OBR published economic effects of policy: AS2023–AB2025 packages, supply-side scorings, March 2026 decisions (external side)

[vahid-ahmadi]

Harvest-first step of #55 (per the 8/16 steer and the claim comment's execution order): the external side of the population the Macro members will answer — no PE counterparts, no engine registration, no emulator work in this PR.

What was retrieved (all Wayback original bytes; obr.uk is Turnstile-guarded, so the uk_obr/NOTES.md no-bypass recipe applies; URLs + snapshots + SHA-256 in sources/obr-policy-effects/raw/README.md, and the adapter sha256-gates every raw file before parsing)

artifact	content harvested
EFO Nov 2023 Ch.2	Chart 2.A — AS2023 package real-GDP impacts by channel (full expensing, NICs cut, welfare reforms, demand)
EFO Mar 2024 Ch.2	Chart 2.A — SB2024 package real-GDP impacts by channel
EFO Oct 2024 Ch.2	Charts A + B — AB2024 package real-GDP impacts by expenditure component AND by measure
EFO Nov 2025 Ch.3	Chart 3.3 (GDP by component) + Chart 3.4 (CPI by measure) — AB2025
EFO Mar 2026 annex tables	Table B.1 — total/direct/indirect effect of Government decisions on borrowing, nested, GBP bn
Briefing Paper No.10 (Nov 2025)	Table 2.1 — every measure OBR has scored for supply-side effects, per-measure potential-output impact (% GDP) by channel, SB2023 → SS2025

Rows: 266, four metrics

gdp_level_effect 151 · cpi_inflation_effect 36 · supply_side_impact 19 · decisions_effect_on_borrowing 60. Every row carries fiscal_event, basis (post_behavioural / supply_side), scope, verbatim source_column, and — for the nested TB.1 — the aggregate_level/parent double-count guard in the obr-welfare adapter's convention, with the published borrowing sign convention carried verbatim rather than normalised.

A free cross-validation surfaced by the spot pins: Briefing Paper 10's AS2023 employee-NICs supply-side scoring (0.16741543…) is bit-identical to the Nov 2023 Chart 2.A terminal-year NICs value — the inventory restates the original scorings at full precision.

Tests

tests/test_obr_policy_effects_adapter.py: block counts, closed row shape, unit-per-metric, five spot values traced to published cells, TB.1 hierarchy identities (direct+indirect=total, subtotals=Σcomponents, £1 tolerance), byte-stable rebuild from the committed raw bytes. Full suite 248 passed / 6 skipped; ruff clean.

Deferred, honestly

Annex A workbooks for five vintages were fetched, inspected, and NOT vendored (forecast levels only, no policy content — documented in raw/README.md); likewise the Mar 2025 / Nov 2025 / Mar 2026 Chapter 2 and Nov 2025 Ch.4/5 workbooks. Spring 2025/2026 package-effect chart data was not located in this pass; Mar 2026 decisions are covered in aggregate by TB.1. DB ingest waits for the post-#57 ingest shape; counterparts (PE-UK LSR, emulator) are #55 steps 2–3.

Builds #55 (step 1: harvest).

🤖 Generated with Claude Code

[MaxGhenis]

Gate round 1 (reviewed as a pair with #81) — strong vendoring; four findings, three blocking

Clean and verified: all six SHA-256 pins match the vendored artifacts, parsing reads stored XLSX values at full precision (the office:value discipline), a no-write regeneration was byte-identical at 30+20+36+30+35+36+19+60 = 266, and every harvested TB.1 cell is a forecast — no Ledger-boundary violations.

1. Blocking — macro units are semantically conflated. CPI effects are published in percentage points; BP10 supply-side impacts are percent of GDP; both emit as generic percent (adapter.py:247, :300), and the adapter test entrenches the conflation. These are three different quantities (pp deviation, % of GDP, and — in OBR policy effects → DB ingest: 266 claims, four macro-effect metrics #81's mapping — GDP-level % deviation) that a query must never treat as one; each needs a deliberate unit concept, same rule that split GBP_PER_WEEK from bare GBP.

