Skip to content

Add the Autumn Budget 2026 registry and harvest (tranche 1 of #230) - #231

Merged
MaxGhenis merged 53 commits into
mainfrom
uk/ab2026-harvest
Nov 5, 2026
Merged

MaxGhenis merged 53 commits into
mainfrom
uk/ab2026-harvest

Conversation

@MaxGhenis

@MaxGhenis MaxGhenis commented Oct 29, 2026 •

Copy link
Copy Markdown
Contributor

In tranche 1 of #230, we add the Autumn Budget 2026 registry and harvest every producer's figures from primaries we fetch and hash. We model it on the AB2025 port (#136/#137). No code reads the harvest yet; we add the ingest in tranche 2.

Changes

  • Registry (data/uk/ab2026_measures.json, 104 measures), which pipeline/build_ab2026_registry.py writes.
    • Contents: every HM Treasury Table 6.1 line once, Table 6.2 line 4 (Sovereign AI, which has no Table 6.1 line), and five packages (the three totals, the cost-of-living trio, and the seven dashboard measures).
    • Keys: pipeline/ab2026_measure_keys.py writes out each line's key, so we keep the key when we reword its title.
    • Engine pin: policyengine-uk 2.151.0, the version the dashboard recorded in its Budget-day receipts. certified_pairs.json records only the 6.11.0 / 2.132.6 pair. In Certify the policyengine.py 6.11.1 replay results #242, we certified policyengine.py 6.11.1's replay results with 2.151.0, and we add that pair's record in tranche 3.
    • Triage: 11 expressible, 4 partial, 89 not expressible. We split the 89 into 13 policyengine-uk development items and 76 measures outside the engine's scope. For each gap, we record a name search of the pinned engine.
      • 7 that the dashboard executes: Energy Payment, electricity VAT, fuel duty, Carer's Allowance taper, trading allowance, Pension Tax Allowance, property income rates. For each, we name in pe_modifier the module, the function, the variables it replaces and reads, and its receipt (path, file sha256, reform sha256, run identity).
      • 2 parameter deltas on current law: savings rates (off the release) and the BADR lifetime limit.
      • Bus fare cap: partial. The engine already models the £2 cap from January 2027, and nothing records the cap 2027 would have had without the measure. We replace data/uk/bus_fares_gap.json's verdict at 2.89.2 with this one.
    • We record in pre_event_outcomes what became of each of the 14 pre-event keys (Autumn Budget 2026 (28 October): scoreability registry so counterparts can be computed on the day #96).
  • uk_hmt_tables (768 rows, which pipeline/stage_ab2026_hmt_tables.py stages from the two pinned workbooks).
    • The stager gives Table 6.1's lines and Table 6.4's recostings revenue_change, and Table 6.1's three totals decisions_effect_on_borrowing (README rule 10). It keys each Table 6.4 recosting to the earlier event's measure.
    • It gives Table 6.2's nine lines and its total financial_transaction_cash_impact, a proposed metric. In line 6, HM Treasury reprioritises financial transactions budgets to fund line 5's scheme, and it sums lines 1 to 9 into the total. HM Treasury lists neither line 6 nor the total in Table 6.1, so the stager gives each a proposed key.
  • Validator: accepts the executed_by_dashboard and package_of_registry_measures constructions, and --resolve checks replaced and read variables against the engine. We build packages with package_of, as in AB2025. The validator still accepts AB2025 and the pre-event registry.
  • Signs: scorecard_db/signs.py reads every sign convention on the harvest's borrowing rows. It reads a total of rows under two conventions, or an average of absolute values, as unknown. On revenue rows, it now reads the OBR workbook's legend, RF Figure 10's "reduces borrowing" and IFS's "positive = increase in net tax".
    • It reads a convention as unknown where the only phrase that decides it falls in a negative leg: "negative = …" or "negative implies …", up to the next semicolon. It also declares one unknown where it reads two overlapping phrases opposite ways (review A H-4).
    • It reads the conventions of 4,956 harvest rows as one direction or the other, where main's signs.py reads 2,290: 2,538 more rows on revenue_change, 128 more on decisions_effect_on_borrowing, and none fewer.
    • Both versions of signs.py classify all 29,894 exported rows from main's data/populations.json and data/populations-pmd/ identically.
  • Tooling: pipeline/harvest_counts.py writes and checks COUNTS.json.
  • Contract checker: pipeline/ab2026_harvest_contract.py defines the row contract as one function. The tests call it, and a lane runs it on its own family before pinning that family in COUNTS.json. It reads value_raw only if the whole string matches one of the forms the README lists, and reports any other string, so we narrow value_raw to the number. tests/test_ab2026_harvest_contract.py breaks each rule on an otherwise valid row. A test runs every number from 0 to 999 in every spelling our number generator produces. Hypothesis checks the period, quote-length, number, ratio, percentage, money-cell, rule 4 and field-type rules, and that changing or deleting fields never makes the checker raise. We tested the checker with 574 mutations of its reader and rule 4, one at a time, and a test failed for each.
  • Proposed keys: pipeline/ab2026_proposed_keys.py writes PROPOSED_KEYS.md: every proposed key we assign to a staged row, with its rows, families and producers. It finds 86 keys (15 measures, 58 packages and 13 options) on 973 rows. In tranche 2, we register or fold each one.
  • Docs: the harvest README and a seed inventory from the 28–29 October reaction monitor. In the README, we keep the AB2025 rules, with the AB2026 exceptions, list what the contract checker adds to the AB2025 row contract, and add three rules:
    • rule 8: stage estimates, not design. Put rates, thresholds, payment amounts, dates, durations, allocations, financing remits and contingent liabilities in reform_hint or conditions, never in rows. We stage HM Treasury's Tables 6.1, 6.2 and 6.4 and the OBR's policy workbook whole. Elsewhere, we treat an amount a scheme will lend or spend as design, and an effect on borrowing, debt or the financing requirement as an estimate;
    • rule 9: key a part of one measure to that measure, set measure_component, and never sum the part with the measure's total;
    • rule 10: give a producer's stated effect of the Budget's decisions on PSNB decisions_effect_on_borrowing, since models.py defines that metric as PSNB and keeps it apart from revenue_change, a receipts line. Also give it to the effect's identified parts and splits, and to any figure a producer derives by restating or recombining them. Keep revenue_change for receipts totals and tax-change estimates unless their evidence establishes PSNB scope, and for single measures, groups of measures and a total's lines by head unless their producer states the effect on borrowing. Keep a row's metric when a producer restates the row's Table 3.1 group or single measure as an effect on borrowing, and keep revenue_change for a total that includes the receipts effect of an option the Government didn't adopt. Give rows on the current budget deficit, or for an outturn the primary deficit, the proposed metric fiscal_balance_change, and name that deficit in their sign convention.

Invariants (tested)

