Changelog

Notable revisions

Corrections that moved a figure the site had already published. Each links to the merged change so the before-and-after can be checked against the diff. Routine weekly data refreshes are not listed here — they are in the commit history of data/pmg_snapshot.json.

1. 2026-09-09·PR #230

   Gini coefficient 63.0 -> 54.1, and seven indicator series moved to live sources

   The first run of scripts/build_indicators.py --fetch --no-csv replaced seven of the sixteen indicator series with what the IMF WEO DataMapper and World Bank WDI APIs return, each stamped with retrieved_at and a source_url. GDP growth matched the hardcoded series to the decimal for 2014-2025. The Gini series did not: the processed CSV carried 63.0 for 2015, 2018, 2019 and 2022, all carry-forwards of the World Bank's old 2014 figure, while the World Bank's consumption-based series now reads 54.1 for 2022/23 (Income and Expenditure Survey) and 59.6 for 2014/15 (revised from 63.0). The series is now the seven survey years the Poverty and Inequality Platform holds, and the glossary entry that said 'approximately 0.63' says what the series says. Two other published points moved: debt-to-GDP for 2025, 77.3 (Budget Review, fiscal year) to 78.6 (IMF general government, calendar year), and the tariff series now ends at 2022 = 7.31 rather than 2024 = 6.8, because WDI TM.TAX.MRCH.SM.AR.ZS stops there and its overlapping values show the CSV was a different series.

   Figures affected: /indicators (Gini, tariff, debt, fiscal balance, investment, life expectancy, GDP growth) · Idea pages: 'Where the numbers stand' under fiscal_space and trade_openness · /glossary Gini coefficient · Outcome briefs and Ask answers that cite an indicator

2. 2026-08-21·PR #112

   Package aggregates re-derived from the live idea set

   data/reform_packages.json carried idea counts written once and never regenerated: 29/18/42/23/27 = 139 ideas, while the database held 51/26/30/24/40 = 171. Every consumer printed the stale number — package cards, package detail headers, the brief and PPTX exports, the public /api/v1/packages response, and the Reform Index, which divides by idea_count. The index denominators were therefore wrong and every sub-index with them. avg_feasibility, avg_growth_impact and the two status counts had rotted the same way, and idea_ids still named ideas since reassigned to other packages. Fixed structurally rather than by refreshing the literals: the JSON keeps only editorial fields and every countable field is recomputed from the live rows at load time, in both backends, through one shared module. A stored count cannot drift if nothing reads it.

   Figures affected: Reform Index score and all five sub-indices · /packages cards and detail headers · /api/v1/packages · Brief and PPTX exports

3. 2026-08-21·PR #112

   The Reform Index stopped dating itself from a file mtime

   lib/reform-index.ts set last_updated from fs.statSync(dataPath).mtime — filesystem metadata that the deploy pipeline rewrites. The flagship page dated itself 2018-10-20, eight years stale, next to figures it recomputes hourly. The date now comes from the idea data's own MAX(updated_at) and falls back to the newest quarterly snapshot's quarter start. Both are properties of the data; neither is filesystem metadata. The data_as_of and data_version stamps introduced across the site derive from the same rule.

   Figures affected: /reform-index · Site footer stamp · API meta blocks

4. 2026-08-21·PR #117

   Research abstracts repaired at the extractor

   37 of 298 research_papers rows stored a site footer or an author bio as part of the abstract — one row ended in “back to top Join our newsletter … SIGN UP Category Search Search Vacancies”. The cause was in scripts/enrich_paper_abstracts.py: it read a fixed 8,000-character window of HTML from the article container, and a window is a budget rather than a boundary, so on any short article it kept reading into the page chrome. Fixed in the extractor and the affected rows repaired, rather than patched row by row.

   Figures affected: /research · /papers detail pages · Research link judging

5. 2026-08-20·PR #104

   All 55 international comparisons enriched

   Before this, 14 of the 55 comparison rows carried approach, timeline and lessons_for_sa (from migration 008) and 41 did not — a standing validation warning and a visible hole on the idea pages, which render the comparison block from exactly those fields. The 41 were written to the same standard: every claim grounded in the row's own outcome summary or a well-established fact about that documented episode, no new quantitative claim introduced, and lessons written against the linked idea's subject rather than in general terms.

   Figures affected: /comparisons · Comparison blocks on idea pages

6. 2026-08-20·PR #106

   Internal record references stripped from reader-facing prose

   Internal identifiers printed verbatim in published text — “the Energy Bounce-Back scheme (id=20)” — across 21 ideas' descriptions and feasibility notes, including one multi-id form. The prose always names the reform it refers to and the relationships live in the dependency graph, so removing the identifiers lost nothing. A hygiene rule now covers idea and comparison prose fields and fails the validation gate.

   Figures affected: /ideas detail pages · /comparisons

7. 2026-08-20·PR #103

   Per-committee coverage published: ingest separated from linkage

   An absence of recent meetings for a committee had no published explanation, so it read as an absence of parliamentary activity. The /data panel now distinguishes four states — the record has stopped arriving, meetings arrive but none is linked to an idea, older meetings are linked but no recent one is, and both current — computed against the snapshot's own generation date rather than against the time the page was built.

   Figures affected: /data

Curated dataset milestones before May 2026 — new data domains, taxonomy changes, bulk editorial passes — are kept in data/CHANGELOG.md.