2. Blocking — 266 is not source-reconciled. TB.1 carries 66 numeric cells across eleven lines; TB1_ROWS selects ten and emits 60 — the six memo current-budget-deficit cells are silently skipped (adapter.py:381). Choosing four metrics is defensible; the drop has to be tallied and explained (the 272 candidates → 266 claims + 6 deliberate drops accounting), never silent.

3. Blocking — baseline/prose provenance is too coarse. source.json blanket-labels rows as "against each EFO's pre-measures forecast," but rows carry no baseline descriptor or locator — and BP10 itself distinguishes tax/welfare counterfactuals from DEL/regulatory activity baselines, while the WCA reversal scores against the earlier WCA-adjusted world. Per-row registered baseline worlds (the pre_obbba_law / hmrc_indexed pattern) with vendored/annotated supporting pages.

4. Identity vocabulary fail-open. _slug() mints program/subgroup/type identities without a closed source map (unknowns should raise, per uk_aliases), and basis is reused for post_behavioural|supply_side when the standard axis means forecast/outturn — these rows are all forecasts; scoring method needs its own registered axis.

🤖 Generated with Claude Code

[vahid-ahmadi]

All four addressed in ea37f62.

1. Units. GDP-level deviations, CPI effects and supply-side impacts were three quantities under one percent. They now carry percent_of_real_gdp, percentage_points and percent_of_potential_gdp, with gbp_nominal unchanged on Table B.1 — the same rule that split GBP_PER_WEEK from bare GBP. The adapter test pins the mapping per metric instead of entrenching the conflation.

2. 266 is source-reconciled. You were right that the memo line fell through a silent continue. Every value-bearing TB.1 line is now classified — emitted (TB1_ROWS) or a declared drop (TB1_DROPS, with a reason) — and an unclassified value line raises. The build reports and asserts 272 source cells = 266 claims + 6 deliberate drops; the memo row is dropped because it scores the current budget, not PSNB.

3. Baselines per row. Each row carries its own round's pre-measures world (obr_pre_measures_<event>), the counterfactual KIND — BP10 ch.2 splits tax/welfare legislated-parameter counterfactuals from DEL/regulatory activity baselines — and a locator. March 2026 Table B.1 keys obr_november_2025_budget_forecast, its own stated counterfactual, and the March 2025 WCA reversal's locator records that it scores against the WCA-adjusted world AS2023 created.

4. Identity vocabulary + basis. Chart series labels, BP10 measures, types and channels are explicit registries; an unregistered label raises instead of _slug() minting one. basis is now forecast (its repo-wide meaning) and the scoring method moved to its own scoring_method axis.

Also: publication provenance is per artifact, which is the input #81's date fix consumes.

266 claims unchanged; suite 261 passed, ruff format clean.

[DTrim99]

Reviewed at head ea37f62 — ran the suite (13 pass) and recomputed the SHA pins (match). This is a clean, disciplined harvest; same rigor as the other lanes.

Verified:

• Provenance — all 6 OBR workbooks are SHA-256 pinned in adapter.py, source.json, and the README, with a recompute-and-compare gate that runs first in build(); editions are pinned by dated snapshot URLs + per-artifact date (with the Wayback-capture-≠-publication-date distinction asserted), not a latest URL.
• Adapter parses by sheet name + title anchor + a closed series registry (not fragile absolute cell refs), and values pass through verbatim — the only transform is the TB.1 £bn→nominal *1e9 (I checked the raw→nominal value matches the pin). Four distinct unit concepts stop %-vs-£bn / pp-vs-% misfiling.
• Fail-loud on any unmapped series/measure/type/channel, title-prefix mismatch, non-FY period, or unclassified TB.1 line — all tested.
• Accounting reconciles (272 cells = 266 claims + 6 tallied memo drops) with a TB.1 tie-back and hierarchy identities (direct+indirect=total, £1 tol), and determinism holds (sorted iteration, no timestamps, byte-stable rebuild test).

Only cosmetic nits: a hardcoded header_row=25 and a deliberately-truncated TB.1 title prefix are mildly fragile, but the title-anchor + closed-registry + FY guard would catch any real re-shuffle. Approving — merge this before the stacked ingest (#81).