  1. Every Table 6.1 line 1–98 appears once, under key(line). The Table 6.1 rows we stage share that key.
  2. Every HM Treasury cell we stage is a multiple of £5m, and none reaches £10bn. We stage one row for each line, or total, and year in each table.
    • Per head, Table 6.1's lines sum to its printed totals within £25m, and the tax and spending lines together sum to the grand total within £25m. We observe gaps of up to £15m.
    • The tax and spending totals sum to the grand total within £7.5m, and Table 6.2's lines sum to its total within £25m.
  3. Each package consists of its legs and nothing else.
    • The tax and spend totals partition the grand total, and every leg's head matches its package.
    • We mark a package expressible only if every leg is; otherwise we list its expressible legs.
  4. COUNTS.json equals harvest_counts.counts(...): rows, primaries, sources, claims sha256, totals.
  5. Every row's measure_key takes one of three forms:
    • an AB2026 key;
    • for a recosting, a key from the event that recosting_of names;
    • null, with a well-formed proposed key we haven't registered or, on 913 rows with no single AB2026 measure's or package's figure (fiscal headroom, real income growth and TPA's comparisons with earlier Budgets, for example), with none.
  6. Quotes from producers outside the Open Government Licence stay under 15 words, and every quote contains its value_raw.
    • The checker reads value_raw whole or reports it: a string that matches none of the README's forms gets a problem at any value. It reads a bracket only if a percentage restates a ratio or worded fraction to the nearest percentage point. It reads a range only if the range runs upward, or neither number is positive, and doesn't join a percentage to a multiplier or a fraction to a percentage. It reports a year from 2010 to 2099 after a bare "in".
    • value matches value_raw's number within four units in the last place, at the scale the number's own word or normalization gives, failing that 1. The checker allows 0.1% rounding for ratios, worded fractions ("1 in 4", "a third") and whole numbers with a fraction ("two and a half"). It allows the same for a fraction of a multiplier or percentage ("half a million", "half per cent"), and none for digits or number words. At any value, it reports the unread operations the README lists in normalization, such as "x (1e6)", "times 1,000" and "£m×1e6".
    • A test checks every number from 0 to 999 in every spelling our number generator produces. Hypothesis checks that any of the 27 examples in READABLE fails at any value with more around it or a word inside it. It checks that a qualifier, case and spacing change no reading, that a minus flips a reading, and that a multiplier scales only its own number under any normalization.
    • Hypothesis also checks that a ratio in digits or words, signed or not, passes only at its signed share, and that a ratio whose first or second number runs on fails at any value. It also checks that a printed percentage's share matches within floating-point error and fails after a 1e-12 shift, and that a money cell passes only at its exact value.
    • The checker limits non-restated, non-categorical printings of one claim to two, and requires a pair to agree within 0.5%.
    • We name the part and its measure on every measure_component row (rule 9).
    • We state in each family's NOTES.md the scope of every proposed key on its rows.
  7. Each receipt matches the dashboard at 1eca3182: file sha256, reform sha256 and binding, generated_at, certified model version, dataset revision and sha256, and the policyengine wrapper's version.
    • The test runs where a dashboard checkout exists. CI has none, so it skips there.
    • It passes locally for all 8 receipts at 72cdfdc.
  8. pipeline/build_ab2026_registry.py regenerates the registry and REGISTRY_KEYS.md byte for byte. The test runs where the 2.151.0 venv exists. It passed at 6e06907 with policyengine-core 3.34.5 (comment 6639810412), and we haven't changed the registry or its builder since.
  9. Tests check rule 10 across the harvest.
    • On every key we reserve for a stated effect of the Budget's decisions on PSNB, or for one of its splits, the tests require decisions_effect_on_borrowing on every row. The Table 6.1 test also checks each row's metric: the borrowing metric on the three totals, revenue_change on the lines.
    • We key the OBR workbook's Table 1.1 lines by head, and its total of receipts measures, to the total effect of Government decisions. We keep revenue_change on those 288 lines and on the receipts total, and the tests require the borrowing metric, or a proposed fiscal_balance_change, on every other row on that key.
    • We keep revenue_change on receipts totals and tax-change estimates: the OBR workbook's total of receipts measures (6 rows) and IFS's estimate of the Budget's tax change at the forecast horizon (1 row). We also keep it on RF's 11 totals with a continued fuel duty freeze.
    • On any other key, we give the borrowing metric only to seven OBR rows (paragraphs 3.20, 3.21 and 3.43) and IFS's two new-giveaways rows.
    • We propose fiscal_balance_change on 22 rows (red book 16, OBR 6). In each one's sign convention, we name the deficit its producer measures and not the other: the current budget deficit on 21 rows, and the primary deficit on the red book's primary balance outturn. The test reads that deficit from the producer's label, or for Chart 1.6 from the stability rule, never from the sign, and fails if we swap the two names in any row.
    • signs.py reads each borrowing row's sign convention as one direction, or declares it unknown.
  10. Wherever signs.py reads a convention for both revenue and borrowing, it reads it opposite ways. A Hypothesis property samples the phrases signs.py reads as one direction, for either metric, and places them in random filler, with or without a negative leg after them. A second test runs every convention in the feed and the harvests, and signs.py reads at least 30 of them for both metrics.
    • We pin in a fixture how signs.py reads each of the harvest's 76 (metric, convention) pairs, as a reviewer checked them. The test fails on a new phrase, or on an edit that changes a reading, until someone reviews it.

Review fixes

The IFS addendum review, at d5e323d

GPT-6.1 Sol reviewed the IFS addendum at d5e323d on Subfleet, re-reading the local primaries, and requested changes in four findings. We answered them in f4c7c94 and 8b7b23d. In a delta review of both commits, GPT-6.1 Sol requested three more changes, and we made them in dfebe9c, which it approved in a second delta review. In these commits, we changed no value, year or sign a producer printed.

  • f4c7c94. IFS's £8 billion and £4 billion of new giveaways match the OBR's Table 3.1 only after rounding, and IFS doesn't say, in its initial response, that they cover every Government decision since March. We moved them from that total's key to ab2026__package_ifs_new_giveaways, at medium confidence. IFS names no source for the 25% and 40% reserve shares in either primary where they appear, so we marked the restated attribution on their four rows as inferred, also at medium confidence. We also corrected two NOTES sentences, on IFS's reference line at 0.9% of GDP and on our failed fetches of the seven data-item pages and the briefing event page IFS names in the addendum.
  • 8b7b23d. GPT-6.1 Sol asked us to move the OBR's Table 3.1 totals and its two memo lines on direct effects (42 rows) to decisions_effect_on_borrowing. It also asked us to give the OBR's current budget deficit memo (6 rows) the proposed metric fiscal_balance_change, and to name that deficit in the sign of the red book's matching memo (6 rows). We did all three, and wrote README rule 10 to put every producer's totals of the Budget's decisions on decisions_effect_on_borrowing. In that commit, we moved 172 rows to the borrowing metric: OBR 87, HMT tables 18, TPA 40, RF 25 and IFS 2.
  • dfebe9c. In its delta review, GPT-6.1 Sol accepted 165 of those moves and found two kinds of row whose evidence doesn't establish PSNB scope. The OBR sums its receipts heads, from air passenger duty to visa fees, into its policy workbook's total of receipts measures, so we returned those 6 rows to revenue_change, the receipts line models.py keeps apart from PSNB. IFS doesn't tie its 0.2% of GDP, its own estimate of the Budget's tax change at the forecast horizon, to PSNB, so we returned that row too. We rewrote rule 10 to give the borrowing metric to a producer's stated effect on PSNB, its identified parts and any figure a producer derives by restating or recombining it. We keep revenue_change on receipts totals and tax-change estimates unless their evidence establishes PSNB scope. GPT-6.1 Sol also found that we gave the wrong count in the Signs bullet. At 8b7b23d, signs.py read the conventions of 2,520 more harvest rows on revenue_change than f4c7c94's version did, not 2,576. At dfebe9c, it read 2,527 more revenue rows, after we returned the receipts total's 6 rows and IFS's 0.2% to revenue_change. The stagers then moved 165 rows: OBR 81, HMT tables 18, TPA 40, RF 25 and IFS 1.
  • 4f424b9. In the second delta review, GPT-6.1 Sol found that we had written, in a test's docstring and in invariant 9, that the OBR keys its Table 1.1 lines to the total effect of Government decisions. Our stager assigns those keys, so we corrected both sentences.

We applied rule 10 past the IFS review's table in three ways:

  • We moved 59 more rows that restate or split effects GPT-6.1 Sol identified in the IFS review, keeping their keys. We moved the OBR workbook's 36 Table 1.1 effects, the OBR's three figures in paragraphs 3.3 to 3.5 and Table 6.1's grand total (6 rows). We also moved RF's 14 figures for the OBR's direct effect of tax measures and for the total cost of the Budget's decisions. In tranche 2, we can then compare producers' figures for each effect under one metric.
  • HM Treasury sets out, in Table 6.1, the cost or yield of the decisions that directly affect PSNB, so we moved its tax and spending totals (12 rows) with its grand total. We also moved TPA's tax totals (40) and IFS's £7 billion net tax rise, where TPA and IFS restate them. RF adds the cost of a continued fuel duty freeze, from the OBR's constant-duty-rate alternative, to the OBR's direct effect of tax measures. We moved that total too (11 rows), until the cloud session returned it after reviews A and B.
  • We named the current budget deficit in the sign of the red book's other Table 1.3 memo, on underlying forecast differences (5 rows).

