Keep the protected UC LCWRA amount from 2030-31

[MaxGhenis]

Fixes #2282.

Reproducer: policyengine_uk/tests/test_uc_health_element_after_freeze.py::test_protected_amount_is_uprated_by_the_element_index_from_2030_31

Builds on #2284 (the UC Act 2025 s. 3 freeze, merged as 0f20450), which adds the uprating index this PR uses. Retargeted to main.

What was wrong

universal_credit_july_2025_reform (scenarios/uc_reform.py) set the health element only for 2026-27 to 2029-30 (range(2026, 2030)). From 2030-31 the uc_LCWRA_element formula paid every LCWRA claimant elements.disabled.amount, the new-claimant amount, including pre-2026, severe conditions criteria and terminally ill claimants. On main a protected claimant went from £457.15 a month in 2029-30 to £235.86 in 2030-31.

The law

• Universal Credit Act 2025 s. 2(1)(b) only relabels the reg. 36 row for pre-2026, severe conditions criteria and terminally ill claimants; the row has no end date.
• s. 3(1) disapplies SSAA 1992 s. 150(1) and (2) for the LCWRA element (both rows, reg. 27) only "in the tax years ending with 5 April 2026, 5 April 2027, 5 April 2028 and 5 April 2029". The review in 2029-30 governs the April 2030 rates again (s. 150(10)).
• s. 4(1) requires increases to the protected amount only "for any of the tax years 2026-27 to 2029-30".

So both amounts continue from 2030-31 and are uprated under s. 150. For UC amounts that uprating is discretionary (s. 150(2)(b); s. 150(3) does not list them); the model assumes it follows the benefit uprating index, as for every other UC amount.

Changes

• scenarios/uc_reform.py
  • The scenario sets the element from 2026-27 to the end of the data's projection, PROJECTION_END_YEAR (2032), not to 2029-30.
  • A protected amount with no stated value takes the s. 4 floor to 2029-30 (unchanged) and, from 2030-31, the year before's amount times the growth of gov.economic_assumptions.indices.uc_health_element, the index the new-claimant amount follows (Freeze the UC LCWRA element for 2026-27 to 2029-30 (UC Act 2025 s. 3) #2284). The index is the benefit uprating index, held in any year the freeze is in force, and never falls.
  • The new-claimant share is read from a parameter.
  • The scenario's years are one public name, REBALANCING_YEARS (2026 to PROJECTION_END_YEAR), for other code to read (the UK hub's guard for elements.disabled.amount reforms).
• rebalancing/new_claimant_share.yaml (new): 11%, 13%, 16% and 22% for 2026-27 to 2029-30, WPI Economics' estimates for the Trussell Trust (the values that were in the code), held at 22% from 2030-31. That hold is an assumption, not a source: the share keeps rising as the caseload turns over, so holding it understates new claimants and overstates protected amounts from 2030-31. I could not find the WPI publication online, so the parameter cites the Act rather than a link.
• protected_health_element.yaml: the description gives the post-2029 rule.
• Docs: uc-rebalancing.md explains the post-2029 amounts and the share.

Still not covered: from 2033-34, after the data's projection, the scenario no longer sets the element and the formula pays every claimant the new-claimant amount. Removing that limit means computing the element in the formula from the claim status and the parameters, which is #2252's proposal.

Invariants

1. No drop. From 2026-27 to 2031-32, neither a protected nor a new claimant's element falls from one year to the next, for any CPI path; with non-negative CPI their Universal Credit does not fall either (the standard allowance follows CPI down when it falls).
2. Protected path. From 2030-31 the protected amount is the year before's times the element index's growth.
3. Reforms take effect. A 2030-31 change to either amount moves that claimant's element and Universal Credit by the same amount. The scenario projects a protected amount from the changed one; a one-year change to the new-claimant amount gives way to prior law's path the next year, as main's dated reforms define for an uprated parameter. The draw reads the share in every year it sets.

