[MaxGhenis]

Current state (16 Oct 2026)

Title: Public Pension Indexation Under Uncertainty: Fiscal and Distributional Effects (Vahid Ahmadi, Max Ghenis). Draft; not for merge until Max and Vahid sign off.

• Figures (16 Oct): every figure is now Plotly Express in a PolicyEngine template built from the dashboard (Inter type, light grid, legend under the plot, teal logo) with the dashboard's colours, exported as vector PDF. Authors: Vahid Ahmadi first, then Max Ghenis.
• Rewritten from scratch on the published pre-Budget release (main: Pre-Budget certified release: rebuild on policyengine 6.2.6, model-conditional expected value, dashboard #30, Merge main into model-v2 (clears #24's conflict) #31, model-v2: certified bundle, engine port and guarded rehearsal #24). Every figure is read from data/results.json. Keywords: pension indexation; stochastic microsimulation; triple lock; United Kingdom; fiscal cost; distributional analysis.
• PDF on the uk-ai-study LaTeX template (paper/latex/, built by paper/render.sh → build.py, Tectonic), with figures and tables restyled on uk-ai-study's figure style and the dashboard palette. 26 pages.
• Reviewed in two rounds by a referee, a public-economics professor, a UK pensions-policy expert and a copy/figure editor, then fact-checked (180 claims against results.json and sources; 8 fixed). Final referee verdict: accept as a working paper; no blocking issues.
• Headline figures (2039-40): expected saving £11.1bn in State Pension spending, £7.9bn after tax and benefits (UK); other calibrations and model versions £8.7bn–£12.0bn, mostly less; £0.9bn on the OBR central forecast; £10.3bn on the OBR long-term triple-lock assumption; Great Britain £10.8bn (£12.8bn scaled for survey coverage) against DWP's £15bn; about 27% of households with less income; pensioner poverty after housing costs up to 0.8 points higher on a typical path. Provisional until the 28 October 2026 Budget.
• Checks: tests/test_paper.py 50 passed; full suite 1,779 passed, 2 skipped; scripts/check_assumptions.py passes.

Open items

• data/results.json (and the dashboard copy) carry a paper_rule_arithmetic block written by scripts/paper_rule_arithmetic.py (rule arithmetic on the regenerated draws, no simulation). Any rebuild of the results, including the post-Budget binding run, drops it: rerun the script after the rebuild (paper/render.sh notes this; paperdata.py refuses to load without it).
• Deferred to the next full rebuild, because they change src/ hashes: the unnormalised estimator in expected_value.reweighted() (the paper uses the corrected one in paperdata.py) and the editorial bracket in src/triple_lock/dwp.py's definition (the paper no longer quotes it; a test enforces this).
• The HTML manuscript is now built with pandoc (not Quarto); the /paper wrapper is at revision 6.

---

Summary

This adds a working paper, The Burnham plan and the State Pension triple lock, by Max Ghenis and Vahid Ahmadi. It is a Quarto manuscript that builds the answer up in the dashboard's six steps. It is served at /uk/triple-lock/paper/ as a shelled wrapper around a versioned iframe, following the PolicyBench /paper pattern.

Stacked on #7, which stops the results file publishing one survey household's weight. The paper renders from the results rebuilt there (revision 406ee7a; numbers unchanged). Merge #7 first; this PR's diff then shrinks to the paper commits.

What the paper says (every figure from data/results.json)

• On the OBR's central forecast the plan saves £0.66bn in State Pension spending in 2039-40 (£0.43bn net). Earnings lead almost every year after the switch, so the only difference comes from the floor years 2030–31.
• The expected saving over 50,000 calibrated paths, 198 of them run in full, is £8.41bn gross (Monte Carlo SE £0.04bn) and £5.39bn net (SE £0.04bn) in 2039-40. Other calibrations give £6.6bn to £10.7bn gross. DWP's figure is £15bn, on one path through Pensim3, for Great Britain.
• About 20% of households lose. A pensioner at the Pension Credit guarantee loses nothing, because Pension Credit and a lower tax bill replace the loss; one with a £15,000 private pension bears 80%.