Reviews A and B, at 4f424b9

On 4 November, two reviewers fetched their families' primaries from the manifest URLs; review A couldn't refetch two superseded HM Treasury copies, since HM Treasury now serves the current versions at those URLs. Review A covered the registry, the HM Treasury families and the harness (comment 6635870934), and review B the OBR, RF and think-tank families (comments 6635496734, 6635498758 and 6635499259). A cloud session applied their findings in 6a1c06f to 15c0773 and listed each finding with its fix in comment 6639062906. It couldn't fetch primaries, so we finished its three open items from the pinned copies: red book rows 194, 365 and 368 (93d2407) and Table 6.4's footnote 2 (3b1abda). We also pinned each receipt's policyengine version (6e06907).

  • Rows: the cloud session took the harvest from 5,948 rows to 5,952. It added 3 OBR rows (paragraph 3.20's rise in the financing requirement, back on cash_requirement_change, and Box 3.1's two counts), 2 CenTax rows from a second primary and Table 6.2's total (6 rows). It dropped 7 red book rows: four printings beyond the two the checker allows for one claim, and three figures outside the harvest's scope (its NOTES categories E, F and H).
  • Later primaries: the OBR corrected its outlook PDF on 30 October, and HM Treasury republished its Pension Tax Allowance explainer on 29 October. Under rule 7, the cloud session added a manifest line, which no row cites, for each later version, and for RF's landing page as it stood on 4 November.
  • Metrics: the cloud session returned RF's 11 totals with a continued fuel duty freeze to revenue_change. RF builds them by adding the receipts effect of an option the Government didn't adopt, and the OBR states that option's effect on borrowing for 2029-30 only. So the stagers now move 154 rows to the borrowing metric: OBR 81, HMT tables 18, TPA 40, RF 14 and IFS 1. The session also moved eight rows' registered metrics from proposed_metric to metric (IFS 4, misc 3, Trussell 1) and gave RF rows 88 and 211 gross_earnings_change. The balance-row test now selects every row with a proposed fiscal_balance_change: the 17 memo rows, the red book's three Chart 1.6 bridge steps and two Box 1.C outturns.
  • Checks: the cloud session added review A's H-1 to H-3 and H-6 to the contract checker, and tests for H-7's manifest branches. It added H-4's two rules to signs.py, pinned the harvest's 76 sign readings in H-5's fixture, and sorted the registry builder's name search (R1), so --check now regenerates the registry byte for byte.

Delta reviews D1 to D3, at 6e06907

GPT-6.1 Sol reviewed 4f424b9..6e06907 and the staging branch in three parts. It requested changes in D1, on the harness (5 findings), approved D2, on the OBR and HM Treasury families, with one fix, and approved D3, on the think-tank families, with three. In D3, it couldn't spot-check every primary, since we hadn't given it local copies of RF's v2 report and workbook, CenTax's PDFs, TPA's primaries or the Welsh and Scottish pages; review B had fetched them all. We answered the three reviews in ac4d14a and f27b927, and on the staging branch in e928465 and 83d5a32.

  • D1-1. The checker let "1 in 4" pass at 0.01, 1 and 4, and let a £m cell drift within the 0.1% we meant for rounded shares. In f27b927, we made the checker accept a ratio only at its share and read a year after "in" as a year, and capped every other printed number's drift at one part in a billion. The final review found gaps in each (F1 and F2).
  • D1-2. The balance test accepted either deficit's name in any row's sign. It now reads the deficit from the row's label and requires the sign to name that deficit and not the other.
  • D1-3. The checker raised an exception on four kinds of malformed input. In f27b927, we made it report each as a problem, and the final review found two more (F3).
  • D1-4. The cloud session counted a gain of 2,674 (comment 6639062906), including conventions signs.py declares unknown. We count only those it reads as one direction: 2,666.
  • D1's other notes. D1 also found that no test pinned one row per table cell, and that every test still passed if we removed any one of four checker arms. It found too that rule 4's check rejected a Figure 1.A or 1.B locator even when it names the paragraph where HM Treasury prints the number. We added the test and the cases, and narrowed the check. Then we tested 14 mutations of the checker one at a time, and a test failed for each. We ran them locally with a stand-in for Hypothesis that draws 200 seeded examples; CI runs the library.
  • D2-1. When the stager applied review A's C-1, it also gave costings row 584, a count of about 4.9 million properties, a receipts component. It now leaves that row as 4f424b9 had it.
  • D1-5, D3-1 to D3-3. We corrected red book row references in the README, in misc row 8's note and in the misc and devolved NOTES. We also said in the red book NOTES which numbering we use in its review section, and dropped a stale sentence from TPA's NOTES.

The final review, at f27b927

GPT-6.1 Sol reviewed f27b927 on Subfleet and requested changes in 12 findings. We answered F1 to F4, on the contract checker, in fe58796 without changing a staged row. For F5 to F12, we corrected two counts of signs.py readings, two OBR row references, the Plan's readiness box and our account of fetches and test procedures. We also gave "we" or a program the verbs we had given rows and totals.

  • F1. The checker read "1 IN 4" as two numbers, so it accepted 1 and rejected 0.25. It also dropped a ratio's minus sign, and with the normalization "N in M -> N/M" it took the 2000 in "1 in 2000" for a year. It now reads a ratio in any case, hyphenated ("1-in-4") or in words ("one in ten"), gives the share the ratio's sign, and reads a year after "in" as a denominator when we write the ratio in normalization. It now also checks ten rows where a producer prints a ratio in words: six from CenTax and four from Resolution Foundation.
  • F2. With a relative tolerance of one part in a billion, the checker let a £2,625m cell drift by £2. The checker now allows four units in the last place, and each staged value that matches a printed number differs from it by at most one.
  • F3. The pair check raised an exception on a row without value, or with an integer too large for a float. The checker now reports both, and it also reports a line it can't parse as JSON, a file it can't read and a field with nesting too deep to print.
  • F4. Under rule 4, the checker accepted a bar reading with a paragraph after the bar in its locator ("Figure 1.A, decile 3 bar; see paragraph 1.7"), or with a negated data label ("no data labels"). It uses the first reference, up to the first comma or semicolon, to identify where HM Treasury prints the number. It rejects a negated printing reference and checks the whole locator for an affirmative bar, chart or graph reading.

We also rewrote the harvest README in active voice and cut its review history. On f27b927's checker, 37 of fe58796's 123 contract tests fail.

Our own pass, at fe58796

Before the delta review, we probed the checker again and found gaps in two checks, which we closed in 1bf4507 without changing a staged row.

  • Rule 4. The checker missed Figure 1.A or 1.B in other spellings, so it accepted "Fig. 1.A, decile 3 bar" and "Figures 1.A and 1.B, decile 3 bars". It also missed negations such as "missing", so it accepted "Figure 1.A data label missing". It now reads "Fig. 1A" and "Figures 1.A and 1.B", and treats "never", "missing", "absent", "lacks", "lacking" and "n't" as negations.
  • Manifest. The checker accepted a list or a number as a manifest line's url or title. It now requires the url, title and retrieved date as text, and names a line by its place when its url isn't text.

On fe58796's checker, 20 of the 152 contract tests fail. In 72cdfdc, we split a 44-word sentence in the README's rule 4 bullet.

Our own pass, at 72cdfdc

While the delta review of 72cdfdc ran, we widened what the checker reads in 615d6a8. It now reads "1 in 2 million" as 5e-7, "one in every ten" and "3 out of 10" as shares, compound number words in a ratio ("twenty-five", "a hundred and five"), worded fractions ("three-fifths") and number words alone ("four"). Across the 5,952 staged rows, it reads 5,943 as before and compares 9 rows it skipped before, and all 9 pass. On 72cdfdc's checker, 47 of the 208 contract tests fail.