Tests

policyengine_uk/tests/test_uc_health_element_after_freeze.py:

• no drop from 2029-30 to 2030-31, and none to 2031-32, for a protected and a new claimant (element and Universal Credit);
• the protected amount's path from 2030-31 (invariant 2) and the new-claimant amount's;
• a 2030-31 reform to each amount moves Universal Credit by the change in the element;
• Hypothesis over OBR CPI paths for 2024-2032: the protected amount never falls for any path (−5% to +15%); for non-negative paths neither claimant's element nor Universal Credit falls, and the protected amount is never below the new-claimant amount;
• two of Freeze the UC LCWRA element for 2026-27 to 2029-30 (UC Act 2025 s. 3) #2284's tests assumed the formula decides 2030-31 awards; they now state a new claimant, and the reform-dict test changes new_claimant_health_element, the amount the scenario pays;
• the seeded draw reads new_claimant_share in every year to 2031-32, and a reform to it for 2031-32 changes that year's status.

Mutation check. Ending the scenario's loop at 2029 fails 5 tests; keeping the s. 4 floor after 2029-30 fails 1.

Local runs at 727085c (main merged): this file, test_uc_health_element_freeze.py, test_uc_rebalancing.py, test_dwp_published_rates.py, code_health/test_cached_arrays_not_written_in_place.py and test_process_time_reforms.py::test_generated_parameters_are_the_series_processing_builds: 1380 passed, 64 skipped; UC YAML policy tests: 482 passed. Earlier, at b6f3f3f: this file, test_uc_health_element_freeze.py, test_uc_rebalancing.py, code_health/test_cached_arrays_not_written_in_place.py and test_process_time_reforms.py::test_generated_parameters_are_the_series_processing_builds: 84 passed. The impact runs were at 577f549; later commits rename and test, and move the index without changing its values.

Impact

Three full microsimulation runs on enhanced_frs_2024_25.h5@1.58.1, 2029-30 to 2032-33, nothing scaled: main (1a31ef1), #2284 (f1a495d) and this branch (577f549). £bn a year.

	2029-30	2030-31	2031-32	2032-33
universal_credit, this PR minus #2284	0	+5.015	+5.108	+5.159
uc_LCWRA_element, this PR minus #2284	0	+13.982	+14.323	+14.671
gov_balance, this PR minus #2284	0	−5.005	−5.098	−5.149
household_net_income, this PR minus #2284	0	+5.005	+5.098	+5.149
universal_credit, this PR minus main	0	+4.605	+4.700	+4.748
household_net_income, this PR minus main	0	+4.595	+4.691	+4.738

• 2029-30 does not change: the scenario already set it.
• From 2030-31 the uc_LCWRA_element total goes back to its 2029-30 level and rises with it (£31.9bn in 2029-30; £32.7bn, £33.5bn and £34.3bn after), instead of falling to £18.7bn on Freeze the UC LCWRA element for 2026-27 to 2029-30 (UC Act 2025 s. 3) #2284 and £19.9bn on main. uc_LCWRA_element covers every benefit unit with LCWRA (6.55m in 2030-31), on Universal Credit or not, so it moves by more than Universal Credit does.
• Against Freeze the UC LCWRA element for 2026-27 to 2029-30 (UC Act 2025 s. 3) #2284, 2030-31: 1.87m households (3.85m people) gain £5.09bn, £2,726 a year on average: protected claimants keeping the higher amount. 0.033m households (0.16m people) lose £86m, about £2,600 each. On Freeze the UC LCWRA element for 2026-27 to 2029-30 (UC Act 2025 s. 3) #2284 the formula decides 2030-31, and it adds an element for each member flagged with LCWRA (other than the other member of a single claim), with no cap of one per claim, so it pays some benefit units more than one element: in the data, £70m of the £86m is new-claimant benefit units it paid two elements and £16m is protected benefit units it paid three (review round 1, from the saved arrays). The scenario pays one element per benefit unit, as UC Regs reg. 27(4) requires for joint claimants. The formula's multiple elements, in the years the scenario does not set, are UC LCWRA element: joint claimants who both have LCWRA get two elements; reg. 27(4) allows one #2429.
• Against main, 2030-31: 1.75m households gain £4.77bn and 0.55m lose £174m (£318 each). The difference from the Freeze the UC LCWRA element for 2026-27 to 2029-30 (UC Act 2025 s. 3) #2284 comparison is Freeze the UC LCWRA element for 2026-27 to 2029-30 (UC Act 2025 s. 3) #2284's freeze: main uprates the new-claimant amount through 2026-29.
• The new-claimant share comes out at 20.5% of benefit units with LCWRA (weighted) against the 22% parameter: the seeded draw is independent of LCWRA status, so the gap is sampling variation. Holding the share at 22% is the assumption above; a higher share would lower these figures. Max's call on it is queued (d1595).

