[vahid-ahmadi]

Summary

This PR redesigns the whole site in the style of the Autumn Budget 2026 and tourism dashboards: a dark teal header band with four views, a teal title band on every page, a sticky white sub-tab row, then white cards on a grey canvas. Each card has a short title, a one-line lede, the figure or table, and the long explanation behind an expander. The two-sided step panel (numbered steps on the left, sticky figures on the right) remains the main layout for model Overview, Methodology, Validation and Code tabs. DESIGN.md is the binding spec. URLs, numbers and data are unchanged; generated pages were changed through their generators and regenerated.

By page type

• Shell (all 110 pages). The header band has Home, Models, Forecasts and Use, and the chosen view is shown by fill and underline. Every page except home has a pathway line and a title band with the h1, a one-line subtitle and a lede of at most 40 words. Below that is a sticky sub-tab row and a one-row footer (PolicyEngine, GitHub, Contact). Every sticky offset reads --nav-h. The stylesheet cache buster is now 10.
• Home. An intro card with the hero fan, three entry cards (Models, Forecasts, Use), and the verdicts in words, each linking to its Validation tab. The model joins are a step panel, followed by the model map with evidence links and a start card.
• Models hub. In-page sub-tabs: Models, Combine, Compare, Validation, Papers. The verdicts are in words with links. Combine is a step panel. On a phone the chip sits under the question.
• Model pages (7 models × Overview, Methodology, Validation, Code, Paper). The sub-tab rows come from site_nav.SUBTABS.
  • Overview has a one-line verdict callout linking to Validation.
  • Validation has a stat row (the home of each headline number), a step panel of the tests, a limits card that keeps every failure statement, and an "On this page" rail.
  • Every tab ends with a "where next" row.
  • The shared /stepper.js replaces the inline copies.
• Charts. Labels are HTML laid over the SVG (chartkit.py), so they render at 12 px at any width. Ours is solid, the official series is dashed, legends are in HTML, and zero lines are drawn.
• Forecasts, data hubs and topic pages. These use the Forecasts sub-tab row (UK · US │ Overview · Data · topics) and an "On this page" rail. Each fact has one home: data readings live on topic pages only, and forecast rounds and scores live on /forecasts.
• Release notes. One stat tile, a provenance card and a "where next" row. They are regenerable with --rebuild.
• Connect and Contact. /connect is an MCP → CLI → Python step panel. /contact is a single card.

Phase 4 (QA) changes

• Phone reading.
  • Word-heavy tables become labelled cards below 640 px instead of 720 px side-scrolls. Numeric grids and the 41-row series register keep their scroller.
  • Sub-tab rows, table scrollers and method diagrams fade on the edge that hides content, and the chosen sub-tab is scrolled into view.
  • The five-tab model rows now fit at 375 px. The three us-hank pages have a sixth tab, so their row still overflows (407 px at 375) and relies on the fade.
• Method diagrams. On all seven Methodology tabs, labels rendered at about 5 px on a phone. Their smallest text is now 11 px, and below 640 px each diagram keeps its native width and scrolls inside its figure. Every diagram was checked at 1440 after the change, and the labels that clipped were fixed:
  • define: the multiplier diagram is now 640 units wide, so its two bar labels fit (they clipped before this PR too). The clean-room footnote is reworded to fit, with the same facts and numbers.
  • svar: the outputs box is wider, the "2 unidentified" line wraps, and the zero-restriction line is shortened.
  • obr and frb-us: the "converged" arrow label is now "done".
• Long intros. The two engine paragraphs at the top of /forecasts moved, unchanged, into an About expander. Ledes over 40 words were cut to the limit with every caveat kept: /economy, /economy/us, /forecasts/us and the DEFINE-UK replication report. The /economy/us "no US forecast" note moved from the title band into an intro card. Topic ledes keep proper nouns ("official Bank Rate", "UK nominal par yield") instead of lower-casing them. On /economy and /economy/us, the meta paragraph at the top of the topics card ("The rest of this site is organised by model…") was replaced by one sentence. It held no caveat or number, so it was removed rather than hidden.
• Bank Rate wording. /forecasts#limits said Bank Rate was "the only robust win". papers/boe-svar/figures/rolling_evaluation.json contradicts that. Against a drifting random walk, Bank Rate scores 0.793 at h=1 with DM p=0.018. Its Benjamini–Hochberg q is 0.362, and none of the 64 variable × horizon tests reaches q < 0.10 (minimum q is 0.362, or 0.401 on the production spec). The sentence now says Bank Rate comes closest and no gain survives correction for the number of tests run. It links to /svar/validation#forecasters and does not restate the figures.
• Smaller fixes. /models and /connect mark their first sub-tab before the first section is reached. /notes/releases gained its "where next" row. The OBR paper lede names the March 2026 EFO baseline again.