The delta review of 72cdfdc, at 92f364c

GPT-6.1 Sol reviewed 72cdfdc and asked for five changes, which we answer in 92f364c. The checker reads every staged row as 615d6a8's checker did, and finds 0 problems in the 12 families.

  • R1. The ratio reader passed "twenty-one in a hundred" at 0.01, since it started again at "one", and passed "one in twenty-five" at any value, since it found nothing to compare. It read "1 in -4" as +0.25, and missed the generic normalization in capitals ("N IN M -> N/M"). The checker now reads each number in a ratio as one run of tokens, takes a minus on either number and reads the generic normalization in any case. It reports a ratio, or a number in words, that it can't read as a problem. We had drawn the worded property's number words from the checker's own tables, so we now write them out in the tests.
  • R2. Under rule 4, the checker passed "Figure 1.A data label unavailable, decile 3" and failed "paragraph 1.7; not a bar reading from Figure 1.A". It now cancels a bar-reading phrase only when a negation comes right before it. In the first reference, it accepts only a paragraph, table or box by number, or a data label. Sol also asked us to pass "Figure 1.A data label (not missing), decile 3". We fail it instead: the checker treats any other wording as a bar reading, and we stage no locator in that form.
  • R3. No test pinned the match at four units in the last place, and ULPS = 64 passed every test. A test now builds values with math.nextafter, for both signs: one to four units pass, and five fail.
  • R4. The malformed-row property accepted an empty problem list, so source_model: [] passed. type_problems now reports each field of a type the row contract doesn't allow, before any check reads the field, and a new property checks it for every text field, period, each condition and each publication field. We renamed the old property test_changing_or_deleting_fields_never_raises and gave it Sol's docstring, extended to name the rows it pairs.
  • R5. We reworded README rule 1, the rule 4 bullets and two lines of this description.

On 615d6a8's checker, 168 of the 401 contract tests fail, and the tests kill all 41 mutants.

The delta reviews of 92f364c, at a9f81e8

GPT-6.1 Sol and a cloud session each reviewed 92f364c, and both requested changes. The checker scanned value_raw for numbers and accepted any matching reading. It passed "1 in .5" at 1, "4 - 5 in 10" at 0.5 and "two and a half" at 0.5. It also passed "half a million" at 0.5, "four million; five thousand" at 5,000,000 and "2 million in 5 years" at 400,000.

We replaced the scanner instead of patching it again, since we staged few forms. In value_raw, we staged digits on 5,900 of the 5,952 rows, a ratio or worded fraction on 31, number words on 7, a range on 6, a bracket on 1 and a rating on 7. The checker now reads value_raw only if the whole string matches a form the README lists, and reports any other string, so no string passes on a number inside it. We removed the token scanner and the year logic, and every staged row passes unchanged. We also ran both checkers on each row at seven values: as staged, ×1,000, /1,000, negated, ×1.01, ×100 and /100. They give the same verdict on all but 6 of the 41,664 cells, where "four" and "one" no longer pass at ×100, ×1,000 or /100, since number words now take normalization's factor or 1.

  • D1 to D4, A1 to A5. The checker now reads the strings in these findings whole, or reports them at any value. It reads ".5 in 4" as 0.125, "1e3 in 4" as 250, "-four" as −4, "two and a half" as 2.5, "half a million" as 5e5 and "one out of every five" as 0.2. A5 asked us to pass ".5 in 10" at 0.5, and the checker reads the whole string as a ratio, so it passes at 0.05. It reports "one in zero", "4 - 5 in 10", "1 in 4 to 5", "twenty-one fifths", "one thousandth", "four million; five thousand" and "2 million in 5 years" at any value.
  • D5, A6. Under rule 4, the checker now looks for a bar-reading phrase with hyphens and dashes as spaces and runs of spaces collapsed, and catches "bars read" and "estimated from the chart". It accepts "paragraph 1.7 and Figure 1.A", "paragraphs 1.7–1.8; Figure 1.A" and "Figure 1.A: data label for decile 3". It rejects "Figure 1.A, data label decile 3", since it reads the first reference up to the comma, and its problem message now says how the locator should start.
  • Brackets and ranges. We also tightened two forms the closed list would otherwise let through. The checker now reads a bracket only if a percentage restates a ratio or worded fraction to the nearest percentage point, and reports "£5bn (2027)" at any value. It reads a range only if the range runs upward in value or in size and doesn't join a percentage to a multiplier, and reports "2025-26" and "4% to 2 million" at any value.
  • D6, B1 to B5, C1 to C3. We reworded README rule 1, the contract bullets and parts of this description, and dropped B1's "£5 in 2027" example with the year logic.

On 92f364c's checker, 203 of the 518 contract tests fail, and the tests kill all 194 mutants.

The delta reviews of a9f81e8, at 3c65fd7

GPT-6.1 Sol and a cloud session each reviewed a9f81e8, and both requested changes. In 9fd73b1 and 3c65fd7, we tightened how the checker reads a number's scale, ranges and years. We also made it mark a bar reading on any bar, chart or graph in a locator. We changed no staged row, and the checker found 0 problems in the 12 families.

  • Scales (N1, M3). The checker let digits with no word of their own match at 1, 10³, 10⁶, 10⁹, 0.01 or 100, so "£5" passed at 5 billion and "5" at 500. It now gives them normalization's factor, failing that 1. All 737 staged rows with digits, no word of their own and no factor in normalization match at 1.
  • Years and ranges (M1, A-L2). The checker read "5 in 2027" as 5 / 2027, and passed "2 to -3" at 2 and "a third to 40%" at 0.4. At any value, it now reports a year from 2010 to 2099 after a bare "in", a range whose first number exceeds its second unless neither is positive, and a fraction joined to a percentage.
  • Rule 4 (N2, N3, N6, M2). The reviewers found more wordings our list of phrases missed, including "read directly off the chart". They also found "read off of the chart in Figure 1.A" and "taken from the chart in Figure 1.A". The checker now marks a bar reading wherever we mention a bar, chart or graph in the locator, or write "off" or "from" before "figure" or "figures" there. It still ignores a phrase right after one negation, and now fails "not without a bar reading", which has two. It also accepts any dash in a printed place, as in "paragraphs 1.7—1.8; Figure 1.A" and "Figure 1–A data label". Under the new rule, it marks none of the 56 locators on HM Treasury's Impact on households rows as a bar reading.
  • Fractions (N4). The checker now reads a fraction with a direct percentage word ("half per cent") or multiplier ("half million"). It also reads a whole number in words up to the hundreds with a fraction ("one hundred and a half"). On a unit_concept: percent row, it compares a fraction of a percentage at the same values.
  • Dashes for nil (N5, A-L3). We couldn't stage nil when a producer printed it as a dash, since the checker read no dash. It now reads a lone dash as zero where normalization mentions nil, and reports it elsewhere, since a producer may print one for "not available".
  • Tests (N7, A-L1). We added probes at ±0.09% and ±0.11% for both signs and both kinds that round, conflicting normalizations for worded multipliers, wrong shares in the exhaustive ratio test, and rule 4 cases for each arm a mutant reached. We also grew READABLE to 27 examples.
  • Forms we left out. The checker still reports "65p", "£ 5", "½", "1/4", "page 8, paragraph 1.7; Figure 1.A", "Table A.1; Figure 1.A", "Figure 1.A data labels for deciles 1 to 3" and "paragraph 1.7 & Figure 1.A". We stage none of them, and we would narrow the first two to their digits.
  • N8, N9, B-L1 to B-L4, C-L1. We reworded the README's contract bullets and corrected this description's claims about the tests, A5 and rule 4.

