UK pensions: calibrate State Pension, Pension Credit and pension contributions to the resident administrative facts (PolicyEngine/chronicle#305), pin the State Pension age boundary, fix the pension spine stages, and the engine pieces a State Pension uprating analysis needs

[juaristi22]

Why

Measured on the #1045 head (arm ch3, c306cc07d, the latest candidate; every spine and terminal gate passes) with policyengine-uk 2.102.3 at FY2025-26, against the pension facts on the feed and in PolicyEngine/chronicle#305. The #979 national build, the v21c measurement and the eFRS 1.57.3 file were measured alongside.

1. State Pension is about 9% short of what UK residents receive, and the calibration cannot see it.
   • Engine: £131.1bn paid to 11.33m people.
   • UK residents: about £144bn and 12.45m. That is GB residents from DWP's Spring 2026 table less the "paid outside UK" line (£146.07bn − £5.63bn and 13.20m − 1.09m), plus Northern Ireland (NISRA: 0.34m, about £3.8bn).
   • The only bound level is obr.state_pension, at −9.7%. It has drifted down across builds (−8.0%, −9.0%, −9.5%, −9.7%) and is a single row among 1,090.
2. The missing State Pension sits with non-taxpayers.
   • SPI 2023-24 Table 3.6 bands are +6.4% (£20–30k), +16.6% (£30–40k), +8.3% (£40–50k) and +25.2% (£50–70k, deferred under D6 to 2026-10-21), while the counts fit.
   • The taxpayer bands sum to £96.2bn and 8.36m recipients; the model carries £102.3bn and 8.33m there. The resident total therefore implies about £48bn and 4.1m non-taxpaying recipients, and the model has £27.9bn and about 3.0m. The shortfall is almost entirely below the tax threshold.
   • Pensioners sit too high in the income distribution. Shares by quintile before housing costs, bottom to top: 18.7 / 23.3 / 22.2 / 19.4 / 16.5%, against HBAI FYE 2025's 22 / 25 / 21 / 18 / 14.
   • Pensioner relative poverty is 13.3% before housing costs, against 16%. The UK CGT: interpolate the sub-exempt remainder, anchor the clone incidence, and fence projected entrants (#970) #979 build matched HBAI closely (23 / 25 / 23 / 16 / 14; 16.7%), so the move came with the income anchors (UK income anchors at the calibration year and reserved top-tail income rows (income-anchor lane, PolicyEngine/chronicle#280) #1006) and the SPI-first reorder (UK spine: make SPI support households internally consistent (SPI-first order, housing shell) #1012).
3. The State Pension age boundary is free.
   • Only five-year bands (65–69) are bound. Single ages 64–68 are 0.73 / 0.98 / 0.64 / 0.73 / 0.59m, against ONS 0.80 / 0.77 / 0.75 / 0.71 / 0.68m: age 65 is 27% high and age 66, the first year of State Pension age, 15% low.
   • 0.25m people under State Pension age (mostly exactly 65) carry state_pension_reported, which the engine drops.
   • The State Pension age population swings by build: 12.19m on ch3 against 12.51m on UK CGT: interpolate the sub-exempt remainder, anchor the clone incidence, and fence projected entrants (#970) #979.
   • New State Pension recipients are 4.23m against DWP's 5.05m (−16%).
4. Pension Credit pays too many small awards.
   • Spending fits (−0.1%), but the caseload is 1.665m against 1.382m (+20%).
   • Savings Credit only is 0.298m against 0.176m (+69%), and Savings Credit spending £0.52bn against £0.36bn.
   • Take-up is one flat draw at 0.70 (DWP FY2019-20).
5. Pension contributions are untargeted.
   • Employer contributions are three times employee contributions: £127.8bn against £35.2bn. DWP's savings trends put eligible savers' 2025 contributions at 61% employer, 27% employee and 12% tax relief of £166.1bn, i.e. about £101bn employer and £45bn employee.
   • Salary-sacrifice amounts are free: £22.5bn now, against £30–32bn on earlier builds. Only user counts are bound.
6. The engine cannot run the analysis as it stands (policyengine-uk 2.102.3):
   • Dict reforms of the triple-lock switches do nothing.
   • State Pension age stays at 66, with no rise to 67 over 2026–28.
   • Basic→new reclassification double-counts additional State Pension: +£5.8bn by FY2030-31.
   • The Pension Credit guarantee for FY2025-26 is £221.85 against the published £227.10.
   • The April 2027 triple-lock rate is 3.4%, against the OBR's 3.7%.
   • Additional State Pension is uprated by the triple lock rather than CPI.
7. The distribution of a State Pension reform is not yet stable. In an illustrative CPI-only-from-April-2027 probe:
   • The budget effect is stable across builds: +£1.07bn (FY2027-28) to +£2.70bn (FY2030-31), with a bootstrap standard error of at most £0.13bn.
   • The decile pattern moves with the data. The bottom three deciles carry 27% of the loss on ch3 against 31% on UK CGT: interpolate the sub-exempt remainder, anchor the clone incidence, and fence projected entrants (#970) #979, and decile 7 carries 14.5% against 9.4%.
   • Decile 4 rests on one heavy record (standard error 0.06pp against 0.01–0.02 elsewhere).
   • The effective sample size among people aged 66+ is 2,303.