Round 2 (tab-by-tab economist and designer review, five areas)

• IA. Forecasts sits ahead of Models in the header. The home page tells the story in order. /models is a router.
• Model pages. All seven models have one model card (what it is, access, vintage) and one trust card (what is tested, what is not). Verdict wording is neutral. Each fact has one home; repeats became links.
• Charts. Strict colour roles (ours solid, official dashed), unit lines, label halos, outturn dots on fans, a log dot strip. The 2028–29 OBR reform bar is drawn as an outline because it is interpolated.
• Forecasts. The scoreboard leads and carries round, archived band, outturn snapshot link, signed error, band verdict and the no-change and published benchmarks.
• Release notes. 53 per-series notes are folded into 8 daily digests. Every retired URL has an explicit redirect in vercel.json to its digest anchor.

Round 3 (layout and density)

• Model names as h1s; the old slogans became the one-line description. Human model names lead (OBR emulator, BoE SVAR, PolicyEngine microsimulation…); code ids are secondary text.
• Prose longer than about 45 words moved behind "About this…" expanders. No caveat or number was deleted.
• Step panels are used only where every step has a right-panel card; elsewhere sections are one full-width column. Charts span the card.
• Section headers are an icon that means something, a short title and one grey line. Eyebrows dropped.
• One card level; no nested cards. The /models comparison table is transposed to five columns.

Screenshots (before → after)

Full-page shots at 1440 and 375 px were taken for home, /obr, /obr/validation, /models, /forecasts, /economy, a release digest and /svar/validation, before (origin/main) and after (this branch). They were reviewed locally and are not attached. What changes, page by page:

• Home. Before: long intro, verdict table. After: intro card with the hero fan, three entry cards, verdicts in words linking to Validation.
• Model Overview (/obr). Before: slogan h1, 100–150-word paragraphs, a step panel whose right side was empty after step 1. After: model name h1, a model card, short steps with a figure per step, detail in expanders.
• Validation (/obr/validation, /svar/validation). After: a stat row as the home of each headline number, the tests as rows, a limits card that keeps every failure statement.
• /models. Before: an eight-column comparison table clipped at 1440. After: five columns, models as rows.
• /forecasts. After: the scoreboard is the full-width lead card; limits are label/value rows.
• Phone (375 px). No page-level horizontal scroll; word-heavy tables stack into labelled cards; chart labels at least 11 px.

Content flags (published record or wording that changed, beyond layout)

1. Unemployment 2026Q2 outturn is now scored on /forecasts (it was never scored before). It falls outside the archived 90% band, with a smaller error than no change.
2. BoE MPR and SEF benchmarks were added to the scoreboard. A false claim that the BoE/SEF paths were stored in the round files was removed.
3. Bank Rate: "the only robust win" → comes closest; no gain survives multiple-testing correction (papers/boe-svar/figures/rolling_evaluation.json).
4. FRB/US: the Validation claim that the GDP trough had the "right sign, size and timing vs the Board's documented properties" was corrected to papers/frb-us/sections/comparison.tex (no published GDP figure; core inflation −0.034pp vs ≈−0.08pp). The Methodology claim that both multipliers sit inside published ranges was removed (the pegged year-2 0.90 is below 1.1–1.2).
5. OLG cohorts: Methodology "ages 21 to 100" fixed to 20–99 (S = 80; papers/psl-og/sections/calibration.tex).
6. /models#score: "GDP −0.02% on impact building to −0.06%" deleted (see decision 1).
7. Release notes: 53 per-series URLs now redirect to 8 digest anchors.
8. Model-review corrections: see round 5 below.

Not in this PR, for Vahid: data PR #195 has been open since 27 Aug and adds 20 per-series notes from the old generator, which conflicts with the digests; fetch-vintages.yml never runs ingest or score, so the record stays stalled at 26 Aug.