We also ran both checkers on each row at seven values. They agree on all but 2,196 of the 41,664 cells. In each of those, a9f81e8's checker passed one of 732 rows at ×1,000, ×100 or /100, where 3c65fd7's checker reports it. We store 0 in the other 5 of the 737 rows, and every scale leaves it at 0. On a9f81e8's checker, 93 of the 725 contract tests fail, and the tests kill all 303 mutants.

The delta reviews of 3c65fd7, at 8206882

GPT-6.1 Sol and a cloud session each reviewed 3c65fd7, and both requested changes. In 8206882, we made the checker read each factor in normalization whole and require a complete clause declaring a dash for nil. We also widened rule 4 to abbreviated figures and rarer wordings. We changed no staged row, and the checker finds 0 problems in the 12 families.

  • Rule 4 (A1, A2, R1, R8). The checker missed abbreviated figure citations ("read off Fig. 1.A", "taken from Figs. 1.A and 1.B"). It also missed "on", "in" or "using" before a figure ("estimated using Figure 1.A"). The reviewers also found "estimated from the plot", "visual estimate of Figure 1.A", "Figure 1.A read by eye", "column 3 of Figure 1.A" and "Figure 1(a), decile 3". It now marks all of them and leaves "Table 2.1, column 3" alone. It also counts a phrase after two negations with words between them ("not possible without reading off the chart"). It still marks none of the 56 locators on HM Treasury's Impact on households rows as a bar reading.
  • Factors (R2). The checker read only the start of a factor, so "4" under "x 10^6" passed at 40 and failed at 4,000,000. Under "x 1,000,000 / 100", it passed 4,000,000 and 0.04 and failed 40,000, and under "(x 1e6)" it passed 4 and failed 4,000,000. It now reads each factor whole, multiplies operations in a row and reads a factor in brackets. It reports a factor it can't read ("x 1 million", "x 5^2") at any value.
  • Whole numbers with a fraction (A3, R4). The checker gave this form no factor, so "two and a half" under "x 1,000,000" passed at 2.5 and failed at 2,500,000. It now uses normalization's factor, failing that 1, and it reads a short multiplier ("2 and a half bn").
  • Nil notes (A4, R3). The checker passed a dash at zero under any note with the word "nil", including "dash means missing, not nil" and "nil-rate band". It now requires one whole clause of normalization to declare the dash ("dash for nil", "the table prints nil as a dash"). It also reads the figure dash (U+2012) and the horizontal bar (U+2015).
  • Tests (A5, R7). We added "0 to -3", rounding probes at ±0.105%, the denominator in the ratio test, each factor's spelling at its correct and wrong values, negations in capitals, and rule 4 rows for each arm a mutant reached.
  • Forms we left out. The checker reports a ratio whose denominator falls from 2010 to 2099 ("1 in 2050"), since it reads the denominator as a year. It also marks some printed places as bar readings, such as "read off the Figure 1.A data label" and "calculated from the figures in Table 2.C". We stage none of these, and we would reword the locators.
  • B1 to B4, C1 to C6, R5, R6, R9. We reworded the README's bullets on percentage words, ratios, dashes, scale and rule 4. We also corrected this description's section on the reviews of a9f81e8, and split its long rounding sentence.

We also ran both checkers on each row at seven values, and they agree on all 41,664 cells. On 3c65fd7's checker, 85 of the 857 contract tests fail, and the tests kill all 434 mutants.

The delta reviews of 8206882, at 4707e91

GPT-6.1 Sol and a cloud session each reviewed 8206882, and both requested changes. In 4707e91, we made the checker report more unsupported factor forms. We also widened rule 4 to punctuation before figures and more figure citations. We changed no staged row, and the checker finds 0 problems in the 12 families.

  • Factors (Sol F1). The checker skipped an operation on a bracket, sign or number words, so "4" under "x (1e6)" or "x +1e6" passed at 4. Under "x 1e6 / (100)" it passed 4,000,000, and it read "x 1 hundred" as 1. At any value, it now reports "x (1e6)", "x +1e6", "x £100", "x one million", "times 1,000" and "divided by 100".
  • Rule 4 punctuation (Sol F2). The checker missed a reading with a colon, bracket or quote before the figure, as in "estimated using (Figure 1.A)", "read from [Figure 1.A]" and "read off: Fig. 1.A". It now looks past them, and it still ignores the phrase right after one negation ("not read from [Figure 1.A]").
  • Rule 4 forms (cloud F4, F6, F14). The checker marked "Figure 1(a) data label" and "Figures 1(a) and 1(b) data labels" as bar readings, missed "read off fig1a" and "read off figure1.A", and counted two negations across a full stop. It now accepts those data labels, marks both readings and ends a clause at a full stop. It still marks none of the 56 locators on HM Treasury's Impact on households rows as a bar reading.
  • Tests (Sol F3). We added tests for each short multiplier on a whole number with a fraction and for factors without a space after the operator. We also tested unsupported operands across all five operators, alongside ordinary notes that state no factor.
  • Forms we left out (cloud F1, F3, F5). The checker doesn't mark "paragraph 1.7; Figure 1.A estimate", "read off the y-axis" or "read off the image". It doesn't read a dash under "dash: nil" or "dash-for-nil", and it reads "page 3 / 4" in normalization as a factor. We stage none of these, and we would reword them.
  • Cloud F2, F7 to F13; Sol F4, F5. We corrected the README's bullets on whole numbers with a fraction, ratios, factors, nil notes and rule 4. We also corrected invariant 6, and split long sentences in the README and in this description's sections on earlier reviews.

We also ran both checkers on each row at seven values, and they agree on all 41,664 cells. On 8206882's checker, 264 of the 1,169 contract tests fail, and the tests kill all 502 mutants.

The delta reviews of 4707e91, at 2f60cdb

GPT-6.1 Sol and a cloud session each reviewed 4707e91, and both requested changes. In 2f60cdb, we made the checker report operations in more wordings and operators with no space before them. For rule 4, we also made the checker look past nested punctuation, possessives and a bracketed word before a figure. We changed no staged row, and the checker finds 0 problems in the 12 families.

  • Factors (cloud M1, Sol F1). The checker read no factor from "multiply by 1,000,000", "divide by 100", "scaled by 1e6", "x (one million)", "x '1e6'" or "£m×1e6", so "4" under each passed at 4. Under "x 1e6 / (one hundred)" it passed 4,000,000. It now reports all of these at any value.
  • Rule 4 (Sol F2, cloud L1). The checker missed read from ("Figure 1.A"), "read off ([Figure 1.A])", "taken from HMT's Figure 1.A" and "read from (HMT) Figure 1.A". It now marks them, and it still ignores the phrase right after one negation.
  • Tests (Sol F3). We added rule 4 rows for brackets and quotes before an intervening word, each with its negation. We also added unread factors on a number with its own multiplier, a ratio, a worded fraction and "nil".
  • Forms we left out (cloud L2 to L4). The checker reads "x 10 to the 6" as 10 and "x 1 e6" as 1. It counts "paragraph 1.7; from [figures in Table 2.C]" as a bar reading, and it ends a clause at the full stop in "e.g." We stage none of these, and we would reword them.
  • Text (Sol F1, F4; cloud D1, D2, R1 to R4). We corrected this description's section on the reviews of 8206882, which said the checker reports any operation it can't read. We also corrected the README's instruction on slashes and products, split long sentences in its contract bullets, and reworded a repeated opener in the section on the reviews of 3c65fd7.

We also ran both checkers on each row at seven values, and they agree on all 41,664 cells. On 4707e91's checker, 506 of the 1,718 contract tests fail, and the tests kill all 555 mutants.

The delta reviews of 2f60cdb, at 05265c2

GPT-6.1 Sol approved 2f60cdb with four low findings, and a cloud session requested changes with low findings only. In 05265c2, we made the checker report seven more spellings of an operation and fractions as operands. For rule 4, we also made it look past more punctuation before a figure. We changed no staged row, and the checker finds 0 problems in the 12 families.

