parent gets a fix, and an invented-input finding with zero effect on pinned inputs stands as a stated limit. So only Medium 1 changes code. The fix (8cb397f5):

• _effective_line's exclusion, _REVISION and the "Present" check read title case again, the three lines as 32cc85d0 wrote them. Round 12's Note 4 returns as a stated limit. Each of the 29 pinned POMS pages holds a single revision range, in title case;
• tests pin the review's E4, ES1, V6 and V7 and the lower-case revision line, and replace the four cases that pinned the any-case reading;
• the docstrings and the claims above no longer say what Lows 1–5 and Notes 4 and 5 falsified. Open items lists the rest, one line per form.

All 777 pinned charts' selectors equal 32cc85d0's on 3.13 and 3.14 (28 files the same, 0 differ, 0 refused, on each), and every registered reader's output, cell values included, hashes as at round 8 (874 specs, byte-identical on 3.13 and 3.14). All 40 pinned HTML captures equal html5lib 1.1's tree. The review's 16 regression probes read alike on 8cb397f5 and 32cc85d0. aggregate_tables.py's syntax tree matches 17bdcf00's, docstrings aside, so the html5lib differentials below stand. 68 of the 775 invented POMS tests fail against 32cc85d0's readers, over 10 test functions, none of them one 32cc85d0 defines, and 5 fail against 17bdcf00's, over the four new functions. Of mutants_r13.py's 50 mutants, 46 fail tests, R1, R2 and R3 change code that no longer exists, and Q2, the equivalent round 12 found, passes. The review's twelve mutants pass all six modules, as it predicted. Only aggregate_poms.py's three lines changed, so we did not rerun the older sets.

A census of the pinned captures backs the limits. None of the 40 HTML captures holds a CDATA section. Each of the 29 POMS pages holds one revision range, in title case, and no "effective dates:" in another case. None of the listed date forms appears on any of the 4,263 lines under 80 characters that the reader's own _lines gives outside a table cell. The net also reads heading lines of any length: outside tables, in four readings, 14 lines match the date net's shape, each a page-title h1 of 97 to 100 characters before the first State heading, where no net runs.

The fourteenth review (delta, at 8cb397f5, an Opus 5.5 reviewer after the Sol lanes closed) approved, with no high or medium finding, one low finding and five notes:

• 8cb397f5 restores the three lines round 13's medium named, each equal to 32cc85d0's text. On any page, the head gives 32cc85d0's read or refuses, and the review's own runs read 28 files the same, 0 differ, 0 refused and 777 charts on 3.13 and 3.14.
• Low 1: four stated-limit lines named the right form and misstated its effect. They now read as the Stated limits above read. The head and the parent agree on every row, and no pinned input holds these forms.
• Note 1 adds CDATA in HTML content to the limit. Note 2 widens the census line. Note 3 adds the refusal message's limit. Note 4 records that the restored "Present" check is an equivalent mutant: title-case _REVISION captures only "Present" or a date. Note 5 records the reviewer's own two runs above the load gate.
• This entry changes the PR body only, so the head stays 8cb397f5.

The earlier slot turn merged master 18082859 (ce188f20) and recounted the tiers (5285c76d). Master's tier files replaced the branch's, and the recount added D4's 794 unit and 256 artifact tests. tests/tier_counts.json then read unit 20,458, artifact 4,527, integration_psid 1,471, reproduction_legacy 520 and oracle_policyengine 308, 27,284 in all. CI's tier-policy test at 8cb397f5 collected those unit and artifact counts on the PR's merge with master, against a manifest of 19,664 and 4,271 (run 41344829714).

A comparison of blob hashes from 8cb397f5 to 5285c76d finds 46 moved paths: the two tier files and 44 other paths that master changed since the merge base 0f69bbb2. All 44 match master's blobs. None of D4's 79 paths moved, and no path changed on both sides.

CI at 5285c76d passed every check but one test in shard 3, tests/test_track_b_b0_2_counts.py::test_a_failure_from_outside_this_script_prints_no_value, which ran first in that shard. CI at 8cb397f5 failed the same way. #596 ("Flush the real streams before each B0.2 test captures output") merged ahead of this branch as dd9c7be2, and 35859218 carries it. CI at 35859218 passed every check, shard 3 included.

ecc09804 merges master 89d064b5, which brings #573's SSI interface. #573 and D4 each added a block to the same exclusion tuple in scripts/first_estimates_birth_evidence.py. They also each added a block to the same reachability test in tests/estimates/test_birth_evidence_artifact.py, so both files conflicted. The merge keeps both blocks and lists the shared income_resources/__init__.py once, in D4's block. Against master, it adds the 7 and 20 lines D4 added to these files, and it drops only master's copy of that __init__ line. d5dd388f recounts on 89d064b5: unit 21,248, artifact 4,527, integration_psid 1,471, reproduction_legacy 520 and oracle_policyengine 337, 28,103 in all. 35859218 merges master dd9c7be2, which changes one test module for #596 and leaves the tier manifests unchanged, so the recount stands.

From 5285c76d to d5dd388f, 159 paths changed: the two tier files, the two conflicted files and 155 other paths that master changed since 18082859. All 155 match master's blobs. None of D4's other 77 paths moved. From d5dd388f to 35859218, one path moved: the test module #596 changed, now equal to master's blob.

38895135 merges master edba88f7, which brings #563's Track B changes and adds 2,405 unit tests to master's count. Only the two tier files conflicted, and the merge takes master's. ca93c707 recounts on edba88f7: unit 23,653, artifact 4,527, integration_psid 1,471, reproduction_legacy 520 and oracle_policyengine 337, 30,508 in all. From 35859218 to ca93c707, 63 paths changed: the two tier files and 61 other paths that master changed since dd9c7be2. All 61 match master's blobs. None of D4's paths moved.

Tests

D4's six modules (test_aggregate_coverage, _numbers, _pensions, _poms, _poms_invented and _ssa): 1,050 passed on Python 3.13.9 and 1,050 on 3.14.7, up from 944. On the merge of 89d064b5, all 1,050 passed again on Python 3.13. The birth-evidence seal tests' reachability test needs modules this sparse clone leaves out, so CI runs it, as it runs tests/track_u2/test_u2_isolation.py. black -l 79 and ruff are clean.

Tier	Added
unit	794
artifact	256

• Unit tests use invented inputs.
• Artifact tests read the pinned public captures.
• Microdata: D4 reads none and captures no Trustees Report cost table.
• Tier counts: tests/tier_counts.json and tests/README-tiers.md carry the recount on master edba88f7: unit 23,653, artifact 4,527, total 30,508.
• html5lib differentials: round 9's and round 7's generators find no mismatch on 5,000 pages per mode on both versions (round 7's rerun at 858fb546). Round 11's generator, round 10's with the script escape form added from its own random stream, refuses each page whose script html5lib runs on (32 headings and 72 table pages per 5,000) and finds no mismatch on pages without a formatting element left open or misnested (3,793 headings and 2,397 table pages compared); on pages with one, the differences are the adoption agency and reopening rules the tree does not model. The differential counts apart the pages that differ only by a newline right after a pre, listing or textarea start tag in a table cell or caption, which the tree drops, as it does everywhere, and html5lib 1.1 keeps there (18 headings and 90 table pages per 5,000). Round 12's generator, round 11's with svg and math content and self-closed raw text tags added from a third random stream, refuses 149 headings and 555 table pages for raw markup inside svg or math and 140 and 639 for a self-closed tag, and finds no mismatch on the 3,507 headings and 1,869 table pages with no formatting element left open or misnested and no svg or math breakout. Of the 132 headings and 380 table pages with breakout markup, 19 and 83 differ. Of the 19 heading pages, 9 are breakout, each a <math><h4> inside an h3, h4 or h5, which html5lib moves out, and 10 differ by the adoption agency rules alone; all 83 table pages differ by the adoption agency rules alone. At the reader's level, the 8 breakout pages with no formatting markup read the same, and page 919 differs only by an extra empty p. Round 11's generator gives the round-11 review's logs byte for byte on the round-12 tree. Round 13's generator, round 12's with the review's foreign forms added from a fourth random stream (<svg><section> and the other block names HTML keeps foreign, <svg><td> in a cell, and <math><mi> before a heading end tag), draws round 12's pages unchanged where it draws no round-13 form. On the round-13 tree it refuses 616 headings and 2,598 table pages because an element other than svg or MathML starts inside svg or math, 107 and 139 of them round 12's breakout pages, which the tree no longer reads. It refuses 181 and 79 because svg or math ends other than at its own end tag, and finds no mismatch on the 3,104 headings and 1,201 table pages with no formatting element left open or misnested. Of the 637 headings and 507 table pages with one, 133 and 152 differ, by the adoption agency and reopening rules. The 4 headings and 16 table pages it draws with breakout markup that the tree still reads hold no svg or math in html5lib's tree, because raw text took the markup. One of those table pages differs by the adoption agency rules, on a misnested <b><code></b></code> ahead of the noframes that holds its svg, and another differs only by the newline html5lib 1.1 keeps, one of 12 headings and 20 table pages that differ that way. Round 11's generator gives the round-11 review's logs byte for byte on the round-13 tree.

axiom: n/a: pinned public aggregate tables and their readers for microsim validation; no PolicyEngine country-model or RuleSpec encoding changed

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
social-security-model	Ready	Preview	Nov 2, 2026 9:42pm UTC

[MaxGhenis]

The Microcosm Dynamics coordinator merged this PR at ca93c707dab8d26ef676c92b48d1333f5923f7fb with --merge --match-head-commit, as 97003ac27229a59617ff31eb0f15e5e31b2734bd (10:44Z, 2026-11-03).

• Review. GPT-6.1 Sol reviewed 35859218 in Subfleet job 20261102-115756-d4-review-r15 and approved it, leaving no finding of medium or higher severity open. In its three low findings, GPT-6.1 Sol asked for wording in this PR's description, and the income-and-resources lane added that wording at 10:41Z.
• After the review. The lane merged master edba88f7 in 38895135 and recounted the test tiers on the merged tree in ca93c707. Of the 63 files that changed between 35859218 and the head, 61 match master's copies, and the lane had changed none of them before the review. In the other two, tests/tier_counts.json and tests/README-tiers.md, the lane recorded 23,653 unit and 4,527 artifact tests, 30,508 in all.
• CI. gh pr checks 574 exited 0 at the head, with 11 checks passing and deploy skipped (Tests run 41530649476). In that run, tests/test_tier_policy.py compared the committed counts with the collected suite and passed. GitHub reported the PR mergeable.
• The slot. Before the next PR in the slot lands, its owner merges master 97003ac2, or later, into it and recounts the test tiers.