Round 5: correctness pass from the 5 Nov model-repo reviews

Each change cites its repo source in the commit message. Numbers changed (old → new):

Page	Old	New	Source
/frb-us/validation limits	core PCE miss −1.7pp, "2.7% vs 4.4%"	−1.65pp, 2.8% vs 4.4% (2026Q1); −0.65pp (2026Q2)	us-frb-model backtests/longbase_vs_outturns.md, fix in PolicyEngine/us-frb-model#10 (LONGBASE now compounded like the outturn)
/frb-us/validation limits	"GDP, unemployment, funds rate within ~0.3pp"	unemployment and funds rate within 0.2pp; GDP within 0.3pp in 2026Q1, +0.9pp in 2026Q2	same file
/svar/validation coverage	"no Covid dummies … depresses coverage"	labelled as the no-dummy run; the dummy run (the fans) gives 69% (h1) and 56% (h8)	boe-var-model results/coverage_evaluation.json production_spec
/olg showcase	"Government consumption (% of GDP) 44.28%"	"Total public spending (% of GDP) 44.28%" (number unchanged)	baseline is the TME profile (52.3% in 2020); og-model-dashboard review; run dated 28 Apr 2026, pre-MTR-fix (og-model-dashboard#2)
/obr/validation tile	"Weak, report-only"	"Weak, regression-gated"	obr tests/test_model_invariants.py:910-962

Wording withdrawn or corrected (no number changed):

• DEFINE-UK 1.78: "social benefits fall → biased up → upper-leaning" is withdrawn on /define/validation, the replication report and the paper (calibration.tex, limitations.tex, PDF rebuilt). In the cached run social benefits rise (+8.0 of 118.2), so the bias direction is not established. The withdrawn IMF 1.1–1.5 band is gone from the methodology diagram and the kv rows, and the label "DEFINE-UK 1.78" now reads "1.78, to 2040 (our computation)". Source: Correct the multiplier caveat, state units, gate stock-flow consistency, and sweep every cached scenario define-uk-model#15.
• OBR profits: the 82.77% error no longer "traces to the OSHH constant": the OSHH anchor applies in raw solves too. The identity-only floor is 23.9%, from mixing EFO and ONS vintages; the rest is open in obr-macroeconomic-model#42. The site and paper are updated (PDF rebuilt), and OAHHx is described as an asset stock, not "other income". Source: Fix lagged ONS formulas, employment seed units and out-of-window shocks; correct the profits explanation obr-macroeconomic-model#48.
• OBR reform chart: £6.46bn and £7.38bn are microsimulation figures, not an "emulator figure" (make_charts.py, regenerated).
• US HANK: the "Open models" tab is now "HANK survey". The Status line (and the registry, README, hank_summary framing and MCP docstrings) says that only the steady state is validated and IRF magnitudes are not. "Realistic distribution" and "response concentrating among high-MPC households" now say the distribution is calibrated (about 54% hand-to-mouth vs about one third in US data) and that the by-wealth split is an MPC-proportional allocation. Source: Review fixes, scenario sweep tests, pinned G-shock and MPC figures us-hank-model#4 README.
• Each fact once: the frb-us 100bp block (the table with the reference column is kept), the frb-us wage bill −0.65% and the DEFINE 4.66%/1.31% tile (the table carries the source) now have one home each.

Backend: obr_shock passes end = max(2027Q4, last shocked quarter) to run_reform, and the periods cap goes 40 → 25 (the databank ends 2031Q1). Before this, periods=20 reported a 12-quarter cumulative as a 20-quarter one. The new tests fail on the old core.py. No real solve past 2027Q4 was run locally.

Decisions for Vahid

1. /models#score. We deleted "GDP −0.02% on impact building to −0.06%". It is a per-quarter percentage and conflicts with the ratio framing on /obr/validation. Restore it with a units note, or leave it out?

2. Kept duplicates.

   • £6.46bn appears on /obr#answers, /obr/validation, /pe/validation and the pe chart.
   • (frb-us 100bp and DEFINE 4.66%/1.31% were resolved in round 5.)

   Keep these, or reduce each to one home?

3. Word-valued stat tiles. The olg, us-hank and define tiles hold words ("none", "anchors", "1 row"). Keep them as tiles, or turn them into callouts?

4. Same forecast twice. The hero fan on the home page and the boe-svar Overview outlook chart plot the same forecast. Keep both, or replace one with a link?

5. /economy trend figures. They no longer show the latest value, which now lives on the topic pages. Is that acceptable?

