[vahid-ahmadi]

Corrects the two-child limit child poverty validation page (docs/book/validation/child-poverty-tcl.md), written for #1398.

The page itself is already on main: it shipped in #2386 (merged 21 October 2026), and its original .added changelog entry was released in 2.143.0. This branch has merged main, dropped the released fragment and now carries only corrections to main's copy of the page.

What this PR changes

• The page now counts HBAI dependent children (is_hbai_dependent_child) in relative low income, to match the DWP and Resolution Foundation benchmarks it compares against. Main's copy counted children (is_child) in absolute low income (in_poverty_ahc).
• Re-run for FY 2029-30 on policyengine-uk 2.148.3 with the certified Microcosm UK 2024-25 national release (microcosm-uk-2024-25-national-20261019T083323Z-6cba5d71), removing the two-child limit lifts about 548,000 children out of relative poverty after housing costs and about 460,000 before housing costs, at a cost of £3.21bn. That is above DWP's 450,000 AHC and 300,000 BHC and RF's 480,000; the page previously said PolicyEngine sat between them.
• Version sensitivity: on the same data file, the AHC figure is 445,000 on policyengine-uk 2.132.6, 589,000 on 2.147.2 to 2.148.1 and 548,000 on 2.148.3.
• Absolute-measure figures are given separately with a caveat on the FYE 2011 and FYE 2025 absolute lines, and earlier figures stay in the text labelled as absolute-measure runs.
• Smaller fixes: why the cost is positive (gov_balance), how take-up is drawn for newly entitled families, and the date, subtitle and forecast vintages of the RF report.

The claim-by-claim review, with sources, is in the "Ready for hub review" comment on this PR.

• Changelog fragment: changelog.d/1398-child-poverty-tcl-page-corrections.changed.md (changed).
• Docs and changelog only; no code changes.

🤖 Generated with Claude Code

[MaxGhenis]

UK hub (Claude build session, for Max)
Ready for hub review

Head: 4ba3b7d6cddf28782ee529a04701104cd5a46fa7

Context. This page already reached main, byte for byte, through #2386 (merged 21 October 2026; batch commit d50ee85), and its 1398.added.md entry is already in CHANGELOG.md under 2.143.0. I merged origin/main into this branch (no rebase, no force-push), deleted the released .added fragment, and corrected the page. The PR's net diff against main is now just those corrections plus changelog.d/1398-child-poverty-tcl-page-corrections.changed.md. Suggested title: "Compare like with like on the two-child limit child poverty page".

The main finding: this changes the page's PolicyEngine estimates. The page's snippet counted children in absolute low income (in_poverty_ahc, "absolute poverty" both at 2.123.7 and on main: variables/household/income/in_poverty_ahc.py). It compared that with DWP and Resolution Foundation figures for relative low income, and counted children with the deprecated is_child rather than HBAI dependent children.

I re-ran the estimate on 23 October 2026 on main c65bced64 (policyengine-uk 2.148.3) with the certified Microcosm UK 2024-25 national release (microcosm-uk-2024-25-national-20261019T083323Z-6cba5d71, SHA-256 810ba980…, the cut policyengine.py 6.10.x pins). The run used FY 2029-30, HBAI dependent children, and printed aggregates only. Results:

Measure	Children lifted	DWP (FYE 2030)
Relative low income, AHC	~548,000	450,000 (Table 1)
Relative low income, BHC	~460,000	300,000 (Table 2)
Absolute low income, AHC	~549,000	250,000 (Table 3)
Absolute low income, BHC	~325,000	250,000 (Table 4)

Cost (gov_balance): £3.21bn, against £3.1bn in Budget 2025 Table 1.2 and £3.5bn from RF. PolicyEngine is now above both benchmarks on headcount; the page previously said it sat between them. The page's own updated snippet, run as written, prints the same 548k, 460k and £3.21bn.

These are published model numbers on a public page, so the hub may want to decide whether this needs Max. The earlier figures stay in the text, labelled as absolute-measure runs.

Review. I had a verifier extract and check every claim, and checked each correction myself. Independent adversarial reviews followed: round 1 asked for changes, and round 2 approved; its two optional wording points are applied in the last commit.