Protected and new claimants (benefit units with LCWRA, millions; 22% flat): 5.18 protected and 1.34 new in 2029-30; 5.21 and 1.35 in 2030-31; 5.23 and 1.35 in 2031-32; 5.25 and 1.36 in 2032-33.

Sensitivity: a rising new-claimant share

d1595 asks whether to hold the share at 22% or let it rise. This is a fourth real run of this branch with the share on the 2026-29 least-squares trend (11%, 13%, 16%, 22% rise 3.6 points a year): 24.5%, 28.1% and 31.7% for 2030-31 to 2032-33, set through Scenario(parameter_changes=...) on rebalancing.new_claimant_share. Either path is a one-line change to that parameter.

£bn a year	2030-31	2031-32	2032-33
universal_credit, 22% flat minus #2284	+5.015	+5.108	+5.159
universal_credit, rising minus #2284	+4.802	+4.621	+4.390
uc_LCWRA_element, 22% flat minus #2284	+13.982	+14.323	+14.671
uc_LCWRA_element, rising minus #2284	+13.405	+13.186	+12.871
universal_credit, rising minus 22% flat	−0.213	−0.487	−0.769
universal_credit, rising minus main	+4.391	+4.214	+3.979

Rising share	2030-31	2031-32	2032-33
New claimants among LCWRA benefit units	1.54m (23.6%)	1.73m (26.3%)	1.95m (29.5%)
Protected	5.01m	4.85m	4.66m
Against #2284: households gaining	1.80m (£4.88bn, £2,717 each)	1.71m (£4.70bn)	1.62m (£4.51bn)
Against #2284: households losing	0.033m (£86m)	0.035m (£89m)	0.049m (£133m)
Against main: households gaining	1.68m (£4.57bn)	1.60m (£4.41bn)	1.52m (£4.24bn)
Against main: households losing	0.62m (£187m)	0.71m (£208m)	0.78m (£272m)

The weighted share of LCWRA benefit units drawn as new claimants sits below the parameter (23.6% against 24.5% in 2030-31; 20.5% against 22% in 2029-30). The seeded draw is independent of LCWRA status, so the gap is sampling variation in the weighted data, the same on main.

Outputs: ~/reviews/uk-lcwra-s3-freeze/impact2282/results/ (compare.txt, compare_rising.txt and the four runs' arrays).

Review

• Round 1 (GPT-6.1 Sol, head b6f3f3f): approve, with two non-blocking corrections: the loss attribution above (two and three elements, not only two) and the first year the formula decides (2033-34, not 2032-33). Both fixed in 555dc5f, with Freeze the UC LCWRA element for 2026-27 to 2029-30 (UC Act 2025 s. 3) #2284's round 7 note (a reform can set the index directly). File: ~/reviews/uk-lcwra-s3-freeze/review-2428-r1.md.
• Round 2 (GPT-6.1 Sol, head 555dc5f): approve the PR as a whole. Two wording notes: the formula's exclusion of the other member of a single claim (fixed above) and the builder docstring's "other parameters" (fixed in the next commit, with the final review after main is merged).
• Round 3 (GPT-6.1 Sol, head 727085c, the docstring note and the merge of main after Freeze the UC LCWRA element for 2026-27 to 2029-30 (UC Act 2025 s. 3) #2284 landed): approve the PR as a whole. The diff against main is byte-identical to the approved change plus the docstring fix. File: review-2428-r3.md.

Overlaps

• Freeze the UC LCWRA element for 2026-27 to 2029-30 (UC Act 2025 s. 3) #2284: merged (0f20450); adds the index.
• UC rebalancing reforms don't take effect unless the Scenario applies before the data load #2252 (the scenario sets the element at build time): this PR keeps that design and extends its years; the formula route in UC rebalancing reforms don't take effect unless the Scenario applies before the data load #2252 would also remove the limit from 2033-34.
• With UC rebalancing switched off, every LCWRA claimant gets the Act's £217.26 new-claimant amount #2283 (rebalancing.active false gives everyone the Act's £217.26): unchanged here.
• Apply the Universal Credit Act 2025 standard allowance uplifts to 2029-30 #2242 (s. 1 standard allowance uplifts): the s. 4 floor reads the standard allowance, so Apply the Universal Credit Act 2025 standard allowance uplifts to 2029-30 #2242 moves the protected amount to 2029-30 and therefore every later year here.

