[vahid-ahmadi]

Update, 15 Oct: the cross-check now runs on Microcosm's UK release

The 40 paired runs in the 14 Oct update used populace_uk_2023, the June file the repo has labelled "Microcosm". They now use Microcosm's UK national release: microcosm_uk_2024_25.h5 at microcosm-uk-2024-25-national-20261002T230158Z-5c6b3f68 (SHA-256 aa31bdf6…). Everything else stays on the certified 6.2.6 bundle and Enhanced FRS 1.58.0.

• Registry (datasets.py, config.py). Microcosm is pinned by tag and SHA-256 and is uncertified under 6.2.6, because its release declares compatibility with policyengine-uk 2.100.0 only. Its ageing anchor is 2025, since all 1,168 of its calibration targets have period 2025. populace_uk_2023 stays registered for the pilot scripts.
• Two fixes in demography.py, please review. Microcosm's person, household and benefit-unit IDs are 64-bit, and policyengine-uk holds them as 32-bit, so they wrap. The order is the same, and no IDs collide after wrapping.
  • The record-order guard now compares IDs in the model's integer type and requires them to stay distinct there (_same_order, with a test).
  • The birthday draw now hashes the IDs the model holds. Hashing the survey's 64-bit IDs gave other birthdays than policyengine-uk's, and 495 records changed State Pension type, so EligibilityChanged refused the run. With the model's IDs, the draw matches policyengine-uk's months_since_last_birthday for every Microcosm record, and it is unchanged on the Enhanced FRS (whose IDs fit in 32 bits).
• Full rebuild at e089a0f (the cache keys on these files). The Enhanced FRS figures are identical to the 14 Oct build.

Results, 2039-40 (£bn)

	Gross	Net
Expected saving, Enhanced FRS (headline, unchanged)	11.1 ± 0.1	7.9 ± 0.1
The same 40 paths on Microcosm	12.3	8.4
Microcosm − Enhanced FRS, paired	+1.12 (SE 0.01)	+0.49 (SE 0.01)
Previously, on populace_uk_2023	+0.73	+0.13

In 2034-35 the paired difference is +0.31 gross and +0.13 net.

• Checks:
  • pytest: 1,686 passed. The C2 fixture and the dashboard tests now name microcosm_uk_2024_25.
  • check_assumptions.py passes, and the dashboard test, lint and build pass.
  • No record-level keys in the file; provenance shows e089a0f with git_dirty false.
• Next: once Guard model-v2 binding inputs and build admission #29 lands, rebase this onto model-v2; then merge main into model-v2 and send model-v2: certified bundle, engine port and guarded rehearsal #24 to main.

---

Update, 14 Oct: Max's answers applied

• Provisional under ruling (c), not (b). triple-lock-build --pre-budget --uncertainty-ruling c runs the primary through the model-conditional route, then checks that its path design and specs, and both mean-path specs, are byte-identical to the committed C2 dry run (data/pilot/c2-dry-run). It refuses if they differ or if the dry-run primary fails C2. The file records the dry run's C2 provenance (run_kind: "dry_run", expected_value_authorized: false, provisional: true).
  • I re-ran the C2 dry run on the current code. It reproduces exactly: the same outcome and score table (test A suspended: 4 of 6 terminal, 80% annual) and byte-identical paths.
• Enhanced FRS 1.58.0 stays primary. The 40 paired Microcosm runs are done (about 8.5 minutes each on my Mac) and included, and the Microcosm skip no longer applies to this release.
• The Pre-Budget certified release: rebuild on policyengine 6.2.6, model-conditional expected value, dashboard #30 dashboard layout, with:
  • the "provisional" banner and card;
  • the dry-run sentence on the Methodology tab;
  • the ±0.5pp cards kept as sensitivities;
  • no "failed backtest" wording.
• Results (2039-40, unchanged): expected saving £11.1bn ± 0.1 gross, £7.9bn ± 0.1 net. Microcosm − Enhanced FRS on the same 40 paths: +£0.73bn gross (SE 0.01), +£0.13bn net (SE 0.02).
• Checks:
  • pytest: 1,684 passed, 0 failed after one stale test fix. test_ageing_validation pinned the pilot's paired draws; it now reads the committed file's draws, and the pilot pins have their own test.
  • check_assumptions.py passes, and the dashboard test, lint and build pass.
  • No record-level keys in the file.
• Next: once Guard model-v2 binding inputs and build admission #29 lands, rebase this onto model-v2; then merge main into model-v2 and send model-v2: certified bundle, engine port and guarded rehearsal #24 to main.