Work packages

Delivery is one microcosm PR, stacked on #1045 while it is open, plus PolicyEngine/policyengine-uk#1899 and one follow-up engine PR.

WP0: re-pin the Chronicle feed to the post-PolicyEngine/chronicle#305 artifact

• Scope. Build the consumer artifact from Chronicle main at the Add UK pension facts: State Pension, Pension Credit, workplace pensions, salary sacrifice and Winter Fuel Payment chronicle#305 merge. It must be a strict superset of 505e0e7, with the schema unchanged. Move chronicle_feed.json, vendor the resources, and regenerate the references, membership, compile-parity receipts and coverage manifest (the docs/uk-chronicle-feed-repin.md procedure used by UK CGT: BADR-qualifying gains, the Table 3 income joint and wealth-conditioned gain carriers (#1014) #1045).
• Acceptance. Every existing target is byte-identical before and after.
• Base. main still pins 5324aa2, and the 505e0e7 re-pin lives on UK CGT: BADR-qualifying gains, the Table 3 income joint and wealth-conditioned gain carriers (#1014) #1045, so this stacks on UK CGT: BADR-qualifying gains, the Table 3 income joint and wealth-conditioned gain carriers (#1014) #1045.

WP1: State Pension level and shape for UK residents

• Consumer operations first.
  • Extend count_x_mean so it pairs Add UK pension facts: State Pension, Pension Credit, workplace pensions, salary sacrifice and Winter Fuel Payment chronicle#305's count and mean facts (by concept, not by id suffix), annualises weekly means (× 52) and composes with the monthly window.
  • Extend the band-label parser to "£X to under £Y" and "Under £Y", reading each row's own upper bound (all-zero series are dropped, so the next sibling's lower edge is wrong).
  • Add the England-and-Wales and Northern Ireland geography pins in tools/generate_uk_target_references.py.
• Rows (new family dwp_state_pension):
  • GB-resident recipients and amount from the Stat-Xplore cubes over the FY2025-26 points; Northern Ireland from the DfC cube.
  • Recipients by single year of age (66–89, 90+) × gender × type. Fan-out pins one dimension, so there is one target per gender × type.
  • Region × type (the nine English regions, Wales and Scotland).
  • Weekly amount bands × type.
• Diagnostics. obr.state_pension and DWP's "paid outside UK" and GB-plus-overseas lines become diagnostics (the obr.capital_gains_tax precedent). The SPI Table 3.6 £50–70k deferral (D6) is re-measured, then retired or re-ruled.
• Acceptance:
  • Resident level and caseload within 2%.
  • Type × gender within 5%.
  • Single-age counts within 5% for ages 66–75.
  • SPI 3.6 bands no worse.
  • The non-taxpaying remainder within 10% of the implied figure.

WP2: the State Pension age boundary and the pension-age SPI channel (spine)

• Root cause.
  • SPI donors get ages drawn uniformly within their SPI age band (65–74), so a copy aged 65 receives SPI State Pension at the 65–74 receipt rate. The stage-2 bridge then turns that into state_pension_reported.
  • Band carriers pool by income band × region with no age term, which is where the recipients aged 16–54 come from.
  • spi_income and spi_band_donors disagree on the band edges: (65,74)/(74,90) against (65,75)/(75,…).
• Fix.
  • Draw donor ages by ONS single-year shares, with SPI State Pension recipients drawn only at State Pension age or over.
  • The bridge requires age ≥ State Pension age; a postprocess rule zeroes State Pension, Pension Credit and Winter Fuel Payment reports below State Pension age, with a receipt.
  • Band carriers pool by income band × age band.
  • Unify the edges.
• The channel's composition (measured before committing). The SPI support channel is 10,000 FRS households drawn unweighted and given 50% of each region's prior mass, and every adult in it carries taxpayer-only SPI incomes. Stratifying its prior share by region × State Pension-age status keeps non-taxpaying pensioners represented. This is the likely root of both the non-taxpaying State Pension shortfall and pensioners sitting too high (the pattern appeared with UK income anchors at the calibration year and reserved top-tail income rows (income-anchor lane, PolicyEngine/chronicle#280) #1006 and UK spine: make SPI support households internally consistent (SPI-first order, housing shell) #1012).
• Calibration. Bind ONS single ages 64–68 × sex (on the feed; mid-2024).
• Acceptance:
  • Zero recipients below State Pension age.
  • Single ages 64–68 within 3%.
  • The State Pension-age population within 1% of ONS.
  • Spine gates pass.