Structure

1. How the triple lock works: the 2011–2026 inputs, what set each rise, the cumulative index, the plan's rule quoting DWP, what the plan would have paid from 2012, and the past-years check.
2. The OBR's central forecast: the path and its sources, both rules, and why the saving is small.
3. Another path: the random draw, with the middle and 90th-percentile paths for context; the monthly VAR and the drift shift.
4. One pensioner: four example households, with the share of the loss each bears, the Winter Fuel Payment threshold and the council tax reduction gap.
5. Everyone: gross and net by year, the gross-to-net account, households losing, quintiles, poverty, and the single-record concentration check (contribution and share only).
6. Every path: the expected saving ± MC SE; strata and estimator; calibration choice (backtest and past-years check) and sensitivities; Enhanced FRS vs Microcosm (paired) and coverage against DWP; DWP's £15bn; the OBR's +0.6pp premium against the paths' 0.55pp.
7. Method and limitations, with an appendix on the forecast-distribution backtest (CRPS, energy and variogram scores).

Files

• paper/index.qmd: the manuscript. paper/paperdata.py computes every number from data/results.json (and the DWP text in data/raw/). paper/figures.py draws the figures.
• paper/references.bib: 26 entries, each fetched and checked on 30 September 2026. paper/SOURCES.md records what each one confirmed, including every quoted passage. I read these directly: the BBC live page, the OBR determinants spreadsheet (its "Triple lock" note), SPC Regulations 2002 reg. 17(10), the Social Security (Up-rating of Benefits) Act 2021 s. 1 and the 2011 Up-rating Order. The rest went through a research lane with a verification log. Only the words the BBC page still quotes as the Prime Minister's ("adjust"; "it will hold its value relative to earnings over time") are quoted as his.
• paper/render.sh: renders HTML and PDF in one pass and syncs them to dashboard/public/paper/web/.
• dashboard/public/paper/index.html: the wrapper, revision r1-20260930. Also changed: dashboard/proxy.js (/paper → /paper/); next.config.js (keeps the trailing slash, rewrites /paper/ to the wrapper); eslint ignores public/paper/**; the dashboard footer links to the paper.
• tests/test_paper.py (33 tests). It checks:
  • the wrapper: version lockstep, sandbox, relative URLs, and that its figures match the results;
  • that the committed render comes from the current results;
  • that every citation resolves and every bib entry is cited;
  • that the prose types no result number (a lint over the manuscript text);
  • sentence-case headings, banned words and no probability language;
  • that no survey identifier or weight reaches the paper;
  • property tests for the number formatting.

Review

Two independent Opus reviews ran on the first draft: an adversarial numbers-and-claims audit, and a check of brief coverage, method and prose. A third, second-round review then checked each earlier finding against the revision and found new ones; those are fixed too. The only finding declined is rerunning the model to raise the Housing Benefit pension-age allowance with earnings, which is disclosed as a limitation instead. Among the fixes:

• The past-years check is now correctly described as the uncalibrated model.
• The choice of calibration is explained by effective sample size, not bias alone.
• DWP is quoted rather than paraphrased as a definition.
• The random path's position (2nd percentile) is flagged in steps 4 and 5.
• The MC SE is scoped to path sampling.
• Pension Credit's after-tax income test is cited to SPC Regulations 2002 reg. 17(10), which settled a reviewer's doubt: the model matches the law there.
• The council tax reduction take-up claim is corrected (see below).
• The step between an earnings-linked guarantee and a CPI-linked Housing Benefit allowance is disclosed.
• Microcosm is compared on its own 40-path subsample.

Checks

• CI green on the head commit (Pipeline tests, Dashboard lint/test/build, Vercel)
• pytest -q locally: paper and results tests 63 passed; full suite 150 passed before stacking on Publish no survey household's weight: drop max_household_weight from coverage #7
• dashboard: bun run test 47 passed, bun run lint 0 errors (2 existing warnings), bun run build OK
• next start locally: /uk/triple-lock/paper returns 308 to /paper/, which returns 200 (the wrapper); web/index.html, web/index.pdf, the CSS and every figure return 200.