axiom: TheAxiomFoundation/rulespec-uk#367 queued (UC Act 2025 ss. 3 and 4, labelled pe-parity; s. 3 companion cases added there for #2284). rulespec-uk uk/regulations/uksi/2013/376/36.yaml lcwra_protected_claimant_amount holds 429.80 with no later version; the post-2029 s. 150 uprating is the s. 150 module #367 queues.

Test-year fix (a66d1ad)

CI failed test_dataframe_dataset.py::test_an_export_names_the_people_in_each_year_the_rebalancing_records at 727085c: it hard-coded the rebalancing years as 2026 to 2029, but the scenario now runs to REBALANCING_YEARS (2026 to 2032). The test now derives its expected columns from that constant. Two tests with the same hard-coded years are changed the same way (test_benefit_cap_specified_benefits_properties.py, test_uc_rebalancing.py). Test files only; no source, data or parameter changes.

Test results at a66d1ad (-n 4, test fixtures only, no microdata):

• test_dataframe_dataset.py: 78 passed. The same test fails at 727085c with the CI message.
• test_benefit_cap_specified_benefits_properties.py: 26 passed.
• test_uc_rebalancing.py: 17 passed. test_section_4_floor_treats_a_cpi_fall_as_zero hit a Hypothesis too_slow health check under host load in the full-file run; run alone it passes.
• uvx ruff format --check .: 1721 files already formatted; ruff check on the three files passes.

Review of this head is pending with the UK hub. The author of this change did not review it.

🤖 Generated with Claude Code

[MaxGhenis]

Hand-off to the UK hub (session local_51f6a5a6-ef54-422b-aa89-8a8130a675ef).

• Head: 727085c89c36cb5364d7f29f752557fbaf9e4f3a. Ready for review, base main (retargeted after Freeze the UC LCWRA element for 2026-27 to 2029-30 (UC Act 2025 s. 3) #2284 merged as 0f204508), mergeable: MERGEABLE.
• Independent review: GPT-6.1 Sol round 3, APPROVE at this exact head (~/reviews/uk-lcwra-s3-freeze/review-2428-r3.md). It confirms the diff against main is byte-identical to the change approved in rounds 1 (b6f3f3f5) and 2 (555dc5ff) plus round 2's docstring wording. Rounds 1 and 2 and the prompts are beside it.
• CI at this head: pending when handed off (the four Python test shards were still running; the other checks had passed). Nothing has failed.
• Local runs at this head: the affected test files, 1380 passed, 64 skipped; UC YAML policy tests, 482 passed.
• Impact (real Enhanced FRS 1.58.1 runs, in the description): Universal Credit +£5.0bn, +£5.1bn and +£5.2bn a year for 2030-31 to 2032-33 against Freeze the UC LCWRA element for 2026-27 to 2029-30 (UC Act 2025 s. 3) #2284's code, with 1.87m households gaining about £2,700 a year and 0.033m losing (the formula had paid them two or three elements; reg. 27(4) allows one). 2029-30 is unchanged. A rising new-claimant share (the d1595 alternative) gives +£4.8bn, +£4.6bn and +£4.4bn.
• Decision: d1595 (Max, low, due 2026-10-29): hold the new-claimant share at 22% after 2029-30 (this PR's default) or let it rise. It does not block the merge; either ruling is a one-line change to rebalancing.new_claimant_share.
• For the hub's Scenario.from_reform(dict).simulation_modifier silently drops statutory changes on a built simulation #2426 guard: the scenario's years are uc_reform.REBALANCING_YEARS; the index is gov.economic_assumptions.indices.uc_health_element.
• Depends on: nothing open (Freeze the UC LCWRA element for 2026-27 to 2029-30 (UC Act 2025 s. 3) #2284 merged).
• Follow-ups owed (issues, unowned): UC LCWRA element: joint claimants who both have LCWRA get two elements; reg. 27(4) allows one #2429 (the uc_LCWRA_element formula pays more than one element per claim in years the scenario does not set, including 2025-26), With UC rebalancing switched off, every LCWRA claimant gets the Act's £217.26 new-claimant amount #2283 (rebalancing.active false), UC rebalancing reforms don't take effect unless the Scenario applies before the data load #2252 (computing the element in the formula, which would remove the limit from 2033-34).

axiom: TheAxiomFoundation/rulespec-uk#367 queued (UC Act 2025 ss. 3 and 4, pe-parity).

[MaxGhenis]

Hand-off: test years derived from REBALANCING_YEARS

Head: a66d1adbc0b859f40622d2eec305d6127803f499, one commit on 727085c8 (fast-forward push, no history rewritten).

Fix. test_an_export_names_the_people_in_each_year_the_rebalancing_records now expects the status columns for every year in REBALANCING_YEARS and the identifier columns through REBALANCING_YEARS[-1]. Also changed: test_benefit_cap_specified_benefits_properties.py (REBALANCED_YEARS = list(REBALANCING_YEARS)), and test_uc_rebalancing.py (YEARS and three loops use the constant). Test files only.

Results at a66d1adbc (-n 4, test fixtures only, no microdata):

• test_dataframe_dataset.py: 78 of 78 passed, run in 8 foreground chunks, no skips. The same test fails at 727085c8 with the CI message.
• test_benefit_cap_specified_benefits_properties.py: 26 of 26 passed.
• test_uc_rebalancing.py: 17 passed. test_section_4_floor_treats_a_cpi_fall_as_zero hit a Hypothesis too_slow health check in the full-file run under host load. Run alone it passes (1 passed, 33s).
• uvx ruff format --check .: 1721 files already formatted. ruff check on the three changed files: clean.

Audit of hard-coded rebalancing years.

• Changed: the three files above.
• Left unchanged, with reasons:
  • test_uc_health_element_freeze.py FROZEN_YEARS and the assertion near line 263: the law's freeze period (2026-27 to 2029-30) is a fixed legal range, not the scenario's years.
  • code_health/test_cached_arrays_not_written_in_place.py near lines 263 and 383: string snippets fed to the AST detector, not year-dependent.
  • test_dataframe_dataset.py __2029 columns (about lines 620-648): an unrelated fixture year in another test.
  • test_uc_health_element_after_freeze.py: already uses the constant.

Status. CI is pending at this head (Test Python 1-4, YAML, docs, fix evidence). mergeable is MERGEABLE, but the PR is BLOCKED until CI and review finish, so it is not ready. This session did not review its own change. Review and landing belong to the UK hub.

[MaxGhenis]

Landed by the UK hub (merge queue, pinned to head a66d1adbc0b859f40622d2eec305d6127803f499).

• Gates at this head: gh pr checks exits 0 (15 pass), not a draft, no CHANGES_REQUESTED.
• Reviews:
  • GPT-6.1 Sol round 3 APPROVED 727085c8 (the whole change; review-2428-r3.md).
  • The hub's independent Sol delta review (Subfleet 20261016-025747-uk2428-delta-review-b) APPROVES a66d1adbc: a test-only commit that derives the expected years from REBALANCING_YEARS. It ran 122 tests, reproduced the old CI failure and verified the fix. The seeded CPI property passed 200 examples.
• Law: UC Act 2025 s. 2 keeps both reg. 36 rows past 2029-30; ss. 3 and 4 end in 2029-30. Fixes From 2030-31 every UC LCWRA claimant gets the new-claimant amount; the protected amount stops with the 2026-29 modifier #2282.
• Impact, as the body states (1.58.1): UC +£5.0bn, +£5.1bn and +£5.2bn a year for 2030-31 to 2032-33, against Freeze the UC LCWRA element for 2026-27 to 2029-30 (UC Act 2025 s. 3) #2284's code. 2029-30 is unchanged.
• Open methodology call, not blocking: the new-claimant share after 2029-30 is held at 22% (cos d1594).
• Follow-ups:
  • Apply the Universal Credit Act 2025 standard allowance uplifts to 2029-30 #2242 (the standard allowance uplift, in review) lowers the 2029-30 protected amount this PR projects from, so Apply the Universal Credit Act 2025 standard allowance uplifts to 2029-30 #2242 re-measures 2030+ after this lands.
  • Fix silent UC health reform routes #2434 (the guard) reads REBALANCING_YEARS.