WP3: Pension Credit

• Rows. Benefit units by type (Guarantee only / Savings only / both) × partner × age band (GB), the Northern Ireland total, and the Guarantee/Savings spending split.
• Take-up (stage). would_claim_pc is today one 0.70 draw over all benefit units, made before the SPI chain and never refreshed. Restrict it to benefit units with an adult over State Pension age, band it by entitlement type (the uc_reporter_redraw precedent), set the rates against the admin caseloads, and refresh after the SPI chain.
• Prerequisite. Correct the engine's FY2025-26 guarantee and Savings Credit threshold (WP6) before measuring.
• Acceptance: caseload within 3%, each type within 5%, spending within 1%.

WP4: pension contributions

• Employee and personal contributions.
  • Bind SPI Table 3.8 (net-pay and relief-at-source contributions by income band; on the feed, unbound).
  • Remove the personal-contribution clip at the 95th percentile of all amounts. The clip leaves the median and 99th percentile both at £1,441 and the total at £2.5bn, against HMRC's £15.9bn of individual personal-pension contributions.
• Employer contributions (stage). Replace employer = 3 × employee (the incumbent's "rough estimate") with a vendored rate resource: at minimum a ratio fitted to DWP savings trends, better ASHE rate bands by sector and employee-rate band. ASHE counts salary sacrifice as employer contributions, so compare employer plus salary sacrifice against the ASHE and DWP totals.
• Salary sacrifice.
  • Amounts drift (£30.5bn on UK CGT: interpolate the sub-exempt remainder, anchor the clone incidence, and fence projected entrants (#970) #979, about £22bn on UK CGT: BADR-qualifying gains, the Table 3 income joint and wealth-conditioned gain carriers (#1014) #1045) because only headcounts are bound, and the OBR base covers FY2027–30 only.
  • Anchor the amount at a ruled 2025 base, or anchor the conversion stage on headcount and amount together.
  • Vendor salary_sacrifice_anchor.json from Chronicle, per the fixtures-from-Chronicle rule.
• Relief rows (Tables 6.1/6.2): only with a counterfactual route (Calibrate the UK national build from Ledger-backed targets #623, Swap the microcosm-built UK spine into the national build and retire the incumbent feed #757; ruling).
• Acceptance: Table 3.8 bands within 10%; employer and employee totals within 5% of their anchors; salary-sacrifice users unchanged.

WP5: the other pension-age benefits a State Pension change interacts with

• Attendance Allowance recipients by country.
• Winter Fuel Payment 2025/26 recipients (England and Wales; the engine variable is household-level, so it needs a person-count rule). HMRC's count above £35,000 as a diagnostic.
• Pension-age Housing Benefit (UK spine: housing benefit £2.7bn against OBR £12.1bn — pensioner and supported-accommodation HB missing after the UC migration #867).
• Scotland's pension-age payments where the engine models them.
• Each within 5%, or diagnostic with a reason.

WP6: the engine (policyengine-uk)