6. DEFINE diagram label. Resolved in round 5: the label now reads "1.78, to 2040 (our computation)" and the withdrawn IMF band is gone.

7. Site-contract test change. In test_site_contract.py, the min-height: 112px assertion was replaced. It pinned the two-row mobile header from Fix platform audit findings #70, which the single --nav-h band replaces (DESIGN §2.4, §3.1). The test now checks the same intent on the new shell:

   • the header height and sticky offsets read --nav-h;
   • the tab row scrolls inside itself;
   • none of the old 112, 129 or 172 px offsets remain.

   Is that acceptable as the new contract?

8. Method diagrams. On a phone they use a native-width scroller with a fade instead of the chartkit HTML-label overlay. Converting them to the overlay is a larger rewrite of 40 hand-drawn SVGs. Accept the scroller, or schedule the rewrite?

9. Table stacking needs JS. reveal.js decides which tables become cards on a phone, using a word-heavy heuristic with a data-stack override. Without JS every table scrolls. Accept, or write data-label into the markup and generators instead?

10. Paper and report landings restate Validation figures. The /papers/boe-svar result paragraph repeats 0.32pp and the smallest adjusted q of 0.36. /papers/define-uk and /reports/define-uk-replication repeat 4.66% and 1.78. These are each paper's own summary, so they were left alone. Keep them, or link to Validation?

11. Step-panel spacing. Each step is at least 70vh tall so the sticky panel can follow along. This leaves a large blank stretch after short steps, for example after step 2 of "How the models connect" on the home page. Keep it, or lower it to around 55vh?

12. Home lede. The home intro-card lede is 47 words and links all seven models. §7.1 treats it as an intro card rather than a title band, so the 40-word limit was not applied. Trim it?

13. /contact. It has no "where next" row because §7.9 says "nothing else". Confirm.

Tests

• PYTHONPATH=integration/src python3 -m pytest tests integration/tests -q: 1467 passed, 101 skipped (round 5). tests + test_site_contract.py alone: 1172 passed, 19 skipped (the count fell from 1564 when 53 release notes became 8 digests; tests/test_site_integrity.py is parametrized per page, 1000 → 595 cases). Its discovery floor test_the_suite_actually_found_the_site was lowered from 70 to 60 pages to match.
• python3 site_contract.py and python3 site_nav.py pass.
• Every generator --check passes: economy/build.py, economy/topics.py, notes/generate_release_notes.py, forecasts/{make_us_baseline,ingest_outturns,archive,make_open_fans,score}.py, data/build_page.py, data/fetch.py, assets/make_hero.py, assets/make_current_outlooks.py, olg/figures/make_showcase.py, validation/figures/make_charts.py.
• Internal links: a stdlib crawl (rerun 6 Nov: 65 pages, 2,936 links, 0 broken) of the pages and hrefs and srcs, including vercel.json redirects, cleanUrls and #fragment ids, found no broken links.
• Browser QA at 1440 and 375 on 42 pages:
  • no page-level horizontal scroll at 375;
  • no chart or diagram label below 11 px;
  • rail-link anchor targets land below the sticky rows;
  • focus rings show on the header and pathway (checked by screenshot); every sub-tab link has a computed 3 px solid outline;
  • dark theme checked on home, /obr/validation, /forecasts and a topic page;
  • no "poor" or "rich" wording and no tick glyphs;
  • for each fact once, every number in each model's Validation stat row was searched for on all other pages. Only the duplicates listed in decisions 2, 6 and 10 were found. The other §7.0 rows were not swept by script.

reveal.js keeps its ?v=2 query string, which is hard-coded on the pages. A cached copy only means tables scroll and rows don't fade.

🤖 Generated with Claude Code

[vahid-ahmadi]

Merge plan for the round 5 stack into #200 (checked 6 Nov 2026)

Tips tested: #200 e4f7500e · #204 83649b0f · #206 baf999a9 · #205 6e6f2027 · #219 7cd00930 · #216 6ca535de · #226 2cdff6bd · #229 a7ad367a · main 153cf419. If any tip moves, re-run the simulation before merging:

git fetch origin && git worktree add --detach /tmp/r5sim origin/vahid/redesign-r2-integrate && cd /tmp/r5sim
for b in r5-ui-home r5-ui-models r5-ui-data r5-compare r5-api-scenarios r5-obr-pin r5-drift-check; do
  git merge --no-ff --no-edit origin/vahid/$b || { echo "conflict at $b"; break; }; done
PYTHONPATH=integration/src python3 -m pytest -q tests integration/tests && python3 site_contract.py && python3 site_nav.py
cd - && git worktree remove --force /tmp/r5sim

