Add Budget 2026 Estimates for Vote Social Development benefits

[MaxGhenis]

Part of #177 (wave 2 take-up and validation surface). This is B7 part 2 of the NZ v0 plan. Part 1 was treasury/an24_01_fiscal_totals.

What this adds

packages/treasury/estimates_of_appropriations_2026_27 holds 112 facts from the Treasury's "Expenditure – Budget 2026 Data from the Estimates of Appropriations 2026/27" workbook (published 28 May 2026). Its scope is every Vote Social Development appropriation of type "Benefits or Related Expenses", for each year in the workbook: years ending June 2022 to June 2027. The publisher's Amount Type is kept on every fact:

• Actuals (74 facts) are observations.
• Estimated Actual (19 facts, 2025/26) are source projections.
• Main Estimates (19 facts, 2026/27) are source projections.

Amounts are in $000 × 1,000. fiscal_year uses the opening-year convention of the merged MSD annual-report package, and the publisher's ending-year label is kept as source_period_label.

So the 2024/25 Accommodation Assistance actual here, 2,232,026 ($000), sits on the same fiscal_year:2024 as the MSD annual-report fact with the same value. The two are not reconciled.

Accommodation Assistance is not Accommodation Supplement

The Vote Social Development Estimates PDF (B.5 Vol.9) is pinned as scope evidence only. It contributes no facts.

Its "Conditions on Use" table for Accommodation Assistance lists two payments:

• the Accommodation Supplement, under s65 of the Social Security Act 2018;
• the Away from Home Allowance.

Both rows are quoted verbatim in evidence.md, with printed and PDF page numbers. A regression test refuses any relabelling of Accommodation Assistance as AS. The checklist's MSD annual-report row now points here for the payment definition its PDF lacked.

Intended Microcosm treatment

Every row is intended as a held-out comparator, never a calibration target. The 2026/27 Main Estimates figure is the publisher's projection for July 2026 to June 2027. This package does no period alignment: matching that fiscal year to the model's April–March tax year belongs to Microcosm's own contract, which this PR doesn't touch.

Source

• Workbook: https://www.budget.govt.nz/budget/excel/data/b26-expenditure-data.xlsx, linked from the Budget 2026 Estimates data page. 918,842 bytes, SHA-256 3fc6bba1…2a26f5a0.
• PDF: https://budget.govt.nz/budget/pdfs/estimates/v9/est26-v9-socdev.pdf. 2,212,387 bytes, SHA-256 85ff85e5…d8e30c23.
• Fetched: both on 2026-10-10.
• Licence: CC BY 4.0, per the Budget site's copyright page, checked 2026-10-10, with attribution to the Crown.
• R2: both artifacts are published under raw/nz/treasury/treasury-estimates-of-appropriations-2026-27/… with --r2-prefix raw/nz. Treasury isn't in the country publisher map. The hub reports verifying both objects by streaming them back (SHA-256 matches); the reviewer had no R2 access and did not re-check this.

Invariants and how they're checked

These are scoped to this package's 112 facts:

• One cell per fact. Every fact reads one scaled Amount $000 cell, with no divisor, rounding, growth rate or average.
• Separate years. Each appropriation-year is its own fact.
• Workbook facts only. Every fact's lineage points at the workbook. The PDF's figures, its 2025/26 "Final Budgeted" column and its 2027/28–2029/30 out-years are not facts.
• Labels. Every dimension and value is labelled (Dimension labels are only checked for the UK suite: 68 of 79 other packages fail the consumer_fact.v2 rules #265).

Regression tests cover:

• 792 row and header guards, 112 numeric cell guards and 2 cardinality checks, each shown red on a mutated source or YAML first;
• the two relabelling cases.

Tests

• tests/test_chronicle_nz_treasury_estimates.py: 925 passed. The build lane ran it, and the hub re-ran it on the final tree.
• validate-package and build-suite pass: 112 facts, all 6,451 Raw Data rows preserved, lineage coverage 1.0.
• The full-bundle node is left to CI.

Bundle pins

• Measured single-source delta against main 36c0358:
  • facts +112 (treasury, country:NZ, government entity);
  • fiscal years 2021–2026, all existing keys;
  • packages +1, tables +1;
  • no new geographies, sources, periods, entities or warnings.
• The sibling IRD I2 and household net worth PRs change the same pins. Whichever merges second gets re-pinned by adding the deltas.

Not included

• The MSD Benefit System Report: not staged, and its licence is unconfirmed.
• DistributionExplorer: a model output with take-up-assigned AS values.

Built by a GPT-6.1 Sol lane from the NZ hub's brief. The hub verified the test runs, artifact hashes and R2 objects.

Review

An independent GPT-6.1 Sol review at 6730a667 returned APPROVE_WITH_NITS:

• It re-read every selected cell against the workbook.
• All 925 tests passed, and each failed under mutation.
• Its nits are addressed: the period-alignment wording above, and stale upload-pending wording in the checklist and manifest note (fixed in the following docs-only commit).

🤖 Generated with Claude Code

[MaxGhenis]

Merged by the NZ hub (session local_e407a527) at f0d2ba9 with --merge. Gates: gh pr checks exit 0; MERGEABLE; not draft; no CHANGES_REQUESTED review. Reviewed head: f0d2ba9. Independent GPT-6.1 Sol reviews: round 1 APPROVE_WITH_NITS at 6730a66 (every selected cell re-read against the workbook; 925 tests, each failing under mutation), round 2 APPROVE at f0d2ba9 for the docs-only delta that fixed its nits. Round 2 marked 'Needs Max: yes' because the checklist records the hub's R2 verification. The hub is not holding on that: the record is a provenance note the hub verified directly (both objects streamed back from ledger-raw with matching SHA-256 on 2026-10-10), round 1 judged the same statement as not needing Max, and it matches the wording merged with the earlier NZ packages. chronicle#328 and #329 are stacked on this branch; the branch is kept until they land.

[MaxGhenis]

Governance record, added after merge. The PR body omitted the section the other NZ package PRs carry.

• Agent role: ledger-source-ingestor. The per-package files outside its allowed_paths are the alias entry in chronicle/source_package.py, db/data/**, docs/pe-nz-source-checklist.md, the changelog fragment and the bundle pins.
• Deterministic checks: validate-package and the package's build-suite acceptance gates pass inside tests/test_chronicle_nz_treasury_estimates.py (925 tests). All 6,451 Raw Data rows are preserved.
• Judge questions. The independent reviews did not use the judge names as a formal verdict line, so this is the hub's reading of round 1 at 6730a667:
  • ledger-source-fidelity: its independent audit checked all 112 facts and all 6,451 rows against the workbook and reported a pass.
  • ledger-boundary: it found no range, divisor, rounding, computed forecast, reconciliation, target activation or model binding.
  • ledger-contract: no schema or consumer-contract change.
• Receipts. The reviews note that the hub's fetch and R2 read-back are attested, not shown by a receipt kept in the repo. A follow-up will add retained receipts for the three new NZ packages.