• State Pension age.
  • Set State Pension age from date of birth, including the rise to 67 policyengine-uk#1899 (open) models the rise to 67 by date of birth.
  • It deletes gov.dwp.state_pension.age.{male,female}, which microcosm reads in uk_runtime/frs_take_up.py:132-133. Migrate microcosm with the engine bump.
  • On the enhanced FRS it lowers State Pension by £1.6bn (FY26-27) to £7.7bn (FY30-31).
• Follow-up PR after Set State Pension age from date of birth, including the rise to 67 policyengine-uk#1899:
  1. Reforms of triple-lock inputs propagate.
     • Dict reforms edit the tree after processing and do nothing.
     • Scenario(parameter_changes=...) re-processes, but a "2027" key lands a year late.
     • The fix is one reform path that re-runs the derivation when its inputs are touched.
  2. The basic→new reclassification double count, measured at +£5.8bn by FY30-31.
  3. Additional State Pension uprated by CPI, not the triple lock.
  4. Pension Credit FY2025-26 rates (guarantee £227.10 / £346.60, Savings Credit threshold £198.27 / £314.34, carer and severe disability additions), a CPI-uprated Savings Credit maximum with a derived threshold, and the guarantee's uprating rule.
  5. The April 2027 rate: the OBR's 3.7% assumption until the announced rate (May–July earnings 3.9%).
  6. A switch for the promised income-tax easement for pensioners on State Pension alone from FY27-28.
  7. A guard against datasets that carry a state_pension column.
• Until the follow-up merges, reforms must pass the basic and new State Pension amounts for every year as dict keys "2027".."2030".

WP7: validation

• HBAI: pensioner quintile shares and pensioner relative poverty against HBAI FYE 2025, as validation only (HBAI is FRS-derived).
• DWP path: the baseline State Pension and Pension Credit path against DWP's forecast to FY2030-31, after WP6. Ages stay static across years, so it is read with that caveat.
• Reform stability: a State Pension uprating reform's budget and decile effects, with bootstrap standard errors, on the final build and a second calibration seed. Acceptance: budget within 3%, decile loss shares within 1 point.

Stacking and order

1. The engine: Set State Pension age from date of birth, including the rise to 67 policyengine-uk#1899, then the follow-up PR, released. Then the microcosm engine bump with the State Pension age migration. The FY2025-26 Pension Credit fix moves the 2025 calibration, so it precedes WP3's measurement.
2. WP0 → WP1 (consumer operations, then rows) → WP2 → WP3 → WP4 → WP5, one commit series.
   • Every stage change regenerates the coverage manifest, the H2 parity fixture and the digests at its own commit.
   • Each step is measured on a licensed arm against the ch3 control, one job at a time.
   • WP1's single-age rows and WP2's stage fix both move mass around ages 65–66. Measure them separately so the effects are not confounded.
3. The branch stacks on UK CGT: BADR-qualifying gains, the Table 3 income joint and wealth-conditioned gain carriers (#1014) #1045 while it is open (no CI on stacked PRs, so run the local funnel plus the targeted shards) and rebases onto main when it merges.

Related, not in scope

• The CGT fence and heavy-row work (UK CGT: sub-exempt gains capped at exactly the annual exempt amount become 10.8m taxpayers under uprating; interpolate the remainder below the threshold and anchor incidence #970, UK CGT: impute BADR gains, bind the Table 3 gains-by-income joint, and condition gain carriers on wealth (post-#979 level and income-interaction findings) #1014, UK CGT: BADR-qualifying gains, the Table 3 income joint and wealth-conditioned gain carriers (#1014) #1045, UK spine follow-ups after #1045: SPI band donors as a mass-conserving support channel, WAS tenure coherence, the residential draw's bound and Table 7 composition, then a national build #1063).
• Pension wealth (UK source packages: ONS WAS Round 8 household wealth distribution (total wealth, components, percentile aggregates) chronicle#205).
• Winter Fuel Payment recovery modelling (Model Winter Fuel Payment high-income recovery separately from entitlement policyengine-uk#1760).
• Employer NI stopping at State Pension age (Employer NI is zero for employees over state pension age policyengine-uk#1894 and Charge employer NI on employees over state pension age policyengine-uk#1895).
• Multi-year runs, to be a separate piece. The engine copies the 2024 cross-section forward with static ages and total-population weight growth: no ageing and no cohort turnover. The candidate there is year-specific weights calibrated to the ONS 2024-based single-age projections and scored against DWP's caseloads to FY2030-31 (see Fix age progression in multi-year projections for two-child limit policyengine-uk#1351, closed, for why incrementing ages was rejected).
• Any reform costing: this issue makes the data and the engine fit for one.