[NEEDS CITATION] markers left in the paper

None. The first draft had two: the April 2022 suspension and April 2011's rise. Both now cite legislation I verified: the Social Security (Up-rating of Benefits) Act 2021 s. 1, and SI 2011/821. The claim that April 2011 "followed RPI" was dropped because the Order doesn't state it; the paper now says only that that year's up-rating order set a larger rise than the rule gives.

Things in the results that look wrong (not changed here)

• Council tax reduction take-up. METHOD.md and method_limitations say Housing Benefit and council tax reduction respond only for existing claimants.
  • That holds for Housing Benefit (housing_benefit_eligible requires housing_benefit_reported).
  • It does not hold for council tax reduction: would_claim_council_tax_reduction = claims_all_entitled_benefits | reported_ctr, and the Enhanced FRS supplies neither as an input.
  • The paper prints the file's limitations with this one corrected in place.
• Resolution Foundation benchmark. data/benchmarks.csv describes it as "higher than under a smoothed earnings link since 2012". The RF page says the £12.6bn is the OBR's estimate against earnings-linked uprating "this year". The paper words it from the source.
• Pension-age Housing Benefit in the Enhanced FRS is £13.4bn against DWP's £7.3bn, and Pension Credit claims are 1.82m against 1.31m. Both inflate the offsets.
• Guarantee against Housing Benefit allowance. The runs raise the Pension Credit guarantee with earnings, while policyengine-uk raises the pension-age Housing Benefit allowance with CPI. On high-earnings paths the guarantee overtakes the allowance, so crossing onto guarantee credit jumps Housing Benefit to the maximum. On the 90th-percentile path the example council renter gains £873 a year from a £1,552 pension cut. This is the likely source of the single records that move net figures by up to £2.6bn.

🤖 Generated with Claude Code

[vercel]

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

Project	Deployment	Actions	Updated
uk-triple-lock	Ready	Preview	Oct 16, 2026 9:06am UTC

[juaristi22]

Program review

Base repository: PolicyEngine/uk-triple-lock
PR number: 8
Reviewed head SHA: c9b192c
Merge base SHA: 3d43e24
Mode: full
Scope: full audit of main 12ada9a (tree-identical to PR base), including model, dashboard, outputs and tests; plus PR #8 paper, route, render and data references
Source manifest: retained with the local review artifacts
Review status: PARTIAL

Source documents

The DWP policy publication was checked live. Two original OBR workbooks were byte-authenticated; the OBR long-term determinants and DWP benchmark workbooks were inspected as committed copies, without publisher-byte authentication. The full local report preserves their SHA-256 manifest and provenance.

Critical

None.

Should Address

A1 — SHOULD ADDRESS — MAIN model; PR8 headline inherits conditional net estimate (OPEN)

Location: src/triple_lock/engine.py:695-714, docs/METHOD.md:33, paper/index.qmd:82,415-421,604-641. Trigger: Interpret the £5.392bn expected 2039–40 net saving or the 20.2% households-losing estimate as nationally calibrated impacts. Expected: The means-tested offsets and household impacts should be robust to reasonable baselines, or the figures should be labelled prominently as conditional on the selected survey/model. Observed: Committed data/results.json has Enhanced FRS pension-age Housing Benefit £13.448bn in 2026–27 versus £7.269bn in the committed DWP workbook (+85%); Pension Credit caseload 1.821m versus 1.312m (+39%). Microcosm gives £3.019bn pension-age Housing Benefit and 1.258m Pension Credit claims. Its paired-path 2039–40 net estimate is £5.864bn versus £5.392bn on the primary dataset, a £0.472bn difference, against a primary path-sampling SE of £0.036bn. The primary net estimate includes £0.965bn extra Pension Credit, £0.136bn extra Housing Benefit, and £1.864bn less income tax, so at least one of the materially misaligned baselines bears directly on the reported offset. The paper describes these discrepancies, static rents, take-up gating, and a possible single-record cliff; the new limitations prose at paper/index.qmd:769-792 makes this clearer, but disclosure does not establish calibration. The DWP figure is GB and includes overseas spending in its flat-rate comparator while the model is UK; the paper flags this, so the differences above are diagnostics, not a direct proof of wrong policy effects. Impact: The net and household headline precision exceeds demonstrated baseline robustness; the fiscal gross figure is better anchored mechanically than the offsets. Show net/household sensitivity beside the headline and investigate the Housing Benefit/Pension Credit population and claim logic before treating the net point as an external costing. Evidence: data/results.json coverage.rows, expected_value.estimates, expected_value.paired_difference; src/triple_lock/dwp.py:22-44,65-89; DWP's State Pension uprating explicitly covers GB, direct AME and a dynamic population model. The DWP benchmark workbook is a committed copy; publisher byte identity is unverified.