Rules

• Use "Create a merge commit" for every PR in this stack. Do not squash or rebase. Model pages: one-column methodology, one overview order, names not ids #206, Round 5 ui-data: forecast-evaluation scoreboard with full-height rail, flat archive, chart units and sources #205 and Models: 'Same question, every model' comparison view (/models/compare) #219 already contain merges of r5 UI: question router on /models, copy-paste /connect, glossary, home and contact #204's branch, and Decision needed (#218): bump the hosted OBR pin to obr-macroeconomic-model main (e4d7faa) #226/Weekly drift check: hosted model output vs the numbers the site quotes #229 contain Integration layer: scenario sweep fixes (fiscal-year OBR window, SVAR simple y/y, units, fail-fast validation) #216's. When r5 UI: question router on /models, copy-paste /connect, glossary, home and contact #204 is squashed, the test merge of Model pages: one-column methodology, one overview order, names not ids #206 conflicts on 10 pages (glossary, home, /models, 6 methodology pages). With merge commits, every step is clean.
• Branches are not auto-deleted when a PR merges (delete_branch_on_merge is off), so GitHub does not retarget stacked children. Retarget each child by hand before you merge it. Do not delete any head branch until its children are retargeted. If a base branch is deleted outside GitHub's post-merge button, the child PR closes.
• After each merge, wait for the checks on Redesign the static site: Autumn Budget look, one fact per home, readable layouts #200's branch to go green before the next one.
• Decision needed (#218): bump the hosted OBR pin to obr-macroeconomic-model main (e4d7faa) #226 needs your decision on Hosted OBR adapter pins obr-macro-model at a commit before reforms were scored around the anchored baseline #218 before it merges. Weekly drift check: hosted model output vs the numbers the site quotes #229 contains Decision needed (#218): bump the hosted OBR pin to obr-macroeconomic-model main (e4d7faa) #226's commits.

Order and commands

R=PolicyEngine/macro; B=vahid/redesign-r2-integrate
gh pr ready 204 -R $R && gh pr merge 204 -R $R --merge     # r5-ui-home
gh pr ready 206 -R $R && gh pr merge 206 -R $R --merge     # r5-ui-models
gh pr ready 205 -R $R && gh pr merge 205 -R $R --merge     # r5-ui-data
gh pr edit 219 -R $R --base $B                              # was vahid/r5-ui-home
gh pr ready 219 -R $R && gh pr merge 219 -R $R --merge     # r5-compare
gh pr ready 216 -R $R && gh pr merge 216 -R $R --merge     # r5-api-scenarios
# --- only if you choose option 1 on #218 (bump the OBR pin) ---
gh pr edit 226 -R $R --base $B                              # was vahid/r5-api-scenarios
gh pr ready 226 -R $R && gh pr merge 226 -R $R --merge     # r5-obr-pin
gh pr edit 229 -R $R --base $B                              # was vahid/r5-obr-pin
gh pr ready 229 -R $R && gh pr merge 229 -R $R --merge     # r5-drift-check
# --- then the redesign itself ---
gh pr ready 200 -R $R && gh pr merge 200 -R $R --merge     # into main

In the GitHub UI, the same steps are: "Ready for review", then "Create a merge commit", in this order. For #219, #226 and #229, first click "Edit" next to the title and change the base to vahid/redesign-r2-integrate.

The UI group (#204, #206, #205, #219) and the integration group (#216, #226, #229) do not touch each other's lines. I merged them in both orders (UI first, then integration first) and got byte-identical trees. #221 and #228 form a separate chain onto main. They still merge cleanly after #200 lands.

If you choose option 2 on #218 (wait): leave #226 and #229 unmerged. #229 is built on #226, so it carries the pin bump. You can merge #200 without them.