  • Factors (cloud A1, Sol F1). The checker read no factor from "£m→x 1e6", "£m→/100", "× by 1e6", "* by 1e6", "÷ by 100", "/ by 100" or "times by 1,000", so "4" under each passed at 4. It did the same with fractions as operands, in "times half", "/ half", "/ a quarter" and "scaled by half", and under "x 1e6 / half" it passed 4,000,000. It now reports all of these at any value.
  • Rule 4 (cloud A2, Sol F2). The checker missed read off the "Impact on households" Figure 1.A, read from "HMT" Figure 1.A and "read from ((HMT)) Figure 1.A". It now marks them, and it still ignores the phrase right after one negation.
  • Figures that mean numbers (cloud A3). The checker marks "from HMT's figures" as a bar reading, since it can't tell a figure in a document from a number. We say so in the README, and we would reword such a locator.
  • Tests (cloud C2, Sol F3). We added negated rows for read from ("the Figure 1.A") and read from ( the Figure 1.A, and rows for each new form. We also added "two and a half" and "half a million" under an unread factor.
  • Forms we left out (cloud A1, A2). The checker reads no factor from "divide it by 100", "scaled up by 1,000", "£m · 1e6" or "£m;times 1,000", so "4" under each passes at 4. It doesn't mark "read from HMT, Figure 1.A", "read from HMT/Figure 1.A", "read from {Figure 1.A}" or a figure five words after "from". We stage none of these, and we would reword them.
  • Text (cloud C1, D1, D2, R1 to R4; Sol F4). We corrected invariant 6 and this description's section on the reviews of 4707e91. In the README, we reworded the bullets on nil notes, slashes and rule 4, and specified that the checker reads the locator.

We also ran both checkers on each row at seven values, and they agree on all 41,664 cells. On 2f60cdb's checker, 359 of the 2,103 contract tests fail, and the tests kill all 574 mutants. We then merged main, with #251, in 4dd5281, so #247 can take #251 and still show only tranche 2.

The review of 4dd5281

GPT-6.1 Sol approved 4dd5281 with three low findings, and we merge at this head. The checker reads a numeric mixed fraction as a partial factor, so "4" under "x 2 and a half" passes at 8 (F1), and we test no locator with repeated closing square brackets (F2). We stage neither form, and we answer both in a follow-up PR.

Plan

  • Add the registry, HMT tables, README, seed inventory and tests.
  • Stage 12 producer families, 5,952 rows (450 restated), from hashed primaries on Subfleet build lanes, and pin their counts in COUNTS.json. We list each family's rows and primaries in sources/harvest-uk-ab2026-2026-10-29/README.md, with the four families that have no rows (uk_cpag, uk_smf, uk_niesr and uk_fai_sfc).
  • Answer the IFS review of d5e323d (f4c7c94 and 8b7b23d, under Review fixes).
  • Get a delta review of f4c7c94 and 8b7b23d, including rule 10's reach past the IFS review's table, and answer it in dfebe9c.
  • Get a delta review of dfebe9c, and answer it in 4f424b9.
  • Run review A (the registry, the HM Treasury families and the harness) and review B (the other families) at 4f424b9, fetching each primary from its producer, and apply their fixes.
  • Get delta reviews D1 to D3 of those fixes, and answer them in ac4d14a and f27b927.
  • Get a final review of f27b927, and answer it in fe58796.
  • Probe the checker again before the delta review, and close the gaps in 1bf4507.
  • Get a delta review of 72cdfdc, and answer it in 92f364c.
  • Get delta reviews of 92f364c from GPT-6.1 Sol and a cloud session, and answer them with the closed reader in a9f81e8.
  • Get delta reviews of a9f81e8 from GPT-6.1 Sol and a cloud session, and answer them in 9fd73b1 and 3c65fd7.
  • Get delta reviews of 3c65fd7 from GPT-6.1 Sol and a cloud session, and answer them in 8206882.
  • Get delta reviews of 8206882 from GPT-6.1 Sol and a cloud session, and answer them in 4707e91.
  • Get delta reviews of 4707e91 from GPT-6.1 Sol and a cloud session, and answer them in 2f60cdb.
  • Get delta reviews of 2f60cdb from GPT-6.1 Sol and a cloud session, and answer them in 05265c2.
  • Mark the PR ready once a review of 4dd5281 approves it.
  • In tranche 2, we register or fold each of the 86 keys in PROPOSED_KEYS.md, then ingest. Before the ingest, we also:
    • teach signs.py the conventions it can't yet read: 14 OBR prose rows on revenue_change and 12 rows on other metrics;
    • put IFS's shares of GDP (percent with a GDP denominator) on the OBR's and TPA's percent_of_gdp;
    • settle HM Treasury figures we stage twice, such as red book Table 6.3 rows 345–356 against uk_hmt_tables lines 45 and 46;
    • choose one metric for financial transactions' cash effect, where we propose financial_transaction_cash_impact for Table 6.2 and give OBR row 2684 the registered cash_requirement_change;
    • give the continued fuel duty freeze's £4.3bn one metric: the OBR states it as an effect on PSNB (row 2740), and RF as a receipts cost (rows 154 and 397–401);
    • choose one metric for the OBR workbook's total of receipts measures, which we stage on revenue_change from the OBR (rows 2190–2195) and on the borrowing metric, under the tax decisions key, where RF restates it (RF rows 385–390);
    • decide whether to stage the OBR estimates that review B raised and we left out (Tables 4.2–4.7, 5.2, 5.8 and 6.3, and paragraphs 4.30 and 7.20), and FAI's 29 October table of Scotland's Barnett consequentials;
    • sort the name search in build_ab2025_registry.py, as we did for AB2026 (R1), and decide whether to rename the one-row welfare-savings key (RB-8).
  • In tranche 3, we add PolicyEngine counterparts from the dashboard receipts and a 6.11.1 certified-pair record.

🤖 Generated with Claude Code

…ables 6.1, 6.2 and 6.4

Tranche 1 of #230, following the AB2025 port (#136).

- data/uk/ab2026_measures.json: 104 measures. Every Table 6.1 line once,
  under a key written out per line in pipeline/ab2026_measure_keys.py, plus
  Table 6.2 line 4 (Sovereign AI) and five packages. The engine pin is
  policyengine-uk 2.151.0, which the dashboard's Budget-day receipts record.
  The seven dashboard measures carry pe_modifier with their receipt hashes.
  Savings rates and the BADR limit are parameter deltas; the bus fare cap is
  partial. 89 gaps each carry a computed name search at the pin.
- pipeline/stage_ab2026_hmt_tables.py stages 762 rows from the two pinned
  workbooks. Table 6.4 recostings carry the earlier event's key.
- The validator accepts executed_by_dashboard and package_of_registry_measures
  constructions and resolves replaced and read variables against the engine.
- pipeline/harvest_counts.py writes and checks COUNTS.json.
- The harvest README, the seed inventory and the tests for the registry and
  the harvest contract.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@vercel

vercel Bot commented Oct 29, 2026 •

Copy link
Copy Markdown

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

Project Deployment Actions Updated
policyengine-scorecard Ready Ready Preview Nov 5, 2026 1:40pm UTC

Request Review

… its own family

pipeline/ab2026_harvest_contract.py holds the row contract the tests applied
inline, so a family being staged can be checked before COUNTS.json pins it.
Only HM Treasury and OBR quotes run whole (README rule 6).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…re keyed

Rule 8 keeps a measure's design (rates, thresholds, payment amounts,
dates, durations, allocations, financing remits, contingent liabilities)
out of the rows. Rule 9 keys a printed part of one Table 6.1 measure to
that measure, marked as a component that is never summed with the total.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… for tranche 2

The contract now also checks that a quote contains the value as printed,
that a measure_component row names its part and its measure (rule 9), and
that a family's NOTES.md gives the scope of every proposed key its rows
name. tests/test_ab2026_harvest_contract.py breaks each rule on an
otherwise valid row, and Hypothesis checks the FY-end-year and 15-word
quote rules for every input. pipeline/ab2026_proposed_keys.py writes
PROPOSED_KEYS.md, the one list tranche 2 registers or folds.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Table 6.1 line 2 costs the whole electricity VAT cut from 1 October 2026,
so a package of the energy payment, the VAT cut and the fuel duty cut is
exactly lines 1 to 3; the old example said otherwise. Rule 2 now names
the three kinds of proposed key (a measure with no Table 6.1 line, a
package named by its scope, a producer's option), and PROPOSED_KEYS.md
shows each key's kind.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@MaxGhenis

Copy link
Copy Markdown
Contributor Author

Picking up tranche 1 for the Budget coordinator: finishing the red book and IFS reworks, applying the five harvest reviews, and landing the families from uk/ab2026-harvest-staging.