Coordinator assessment: Code review separately traces the inherited Housing Benefit claimant gate and modelled council-tax-reduction response in src/triple_lock/engine.py:420-548; the committed example exposes the latter. This supports the need for an offset sensitivity, while the direction of combined bias remains unquantified.

A2 — SHOULD ADDRESS — MAIN uncertainty model; PR8 interpretation (OPEN)

Location: src/triple_lock/ts_monthly.py:106-139,191-225,351-380, src/triple_lock/expected_value.py:91-107,268-348,403-450, paper/index.qmd:82,502-531,553-581,764-804. Trigger: Read the 50,000 paths and £8.411bn gross / £5.392bn net “expected saving” as a well-calibrated predictive joint distribution. Expected: The stated uncertainty should include or bound uncertainty in the macro process, OBR long-term mean path, and data/model response, especially because the policy payoff depends on CPI/earnings reversals and tails. Observed: The primary uses one BIC-selected VAR(1) with 297 fitted residual pairs, residual bootstrap shocks, and a deterministic smooth drift to OBR calendar means. Only 200 sampled paths (198 unique) get full model runs; 40 get paired Microcosm runs. The quoted £0.042bn gross and £0.036bn net SEs measure this sampling only. The older statutory backtest failure has improved but still gives only 1/6 realised gaps inside the primary monthly-shift middle-80% band when April 2022 uses published earnings, versus 4/6 under the legally suspended earnings leg. The paper correctly notes six overlapping origins, dependence on the April-2022 treatment, historical revision, and that no method supports a reliable 80% predictive interval. Its dynamics-tilt sensitivities on the same paths span £6.61bn–£10.66bn gross, with the highest row's net SE about £3.40bn because effective full runs fall to 44. Impact: The numerical “expected” value remains conditional on one macro law and one OBR mean path; the backtests do not support a tightly calibrated predictive expectation of the reform's impact. Carry that conditionality into the abstract/dashboard headline and avoid using the tiny path-sampling SE as the apparent total uncertainty. Consider a distinct model-form/mean-path sensitivity or an explicitly bounded scenario analysis. Evidence: Committed data/results.json expected_value.draws.model, strata, estimates, sensitivities, trajectories.backtest.statutory; paper appendix. These are diagnostics, not proof that an alternative model would give a particular answer.

A3 — Fiscal and distributional estimates hold the pensioner cohort and pension type fixed through 2039 (OPEN)

[MAIN; P2 modeling validation gap, open]. src/triple_lock/engine.py:283-305 pins each person's survey-year state_pension_type, and :449-455 applies it to every run/year. The committed coverage has 8.262m BASIC and 3.873m NEW type people, with survey ages top-coded at 80; the same type mix persists to 2039 while survey weights change. dashboard/src/components/SummaryTab.jsx:221-226 acknowledges the limitation. The cost of a basic versus new flat-rate cut and its household incidence differ, so the £8.41bn gross and £5.39bn net expected savings are estimates for this fixed-composition population, not a projected 2039 pensioner cohort. Comparing with DWP's dynamic £15bn is therefore not like for like. The published results give no ageing/type-mix sensitivity to show how much this changes either number. This is a material uncertainty limitation, not proof that the arithmetic on the frozen sample is wrong.