Checks on the final tree (base + 204 + 206 + 205 + 219 + 216 + 226 + 229, merge commits)

• PYTHONPATH=integration/src pytest tests integration/tests: 1686 passed, 103 skipped.
• python3 site_contract.py OK; python3 site_nav.py OK.
• python3 <generator> --check passes for all 16 node-free generators: make_current_outlooks, make_hero, build_page, data/fetch, economy/build, economy/topics, forecasts/archive, ingest_outturns, make_open_fans, make_us_baseline, forecasts/score, release notes, make_showcase, make_charts, models/compare/build, run_battery.
• tools/prerender_math.py --check needs node, which this Mac does not have. The node-free verify_static() returns no problems. The 6 methodology pages, vendor/mathjax/chtml.css and tools/ are byte-identical to Model pages: one-column methodology, one overview order, names not ids #206's tip, where CI's full check passed ("Site provenance" on baf999a).
• Linkcheck over local href/src links and anchors: 71 pages, 3056 links, 0 broken.

Fixes made at the source (no hand resolution)

• Integration layer: scenario sweep fixes (fiscal-year OBR window, SVAR simple y/y, units, fail-fast validation) #216 6ca535de: in the HANK adapter caveat, CLI help, module comment, MCP tool docstring, summary framing and shock framing, "validated replication" / "steady state validated" became "partial replication: six of seven solved steady-state parameters match; IRF magnitudes are not validated". The registry status strings and the README table were left alone on purpose. Round 5 ui-data: forecast-evaluation scoreboard with full-height rail, flat archive, chart units and sources #205 and Model pages: one-column methodology, one overview order, names not ids #206 own them, the seven model cards must quote them verbatim (test_model_cards_quote_the_capability_registry), and economy/topics.py renders them. Changing them on Integration layer: scenario sweep fixes (fiscal-year OBR window, SVAR simple y/y, units, fail-fast validation) #216 as well would have forced card edits and a topics regen that collide with Round 5 ui-data: forecast-evaluation scoreboard with full-height rail, flat archive, chart units and sources #205 and Model pages: one-column methodology, one overview order, names not ids #206. The registry is the same in either merge order because Integration layer: scenario sweep fixes (fiscal-year OBR window, SVAR simple y/y, units, fail-fast validation) #216 touches only its own source_url/package lines.
• Weekly drift check: hosted model output vs the numbers the site quotes #229 a7ad367a: canaries.json pinned each quote to page:line. Once the UI PRs were in, the lines had moved, and 7 drift tests failed on the merged tree. quoted_at now names the page only. The line is found from quoted_anchor, and the test requires the anchor exactly once on the page. No expected values changed.

Follow-ups and decisions, by PR

• Redesign the static site: Autumn Budget look, one fact per home, readable layouts #200: decisions 1–5 and 7–13 in its body (Broaden site copy beyond UK-only scope (US models planned) #6 is resolved). Data PR Refresh official data vintages #195 conflicts with the digest notes. fetch-vintages.yml never runs ingest or score.
• Model pages: one-column methodology, one overview order, names not ids #206: the comment above the us-hank status in capabilities.py still says "only the solved steady-state parameters are validated". Fix it on Model pages: one-column methodology, one overview order, names not ids #206, because it sits next to Model pages: one-column methodology, one overview order, names not ids #206's status hunk. The svar overview chart title says "Current boe-svar …" (in Round 5 ui-data: forecast-evaluation scoreboard with full-height rail, flat archive, chart units and sources #205's file assets/make_current_outlooks.py).
• r5 UI: question router on /models, copy-paste /connect, glossary, home and contact #204 / Model pages: one-column methodology, one overview order, names not ids #206 / Round 5 ui-data: forecast-evaluation scoreboard with full-height rail, flat archive, chart units and sources #205 (vocabulary, Referee review of the model pages (6 Nov 2026) #227): reports/define-uk-replication still quotes "PARTIALLY VALIDATED". Per-dimension registry grades ("implementation fidelity: strong" for HANK and SVAR) and band provenance are unchanged.
• Integration layer: scenario sweep fixes (fiscal-year OBR window, SVAR simple y/y, units, fail-fast validation) #216: SVAR future rounds move to the simple y/y basis. On the archived 2026-07-29 round, CPI medians would rise by 0.03–0.05pp. No archived record changes.
• Decision needed (#218): bump the hosted OBR pin to obr-macroeconomic-model main (e4d7faa) #226: needs your choice on Hosted OBR adapter pins obr-macro-model at a commit before reforms were scored around the anchored baseline #218. Option 1 (recommended in the PR) is to bump the pin and redeploy Modal. After that, the CGIPS sentence (Site figures with no committed source: OBR 2031Q1 drift, Bank Rate ±100bp path, FYCPR, CGIPS sweep, HANK and OLG runtimes #207) and the /obr window wording are follow-ups.
• Weekly drift check: hosted model output vs the numbers the site quotes #229: the OBR canaries fail until Decision needed (#218): bump the hosted OBR pin to obr-macroeconomic-model main (e4d7faa) #226 is merged and the MCP server is redeployed. The weekly cron runs from main only after Redesign the static site: Autumn Budget look, one fact per home, readable layouts #200 lands.