The original description follows. Where it differs from the update above, the update applies.

---

@MaxGhenis this PR collects all of our pre-Budget work on top of model-v2 (merged up to bf0e6ee): a certified rebuild, its results and the dashboard changes. Nothing is merged. Below: what we did, the decisions that are yours, and the runs still to do.

Supersedes #27 and #28; this branch contains both, so they can be closed once this one is settled.

What we did

• Certified rebuild on my Mac. Built with policyengine 6.2.6 (pe-uk 2.125.1, core 3.33.0) on Enhanced FRS 1.58.0. The bundle and Housing Benefit passport preflights both passed. The build ran from a clean tree.
  • 161 primary-path runs, plus 322 earnings ±0.5pp mean-path runs, the central path, the OBR wedge, the history replays and Microcosm coverage.
  • Inputs: OBR March 2026, with August 2026 CPI as April 2027's CPI leg (provisional).
• model-v2: certified bundle, engine port and guarded rehearsal #24's CI fix (from Fix #24 CI: render the HB passport footnote from the renderer #27). The renderer now renders the HB passport footnote. Your bf0e6ee fixed the same files, so the merge keeps your version.
• --pre-budget release mode (from Pre-Budget release mode: scenarios-only certified build, clearly labelled #28):
  • writes a release block and shows a dashboard banner;
  • records your ruling (c) and says it is "not a new d955 ruling";
  • certified provenance is checked against the recorded bundle;
  • an end-to-end assembly test covers the pre-Budget path.
• Ruling (b) under --pre-budget: shows a model-conditional expected value now. The banner says the primary failed C1, and that the binding C2 score and a new release follow the 28 Oct Budget.
• Microcosm paired subsample skipped in this release only (pre_budget and ruling == "b"). It is recorded as expected_value.paired_sensitivity.status: "omitted". The normal and binding builds are unchanged, and a test covers that.
• Dashboard:
  • Budget impact leads with the distribution across paths, labelled model-conditional, beside the scenario envelope (cards plus a new dot chart, not a distribution);
  • the scenario section is no longer duplicated;
  • no "unavailable" boxes;
  • expected-value wording kept out of the scenarios-only view.
• Tests follow the file. Historical publication pins now come from archived pilot bytes (tests/fixtures/historical-publication/). Assumptions tests follow the recorded population treatment.
• Checks:
  • full pytest: 1,682 passed, 0 failed;
  • check_assumptions.py: passes;
  • dashboard: test, lint and build pass;
  • no record ids, weights, single-record figures or counts of 1–9;
  • Burnham ≤ triple lock on every path, and net ≤ gross.

Results (UK, £bn, 2039-40, gross / net)

	Gross	Net
Expected saving, model-conditional (160 paths)	11.1 ± 0.1	7.9 ± 0.1
Middle 80% of paths	3.9 to 18.9	2.7 to 13.4
Central forecast	0.9	0.6
Triple lock at the OBR's long-term rate	10.3	7.3
Earnings −0.5pp from 2031	10.4 ± 0.5	7.4 ± 0.4
Earnings +0.5pp from 2031	11.0 ± 0.5	7.7 ± 0.4

Against the uncertified pilot (E both, kept), central gross is −1.3% and net is +7.3%. Most of that comes from the HB passport fix and the 1.58.0 Housing Benefit and Pension Credit data. GB coverage against DWP: pension-age HB is now +3–11% (pilot +71–86%), and PC claims +13–40% (pilot +30–63%). The State Pension bill is about 15% below DWP.

Decisions for you

Method

• Publishing (b) before the Budget. Your recorded ruling is (c): an expected value only if the primary passes C2 on the 28 Oct inputs. This PR shows a model-conditional expected value now, clearly labelled, and a new release follows the Budget. OK, or should the pre-Budget release be scenarios only (ruling a, which this branch also supports)?
• Skipping the Microcosm cross-check in this interim release. OK, or run the 40 paired runs first (see below)?
• Microcosm or Enhanced FRS as the primary dataset. Should the published numbers be on Microcosm rather than the Enhanced FRS? Our understanding is that 6.2.6 certifies Enhanced FRS 1.58.0 and REBUILD treats Microcosm as an uncertified sensitivity, so we kept the Enhanced FRS. If PolicyEngine is moving to Microcosm as the default, the primary would switch, at about 32 GB and 20 minutes per run (roughly 54 hours for 161 runs on one worker).

Dashboard

• Which dashboard ships: this branch's Budget impact tab (distribution plus envelope plus dot chart), or your model-v2 layout. With this file, your getExpectedValue returns null because it requires the Microcosm sensitivity estimates, so the summary and the expected saving show "unavailable". With the scenarios-only file, it renders the envelope block twice.
• The earnings ±0.5pp cards are path-set averages, so they sit close to the expected value. Keep them, or show only the central and OBR-wedge scenarios?
• Banner wording: the "28 October" date is in the component and in pipeline constants.

Code

• Review the narrow pre_budget and ruling == "b" Microcosm skip, and the release block in RESULTS_SCHEMA.md and REBUILD.md.
• Merge route: this PR into model-v2, then model-v2: certified bundle, engine port and guarded rehearsal #24 to main. Close Fix #24 CI: render the HB passport footnote from the renderer #27 and Pre-Budget release mode: scenarios-only certified build, clearly labelled #28.
• The git-history scrub of the single-record diagnostics (Stop publishing single-record diagnostics #26 removed them from main) is still your call.

Runs still to do

• 40 Microcosm paired runs for the dataset cross-check (about 20 minutes and 32 GB each; about 13 hours on one worker). Every other job is cached. After that, rebuild with ruling (b) and the cross-check, or with (c) after the Budget.
• Binding C2 score on the 28 Oct inputs: September CPI (21 Oct), May–July AWE, and the OBR Budget means. Then the certified rebuild under (c), with either an expected value beside the envelope or scenarios only.
• The version bridge split (bundle → data → passport) for the pilot-to-certified change, if your Part J import doesn't already cover this head.

🤖 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 15, 2026 2:47pm UTC

[MaxGhenis]

UK hub FYI for the pre-Budget release: the uk-data age top code, and a change coming in the next uk-data batch.

• The data: uk-data 1.58.1 (enhanced_frs_2024_25.h5, sha256 056e80fd…) has no record over 80. FRS adult ages are top-coded at 80, so 7,718 person records (3.59m people at 2024 weights) sit at exactly 80.
• This repo already handles it. src/triple_lock/demography.py gives top-coded records a represented age from 80 to 105+ by a deterministic draw on the anchor year's ONS single-age shares, in every treatment except legacy. So the projections do depend on ages over 80, but through that stated method.
  • Check that the published text says the over-80 age split comes from ONS shares, not from the survey.
  • Check that any legacy-treatment figures don't appear without that caveat.
• Coming in the next uk-data batch (Max's go, before 29 Oct): the national and regional ONS age targets will merge their top bands into one open-ended 80+ band per sex, because the data can't meet the 85–90 targets (uk-data #593). That changes the calibrated 80+ weights, which this repo's anchor-year weights inherit. A release certified on 1.58.1 stays valid as it is, but a rebuild on the next data release will move the 80+ cells.