Suggestions

S1 — The embedded paper disables the scripts its render uses (OPEN)

[PR8; P3 presentation defect]. dashboard/public/paper/index.html:136-139 sets sandbox="allow-same-origin allow-popups allow-popups-to-escape-sandbox" on the iframe, excluding allow-scripts. The rendered manuscript loads Quarto lightbox, anchor, tooltip and tab scripts at dashboard/public/paper/web/index.html:51-62 and initializes the figure lightbox at :2469; figures link to PNGs with class="lightbox" (e.g. :154). Inside the embedded paper, clicking a figure falls back to navigating the iframe to the raw PNG instead of opening its lightbox, and script-driven document controls cannot function. The standalone HTML link avoids the sandbox. Decide whether to allow trusted local scripts in the embed or render the manuscript without script-dependent controls; verify the route in a browser. This does not change model numbers.

Coordinator assessment: The missing allow-scripts token and script-dependent controls are confirmed from the committed files; the interactive effect should be checked in a live browser before treating this as a functional breakage.

Evidence Gaps

• The private Enhanced FRS/Microcosm microdata and pinned PolicyEngine bundle were unavailable, so the 198 full population runs, offsets and high-weight record could not be independently regenerated.
• The official OBR long-term-determinants and DWP benchmark workbook bytes could not be authenticated against the publisher from this environment; committed copies were inspected with their provenance stated.
• The local Python suite could not collect because scipy and hypothesis were absent, and the dashboard build lacked node_modules; GitHub CI success at the reviewed SHA is recorded separately.
• The paper iframe interaction was inferred from markup and not confirmed in a live browser.
• The LaTeX builder added at PR Working paper: Public Pension Indexation Under Uncertainty: Fiscal and Distributional Effects (/paper route) #8 head c9b192c was inspected but could not be executed locally: pandoc and tectonic were unavailable. Its output is not the dashboard's currently committed PDF.

Notes

• Main and PR Working paper: Public Pension Indexation Under Uncertainty: Fiscal and Distributional Effects (/paper route) #8 base have identical file trees; PR Working paper: Public Pension Indexation Under Uncertainty: Fiscal and Distributional Effects (/paper route) #8 changes paper/routing and results provenance metadata, but no numerical results. Latest commit c9b192c adds a LaTeX builder and clearer limitations prose; the scoped delta review found no changed model numbers.
• A proposed council-tax-reduction documentation finding was withdrawn after the latest paper prose clarified its scope to the survey runs.
• The review took over the workflow's 25-minute outer target because the PR head advanced during assembly and the new LaTeX delta needed a scoped check; timing boundaries use recorded UTC progress and file timestamps.
• Prior issue Follow up on full triple lock dashboard and modeling review #2 A12 rounding and A14 effective-path-support findings are resolved/obsolete in the rebuild. A3 is materially addressed by full net/household path runs; A13 backtest concern remains in altered form. DWP primary policy text is now available but exact CPI/AWE series and anchor year remain assumptions.
• No new arithmetic defect was confirmed; no tracked source files were changed.

Validation Summary

At PR #8 head c9b192c, GitHub Dashboard, Pipeline and Vercel checks: SUCCESS. Local read-only checks at 200553e: AST parse of 30 Python files, results JSON copy and 32 paper assets validated; a 100,000-path synthetic uprating diagnostic found zero Burnham rates/levels above triple lock. The 200553e..c9b192c delta was inspected in a detached snapshot: paper prose and a new LaTeX builder only, with unchanged model/results data. A bounded local pytest attempt could not collect due to missing scipy and hypothesis; the new LaTeX build, full population rerun and dashboard build were NOT RUN (missing dependencies/private bundle). Five cached/committed source items passed manifest byte integrity; only two OBR workbooks were publisher-authenticated. Live DWP page and relevant legislation were checked by policy reviewer.

Timing

setup seconds: 103.00s; scope seconds: 269.00s; parallel review seconds: 933.00s; policy role seconds: 857.00s; code role seconds: 933.00s; adjudication seconds: 175.00s; consolidation cleanup seconds: 574.00s; elapsed seconds: 1639.00s

Review Severity

COMMENT. Open findings: 0 critical, 3 should address, 1 suggestions.