[vahid-ahmadi]

Combined merge order with the pipeline PRs (#221, #228, #195): simulated 6 Nov

Recommendation: merge this stack first if it can land by about 11 Nov. Then #221 → token/setting (#223) → #228 recovery → dispatch fetch-vintages (it rebuilds #195 from the new main) → close/reopen #195 for CI → merge #195 before the 12 Nov GDP release. If #200 can't land by then, keep the pipeline-first order and rebase this PR using the steps below.

What I simulated (local scratch worktrees; nothing pushed or merged): both orders, at tips #204 475867bd, #206 12f0a878, #205 862ec6cb, #219 fab6b2bd, #216 6ca535de, #226 2cdff6bd, #229 a7ad367a. Four of these tips have moved since the merge plan above. The stack still builds with merge commits and no conflicts.

• No source file conflicts in either order. forecasts/score.py (the "Last scored" line), notes/generate_release_notes.py (digests), vercel.json and data/refresh.py all auto-merge. Every conflict is in generated output.
• After the resolution below, both orders end at the same tree, byte for byte.
• Final tree: all 18 node-free --check gates pass, refresh.py's gate run passes (1408 passed, 26 skipped), and the full suite gives 1883 passed, 109 skipped, after the anchor fix pushed to Weekly drift check: hosted model output vs the numbers the site quotes #229 (e1578302). Before it, 1 failed: test_quoted_anchor_is_on_the_page_once[define_green_public_investment]. Model pages: one-column methodology, one overview order, names not ids #206 a4b4fed2 added a second "peak +0.92%" to define/validation. The canary anchor is now "peak +0.92% vs". No expected value changed.

This stack first. #221 and #228 merge clean on top. refresh.py already calls this PR's generator (--vintage, --check). A dispatch with --carry-from data/vintage-refresh produced #195's content with no hand edits: CPI 2026Q3 3.10 and GDP 2026Q2 1.37 added, scorecard 2 → 6 forecasts over 2 periods, every gate passed including the append-only checks. Merging #195 as it is now would conflict on 14 pages, so don't. Let the bot rebuild it instead.

Pipeline first (fallback): rebase this PR. 16 conflicting files, all generated: economy/index.html, 5 UK topic pages, economy/us/index.html, 4 US topic pages, forecasts/index.html, notes/releases/index.html, sitemap.xml, forecasts/outturns.json, forecasts/scorecard.json.

1. Take this PR's side for all of them except forecasts/outturns.json.
2. For outturns.json, keep main's rows (gdp 2026Q2, cpi 2026Q3). Append this PR's unemployment row exactly as it stands (vintage 2026-08-26, release_updated 2026-08-17). Don't let ingest write that row: on the merged tree, ingest re-dates it to the 2026-11-05 vintage / 19 Oct release.
3. Run score.py, then make_open_fans.py, generate_release_notes.py --rebuild, economy/topics.py, economy/build.py, data/build_page.py and assets/make_current_outlooks.py, then score.py again. All gates then pass, and the record shows "Last scored 5 November 2026".
4. Add 20 explicit redirects to vercel.json, /notes/releases/2026-11-05-<series> → /notes/releases/2026-11-05#<series>, plus any later per-series dates. Those pages will have been live. The tests don't catch a missing one.

Pre-emptive commits: one, #229 e1578302 (the drift anchor; see above). None for the cross-stack conflicts. Every conflict depends on data the other branch doesn't have, so committing it ahead would change what either branch publishes.

Full notes: scratchpad/r5/stack-merge.md.

[vahid-ahmadi]

CHANGELOG.md conflicts across the stack: fixed.

The changelog pass had put an [Unreleased] edit on every stacked PR, so they conflicted with each other once one merged, and this PR and #221 both created the heading at the same spot on main.

• This PR (f8e42eee) now holds the only redesign-stack entry: ### Site redesign stack (#200), bullets prefixed by PR number (Redesign the static site: Autumn Budget look, one fact per home, readable layouts #200, r5 UI: question router on /models, copy-paste /connect, glossary, home and contact #204, Model pages: one-column methodology, one overview order, names not ids #206, Round 5 ui-data: forecast-evaluation scoreboard with full-height rail, flat archive, chart units and sources #205, Models: 'Same question, every model' comparison view (/models/compare) #219, Integration layer: scenario sweep fixes (fiscal-year OBR window, SVAR simple y/y, units, fail-fast validation) #216, Decision needed (#218): bump the hosted OBR pin to obr-macroeconomic-model main (e4d7faa) #226, Weekly drift check: hosted model output vs the numbers the site quotes #229, BoE SVAR: label replication figures with their run vintage (boe-var-model 845e2a4) — option (a) for boe-var-model #24 #231, BoE SVAR: replication figures to 2026Q2 data — merge only if boe-var-model #24 merges (option b) #232).
• r5 UI: question router on /models, copy-paste /connect, glossary, home and contact #204, Model pages: one-column methodology, one overview order, names not ids #206, Round 5 ui-data: forecast-evaluation scoreboard with full-height rail, flat archive, chart units and sources #205, Models: 'Same question, every model' comparison view (/models/compare) #219, Integration layer: scenario sweep fixes (fiscal-year OBR window, SVAR simple y/y, units, fail-fast validation) #216, Decision needed (#218): bump the hosted OBR pin to obr-macroeconomic-model main (e4d7faa) #226, Weekly drift check: hosted model output vs the numbers the site quotes #229, BoE SVAR: label replication figures with their run vintage (boe-var-model 845e2a4) — option (a) for boe-var-model #24 #231, BoE SVAR: replication figures to 2026Q2 data — merge only if boe-var-model #24 merges (option b) #232 no longer change CHANGELOG.md (one small commit each). Round 5 ui-data: forecast-evaluation scoreboard with full-height rail, flat archive, chart units and sources #205 also merged r5 UI: question router on /models, copy-paste /connect, glossary, home and contact #204 forward, because it already contained r5 UI: question router on /models, copy-paste /connect, glossary, home and contact #204's old changelog commit.
• Data pipeline: fetch, ingest, score and open one checked PR; stop the nowcast's silent no-op #221 / Recover the 50 orphaned refresh snapshots with repeat days dropped (#222) #228 keep their entry under ### Data pipeline (#221, #228). Both stacks share one [Unreleased] heading from commit 84e9539f, which both branches merge. A comment line separates the two blocks, so they never edit the same lines.
• BoE SVAR: label replication figures with their run vintage (boe-var-model 845e2a4) — option (a) for boe-var-model #24 #231: merged Model pages: one-column methodology, one overview order, names not ids #206 forward. svar/methodology/index.html keeps Model pages: one-column methodology, one overview order, names not ids #206's layout plus BoE SVAR: label replication figures with their run vintage (boe-var-model 845e2a4) — option (a) for boe-var-model #24 #231's pinned 845e2a4 link and label. Now mergeable. BoE SVAR: replication figures to 2026Q2 data — merge only if boe-var-model #24 merges (option b) #232: re-synced; kept its 2026Q2 numbers and added Model pages: one-column methodology, one overview order, names not ids #206's three glossary links.

Checked on scratch branches (merge commits):

1. This PR + 204 → 206 → 205 → 219 → 216 → 226 → 229: no conflicts. Same tree with the UI order reversed and with the integration group first. Full PYTHONPATH=integration/src pytest tests integration/tests on tree (1) (24121b4e): 1691 passed, 103 skipped; site_contract.py, site_nav.py, verify_static() OK. CI green on all 12 PR heads.
2. BoE SVAR: label replication figures with their run vintage (boe-var-model 845e2a4) — option (a) for boe-var-model #24 #231 → BoE SVAR: replication figures to 2026Q2 data — merge only if boe-var-model #24 merges (option b) #232 on top: clean.
3. main + Data pipeline: fetch, ingest, score and open one checked PR; stop the nowcast's silent no-op #221 + Recover the 50 orphaned refresh snapshots with repeat days dropped (#222) #228 then this stack, and the reverse: no conflicts, identical final tree.

All 12 PRs show MERGEABLE, CI green. Nothing merged.

If #226/#229 (waiting on #218) or #231/#232 haven't merged here when this PR merges, delete their bullets from this PR's entry first.

Notes: r5/changelog-fix.md