Generated by Claude Code

@MaxGhenis

Copy link
Copy Markdown
Contributor Author

Progress at 13:03Z on 3 November: no families have landed yet, and nothing is unpushed.

What happened: after the pick-up comment at 03:09Z I set up (checkouts of uk/ab2026-harvest at 9dbb52d and uk/ab2026-harvest-staging at 7c97f3e, a Python 3.12 venv .venv-pe2151 with policyengine-uk 2.151.0) and read the harvest README, the staging README, COMMON.md, verify-COMMON.md and the contract. The session's worker process then restarted at about 03:15Z and sat idle until your check. No family file was edited, so there is nothing to commit on either branch; both heads are unchanged.

Ownership check: the only pushes in the six hours before pick-up were your two staging commits (7a23c33, 7c97f3e), which are the hand-off itself, so I proceeded.

Next, in the brief's order: uk_hmt_redbook (check the rework against its brief, contract, byte-for-byte re-run, 15 spot checks), then uk_ifs_ab2026 sections 2 to 4, then the reviewed families, pushing both branches after each.

Blocking: nothing. The container and venv survived the restart.


Generated by Claude Code

The 29 October rework cut the first staging's 1,699 rows to 432 under
README rules 8 and 9 and stopped before its report. Checking each
instruction of its brief against the rows found two loose ends, fixed here:
Chart 1.6's six headroom-bridge labels used `component`, which rule 9
reserves for parts of a Table 6.1 measure, so they now name their bar in
`chart_bar`; and two manifest notes named local machine paths, which a
public repository should not carry.

The family passes the contract with 0 problems and its stage.py (kept on
uk/ab2026-harvest-staging) reproduces the gzip byte for byte. No primary
was re-opened: this session's network policy refused gov.uk, so the
15-row spot check the brief asks for could not run, and NOTES.md says so.
COUNTS.json and PROPOSED_KEYS.md are regenerated for the landed family.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6HuQUQ1obM9JvyhJbE
MaxGhenis pushed a commit that referenced this pull request Nov 3, 2026
The landed family's edits sit in an explicit block at the end of the
embedded data: Chart 1.6's labels move from `component` to `chart_bar`,
and two manifest notes lose local machine paths. Keeping the script here
means the gzip on uk/ab2026-harvest can be reproduced from GitHub.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6HuQUQ1obM9JvyhJbE
…7 rows)

The first IFS lane predates README rules 8 and 9. Applying the addendum's
sections 3 and 4 drops 19 rows that are design, not estimates (rates,
payment amounts, thresholds, government targets, scenario inputs, financing
and the DIP funding allocation), and re-keys eight. IFS's three cost-of-
living giveaways are exactly Table 6.1 lines 1 to 3 (5,040 against its
£5bn), so they take the registered package; the full-period VAT row is
line 2; the self-assessment timing and compliance-staff figures match
Table 6.3 items, so they are line 45 components; and the two remaining
proposals use the names rule 2 gives their scopes.

Sections 1 and 2 could not run here. The browser-read pages are listed as
pending for the coordinator, and the seven Flourish data items are listed
as unreached: this session's network policy refused the Flourish hosts and
ifs.org.uk, so no primary was re-opened and no spot check was possible.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6HuQUQ1obM9JvyhJbE
MaxGhenis pushed a commit that referenced this pull request Nov 3, 2026
The addendum's rule 8 drops and rule 9 re-keys sit in an explicit block at
the end of the embedded rows, so the 87-row gzip on uk/ab2026-harvest can be
reproduced from GitHub.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6HuQUQ1obM9JvyhJbE
…inktanks-b review applied

The review approved all three with fixes, applied here as relabels; no
printed value changes.
- The Welsh Table 1 cells now say which are parts and which are totals,
  so an ingest cannot add Resource and Capital to their Total (finding 2).
- The Welsh package key names its scope, the Budget's Barnett
  consequentials, so HM Treasury's figures for the same scope can share it
  (finding 3).
- Real-terms settlement changes get their own metric,
  devolved_funding_real_terms_change, because AB2025's block_grant_effect
  means a policy's effect on the grant with the opposite sign (finding 4).
- The Scottish Government's five rows join the Welsh ones in
  uk_devolved_govts (finding 7), leaving uk_fai_sfc empty.
- CPS's PuFins figure names no Budget decision, so it is context with no
  proposed key (finding 1).
- Trussell's row moves to a new uk_trussell_wpi family with AB2025's slug
  (finding 5).
The README's slug table gains the three slugs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6HuQUQ1obM9JvyhJbE
MaxGhenis pushed a commit that referenced this pull request Nov 3, 2026
uk_devolved_govts' script now writes the Scottish Government's rows too
(moved from uk_fai_sfc, whose script now writes an empty family), and
uk_trussell_wpi's script is uk_thinktank_misc's, filtered to Trussell's
row, so every gzip on uk/ab2026-harvest can be reproduced from GitHub.
uk_niesr's NOTES carries the review's corrected exclusion reason.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6HuQUQ1obM9JvyhJbE
…pplied

The review approved all three with fixes; each is applied in an explicit
block in the family's stage.py (kept on uk/ab2026-harvest-staging).
- TPA's "£93" is the most tax the Pension Tax Allowance removes, the same
  quantity as its "up to £94" maximum, so rule 8 makes it design: the row
  goes (306 to 305) and the figure moves into the 53% row's conditions.
  "About two-fifths" gets its own proposed metric, because the registered
  revenue_share is a group's share of one measure's yield.
- CenTax's £3bn, £0.6bn and £0.4bn are parts of its £4bn and are now
  tagged as components, so an ingest cannot add them to it. Its timely-
  payment derivation matches the red book's Table 6.3 cells exactly. The
  £20 million yield is no longer re-signed, and a 25th-percentile EATR uses
  the registered average_tax_rate.
- IPPR's 0.2pp CPI figure names no scope of its own, so under rule 3 it is
  context rather than a proposed package, and it is an upper bound.
No printed value changes; the families' sites were unreachable from this
session, so no primary was re-opened.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6HuQUQ1obM9JvyhJbE
MaxGhenis pushed a commit that referenced this pull request Nov 3, 2026
Each script now ends with the review's fixes as an explicit block, so the
TPA, CenTax and IPPR gzips on uk/ab2026-harvest can be reproduced from
GitHub. TPA's script reads uk_hmt_tables from its sibling directory.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6HuQUQ1obM9JvyhJbE
The review approved the 597 rows with fixes. p.50's UK tax gap is HMRC's
whole-economy statistic, not a part of Table 6.1 line 45 nor even the
costing's base, so its two rows become context with a null key, as the
script's own context path intended. The p.41 British Business Bank
costing keeps its proposed key, but its notes now warn that it very
probably sits inside a Table 6.1 line and must never be added to one.
The line 45 and 47 component notes now say that their notes do not sum to
the line. No printed value changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6HuQUQ1obM9JvyhJbE
MaxGhenis pushed a commit that referenced this pull request Nov 3, 2026
The review's three fixes sit in an explicit block before build(), so the
gzip on uk/ab2026-harvest can be reproduced from GitHub.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6HuQUQ1obM9JvyhJbE
… 766 rows)