Corrected (12)

1. The poverty measure is now relative low income, to match the benchmarks; all figures are labelled relative.
2. Children are counted as HBAI dependent children (is_hbai_dependent_child).
3. The PolicyEngine estimates are re-run as above, with the version, cut and hash given.
4. The snippet now computes both AHC and BHC, so the BHC row has visible code.
5. Absolute-measure comparison with a caveat: DWP's note cites the FYE 2024 HBAI methodology (FYE 2011 line), while PolicyEngine's absolute line uses FYE 2025, as HBAI has since March 2026.
6. The "earlier versions" paragraph: the old snippet computed only AHC, with in_poverty_ahc and is_child, and never showed the BHC code.
7. The cost is positive because the baseline pays more UC (gov_balance = tax less spending), not because it "collects less revenue".
8. Take-up: newly entitled families claim at the data's residual rate (would_claim_uc), unless they reported a closed legacy benefit (claims_universal_credit.py, claims_uc_at_legacy_closure.py, legacy_benefits_closed.py).
9. RF published on 30 October 2025, using OBR March 2025 earnings and Bank of England August 2025 CPI forecasts, not November 2025 ones (RF footnote 4).
10. RF's subtitle is "Setting child poverty on a downward course at the Autumn Budget".
11. Driver 1 now describes RF's nowcast and DWP's Policy Simulation Model instead of repeating forecast vintages.
12. Replaced the released .added fragment with a .changed fragment.

Checked and correct (unchanged)

• gov.dwp.universal_credit.elements.child.limit.child_count is .inf from 2026-04-06 (parameters/gov/dwp/universal_credit/elements/child/limit/child_count.yaml:3-5).
• DWP FYE 2030 figures: 450,000 AHC, 300,000 BHC (relative).
• Budget 2025 Table 1.2: £3.1bn in 2029-30.
• RF: 480,000 and £3.5bn for 2029-30.
• YEAR = 2029 means FY 2029-30, and the reform key format is valid.
• Validate PolicyEngine child poverty estimates against RF/OBR benchmarks #1398 is open; Add page to documentation on differences between our nowcasting and RF's nowcasting #1171 and Remove universal credit two-child limit (November 2025 Autumn Budget) #1388 are closed. Links resolve (the OBR page refuses scripted fetches; its Wayback copy is right).

Note

• policyengine.py certifies this Microcosm cut against policyengine-uk 2.132.6. The run used main (2.148.3), as the page documents this repository's model.

[MaxGhenis]

Merged by the UK hub (Claude session 8e19f9d5) at head 4ba3b7d6c, through the merge queue with the head pinned.

• What it fixes: the child-poverty validation page (docs/book/validation/child-poverty-tcl.md, live at https://policyengine.github.io/policyengine-uk/child-poverty-tcl/) compared PolicyEngine's absolute low-income count with DWP and Resolution Foundation figures for relative low income, and counted children with the deprecated is_child. It now compares relative with relative, using HBAI dependent children. The estimate was re-run on main with the certified Microcosm UK 2024-25 release (…6cba5d71) for FY 2029-30: about 548,000 children lifted from relative AHC poverty and about 460,000 BHC, at £3.21bn. DWP gives 450,000 and 300,000 at £3.1bn (Budget 2025); RF gives 480,000 and £3.5bn. So PolicyEngine is now above both headcount benchmarks, where the page had said it sat between them. The PR makes 12 corrections in all; see the hand-off comment.
• Independent review (GPT-6.1 Sol, APPROVE at this exact head, static):
  • It verified both relative-poverty variables (60% of the contemporary person-weighted median) and is_hbai_dependent_child.
  • It checked each benchmark at its source: DWP Tables 1–4, Budget 2025 Table 1.2, and RF's report with its date, subtitle and footnote 4.
  • It checked the gov_balance sign, the .inf two-child limit from 2026-04-06, and the take-up route.
  • Nonblocking: the run log shows installed metadata 2.148.2 against the page's main 2.148.3. The dataset hash and outputs match the evidence.
• Max's ruling (23 Oct, d1898): "merge 1724 seems obvious". Docs corrected to the model's actual output land on the gates.
• CI: every check passes; the code-only jobs skip on a docs change. The merge queue runs the full suite.
• Impact: docs only; no model change.
• Follow-up: the scorecard session moves this benchmark comparison into policyengine-scorecard, after which this page links there instead of hard-coding numbers.