Decisions needed

1. The resident State Pension window. FY2025-26 points (May, Aug and Nov 2025, and Mar 2026) or calendar-2025 points. The engine pays FY2025-26 rates in period 2025.
2. The fitted resident source. The Stat-Xplore cubes plus Northern Ireland (with DWP's forecast lines as diagnostics), or the reverse.
3. Single ages. Split the 65–69 band into singles, or bind singles beside it.
4. obr.state_pension. Demote it to a diagnostic. And keep obr.pension_credit bound, or rely on the DWP type rows.
5. Employer contributions. A savings-trends ratio or ASHE rate bands; and whether the DWP totals are bound or diagnostic.
6. Salary sacrifice. The 2025 amount anchor, and whether the relief rows come in through a counterfactual route or stay on Calibrate the UK national build from Ledger-backed targets #623 and Swap the microcosm-built UK spine into the national build and retire the incumbent feed #757.
7. State Pension found below State Pension age. Zero it with a receipt, or reroute it to other social security income.
8. Pension Credit take-up rates by type. Solved against the admin caseloads, or DWP's FRS-based take-up estimates.
9. The Pension Credit guarantee after 2026. Earnings (the statutory floor), or following the triple lock (recent practice).
10. The income-tax easement for State-Pension-only pensioners. Model it behind a switch now, or wait for the mechanism.
11. Completing State Pension receipt for 66+ non-reporters. Optional, evidence first.

🤖 Generated with Claude Code

[juaristi22]

#1084 is merged (fe4e38d). It delivered WP0–WP5 and WP7's HBAI and reform-stability checks; the measurements and rulings are in experiments/1069-uk-pensions-receipts.md. This issue stays open for what's left, most of which is blocked on policyengine-uk.

Blocked on policyengine-uk

• Set State Pension age from date of birth, including the rise to 67 policyengine-uk#1899 (open PR: State Pension age by date of birth, including the rise to 67). It blocks:
  • c0, microcosm's engine bump with the State Pension age migration. uk_runtime/frs_take_up.py reads gov.dwp.state_pension.age.{male,female}, which #1899 deletes.
  • The State Pension type boundary. Until then state_pension_type keys on age and sex. The 70–74 boundary zig-zags, and single ages 69–73 are off by 10–20%.
  • Two items of the engine follow-up below: the basic-to-new reclassification double count, and Additional State Pension uprated by CPI rather than the triple lock.
• The engine follow-up PR (WP6), to be filed after #1899. Four of its items are drafted locally; nothing is filed yet. The items:
  • Reforms of triple-lock inputs propagate.
  • The Pension Credit FY2025-26 rates and the guarantee's uprating rule. Until then Pension Credit calibrates at the engine's current FY2025-26 values, which WP3 wanted corrected first.
  • The April 2027 rate.
  • A guard against datasets that carry a state_pension column.
  • The two #1899-dependent items above, and the income-tax easement switch (decision 10).
• Model Winter Fuel Payment high-income recovery separately from entitlement policyengine-uk#1760 (Winter Fuel Payment recovery above £35,000). Until it lands, DWP's 11.0m recipients stay a diagnostic.
• Pension Age Disability Payment isn't modelled in the engine, and there is no engine issue yet. So Attendance Allowance in Scotland isn't bound.
• WP7's DWP path, the baseline State Pension and Pension Credit against DWP's forecast to FY2030-31, runs after the engine follow-up.

Open in microcosm, not blocked on the engine

• Single ages. WP2's 3% acceptance isn't met: 65 is −17%, 64 −6%, and 62 and 63 are +15% and +12%. A ruling is owed on the R3 split (65, 66, 67–69).
• SPI Table 3.8 contributions are out of the fit, because Table 3.8 splits by relief mechanism and the engine's columns by provider. A ruling is owed on binding the combined amounts against employee plus personal contributions.
• The salary-sacrifice amount. £34.3bn is sacrificed against the £24.6bn implied, because the forest predicts sacrifice for people with no pay. The fix is UK spine follow-ups: SPI donors funded, WAS coherence, residential split, Child Benefit, release-cut blockers (#1063) #1089's c7 (UK spine follow-ups after #1045: SPI band donors as a mass-conserving support channel, WAS tenure coherence, the residential draw's bound and Table 7 composition, then a national build #1063).
• The income-tax relief rate bands wait for a band-test fix, and their exclusions expire 2026-11-25. Bands or total needs deciding before it lands, so nothing double-binds.
• The employer NICs relief is restated from 13.8% to 15% by the rate only; that ruling is to confirm.
• The Pension Credit Guarantee/Savings spending split isn't bound; only OBR's total is.
• The salary-sacrifice anchor isn't vendored from Chronicle yet; the values already match the feed.
• Remaining gaps. The SPI £20–30k State Pension band is +13.1% (8% too many recipients). Non-taxpayer State Pension is short of what HMRC and DWP imply.

Of the decisions listed above, 9 (the Pension Credit guarantee after 2026) and 10 (the easement switch) belong to the engine follow-up; the plan ruled the rest.