The review approved RF's rows with thirteen mechanical fixes, all applied.
The 25 rows for exactly Table 6.1 lines 61 and 62 take the name README
rule 2 gives that scope. Restated rows name their sources; two RF
statistics are pinned to the averages RF compares them with; six halves'
averages carry Figure 8's cash basis; the £358 and 20 per cent rows are
England, like the row beside them; and a workbook cell holding a floating-
point zero (-7.46e-7 bn, with no bar drawn) is dropped. NOTES loses the
passage contradicting README rule 3's corrected example, fixes a hash
typo, and records Figure 8's lower rebate bars in v2 and the VAT-element
wording as a discrepancy. No printed value changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6HuQUQ1obM9JvyhJbE
MaxGhenis pushed a commit that referenced this pull request Nov 3, 2026
The review's fixes are the P_SAV_PROP constant and an explicit block
before write(), so the 766-row gzip on uk/ab2026-harvest can be
reproduced from GitHub.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6HuQUQ1obM9JvyhJbE
MaxGhenis pushed a commit that referenced this pull request Nov 5, 2026
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6KDzCrvKBEjHk8T31b
MaxGhenis pushed a commit that referenced this pull request Nov 5, 2026
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6KDzCrvKBEjHk8T31b
MaxGhenis pushed a commit that referenced this pull request Nov 5, 2026
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6KDzCrvKBEjHk8T31b
MaxGhenis pushed a commit that referenced this pull request Nov 5, 2026
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6KDzCrvKBEjHk8T31b
MaxGhenis and others added 2 commits November 5, 2026 00:49
…the a9f81e8 reviews

GPT-6.1 Sol and a cloud session each requested changes at a9f81e8. Digits with no word of their own now take normalization's factor, failing that 1, where they could match at any of six fixed scales. The checker reports a year after a bare "in", a range whose first number exceeds its second unless neither is positive, and a fraction joined to a percentage.

Under rule 4, the checker now marks a bar reading wherever we mention a bar, chart or graph in the locator, or say there that we took a number from a figure. It cancels such a phrase after one negation and not after two, and accepts any dash in a printed place. It also reads a fraction with a direct percentage word, a whole number in words up to the hundreds with a fraction, and a lone dash under a nil note.

We change no staged row, and all 5,952 pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The list ran to 42 words, and "one out of every five" already shows "out of".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis pushed a commit that referenced this pull request Nov 5, 2026
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis pushed a commit that referenced this pull request Nov 5, 2026
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis pushed a commit that referenced this pull request Nov 5, 2026
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
GPT-6.1 Sol and a cloud session each reviewed 3c65fd7 and requested
changes, and this commit answers both.

- `_factors` reads each operand whole, multiplies operations in a row,
  reads a factor in brackets, and returns None for one it can't read, so
  the checker reports that row at any value (R2).
- A whole number with a fraction takes normalization's factor and reads
  a short multiplier (A3, R4).
- The checker reads a dash as zero only under one whole clause that
  declares it for nil, and it reads the figure dash and the horizontal
  bar (A4, R3).
- Rule 4 reads "Fig." and "Figs.", "on", "in", "using" and "of", plots,
  diagrams, visuals, a figure's column, readings by eye and "Figure 1(a)".
  It also counts a second negation anywhere in a clause (A1, A2, R1, R8).
- The tests add each case (A5, R7).
- The README's contract bullets state what the checker now does (B1 to
  B4, R5, R6, R9).

We change no staged row, and the checker finds 0 problems in the 12
families.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
GPT-6.1 Sol and a cloud session each reviewed 8206882 and requested
changes, and this commit answers both.

- `_factors` returns None for an operation on a bracket, sign, currency
  sign or number words, and for an operation in words, so the checker
  reports that row at any value (Sol F1).
- `_CONTINUES` reads "hundred" too, so "x 1 hundred" no longer reads as 1
  (Sol F1).
- Rule 4 looks past a colon, bracket or quote before a figure (Sol F2),
  reads "fig1a" and "figure1.A" (cloud F4), accepts "Figure 1(a) data
  label" as a printed place (cloud F6) and ends a clause at a full stop
  (cloud F14).
- The tests add each case, each short multiplier on a whole number with
  a fraction, and ordinary notes that state no factor (Sol F3).
- The README's contract bullets state what the checker now does (cloud
  F1, F2, F7 to F10; Sol F4).

We change no staged row, and the checker finds 0 problems in the 12
families.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis pushed a commit that referenced this pull request Nov 5, 2026
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwQq6KnYGrWqigco2AXori
GPT-6.1 Sol and a cloud session each reviewed 4707e91 and requested changes.

- The checker's `_UNREAD` now reports an operation in words with any spacing
  ("multiplied  by 1e6", "x by 1e6"), an operation on a wrapped or quoted number
  word ("x (-one million)"), and an operator with no space before it ("£m×1e6",
  "£m;x 1e6", "%/100", "£m/100"). It still ignores a slash between digits ("2/3",
  "2026/27").
- Rule 4 now reads nested wrappers ('read from ("Figure 1.A")'), a word in a
  bracket or quotes before the figure ("read from [the Figure 1.A]"), and a word
  that holds an apostrophe or ends a bracket ("HMT's Figure 1.A", "(HMT) Figure
  1.A"). A word's first character excludes quotes, so reading time stays flat on
  a long run of quotes, and a test holds it under two seconds.
- New tests fail on 4707e91's checker where it missed these forms, and cover an
  unread factor on a number that owns a multiplier, a share, a fraction or nil.
- The README states the forms the checker reports and tells stagers to write
  "page 3 of 4" for a slash that divides nothing.

No staged row changed, and the checker reports 0 problems on 5,952 rows in 12
families.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis pushed a commit that referenced this pull request Nov 5, 2026
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis pushed a commit that referenced this pull request Nov 5, 2026
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis pushed a commit that referenced this pull request Nov 5, 2026
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis and others added 2 commits November 5, 2026 08:39
GPT-6.1 Sol approved 2f60cdb with four low findings, and a cloud session
requested changes with low findings only.

- The checker's `_UNREAD` now reports a fraction as an operand ("times half",
  "/ a quarter", "x 1e6 / half"), a "by" after an operator ("× by 1e6"),
  "times by 1,000", and an operator right after an arrow ("£m→x 1e6",
  "£m→/100").
- Rule 4 now reads a word before a figure that ends in closing quotes or
  several brackets (`read off the "Impact on households" Figure 1.A`,
  `read from "HMT" Figure 1.A`, "read from ((HMT)) Figure 1.A"). A test holds
  reading time flat on long runs of quotes and brackets.
- New tests cover each form and its negation, and "two and a half" and "half a
  million" under an unread factor.
- The README lists "x half" among the forms the checker reports, says that rule
  4 reads the locator, and states that the checker counts "figures" where it
  means numbers.

No staged row changed, and the checker reports 0 problems on 5,952 rows in 12
families.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#251 sorted AB2025's name search on main. We merge it here so #247 can take it
and still show only tranche 2.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis added a commit that referenced this pull request Nov 5, 2026
This takes #231's final head and main, with #251, which sorted AB2025's name
search.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@MaxGhenis
MaxGhenis marked this pull request as ready for review November 5, 2026 14:07
@MaxGhenis
MaxGhenis merged commit f6edbdf into main Nov 5, 2026
16 checks passed
@MaxGhenis

Copy link
Copy Markdown
Contributor Author

Merged at 4dd5281d with a merge commit (f6edbdf6), after gh pr checks exited 0 and GitHub reported MERGEABLE.

  • GPT-6.1 Sol reviewed 4dd5281d and approved it with three low findings. F3 changed one sentence of the description. A follow-up PR answers F1 and F2, on forms no staged row uses (0 of 5,952 rows).
  • Sol's and a cloud session's reviews of 2f60cdb8 led to 05265c2d, and the merge of main in 4dd5281d changed none of this PR's 55 paths.

🤖 Generated with Claude Code

This branch was successfully deployed

1 active deployment
Preview — 4dd5281d Deployed Nov 5, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants