[vahid-ahmadi]

Demo-critical, draft by decision (see the log below). The hosted OBR model is not demo-safe at its current pin.

Why

Hosted pin e4d7faa solves the published closure. Measured on the live server today:

• obr_shock(CGG, +1250, 12): about 4.5 minutes, 12 of 12 quarters unconverged (Gauss-Seidel stall break), multiplier exactly 1.0000 and flat.
• score_reform(model="obr"), default window: 20 of 20 quarters unconverged, with quarterly GDP deltas jumping (−2.09bn in 2030Q4, then −1.33bn).

obr main 94f08b9 (#54, #55, #57 and #58, merged 15 Nov) makes the nominal-GDP pin the default. I re-measured locally at 94f08b9 through core.obr_shock, one heavy job:

Lever (×12 quarters)	Converged	GDP £m, q1 / q4 / q8 / q12	Time
R +1.0 (Bank Rate +100bp) at 94f08b9	all quarters	−129 / −405 / −480 / −522	78 s
HHDI_ADDFACTOR 1,615.75 (1p basic rate)	all	−150 / −311 / −449 / −533 → ratio q12 0.33 (year-1 mean 0.145)	76 s
CGG +1,250	all	+1,267 / +1,283 / +1,300 / +1,311 → multiplier 1.01 → 1.05	71 s
TCPRO +0.05	all	0 / −77 / −239 / −382	78 s

The Bank Rate path matches the OBR session's independent Actions run (obr-for-macro-243.md).

Change

• Pin e4d7faa → 94f08b9 in modal_app.py and pyproject.toml.
• canaries.json, plus the drift and wiring tests: the Bank Rate expectations and the household ratio 0.33. The stale-pin test now rejects e4d7faa's path.
• CGG caveat (served in every obr_shock payload) and the /obr/code example: "about 1.01 on impact, rising to about 1.05", still against the OBR's 0.6.
• /obr/validation: the Bank Rate card has the new numbers (cut +134..+541, from the OBR session's run). The hosted card, which still described pin 1ed742c and "until Decision needed (#218): bump the hosted OBR pin to obr-macroeconomic-model main (e4d7faa) #226 is deployed" (stale since 6 Nov), now describes 94f08b9 and the score_reform runtime (Hosted score_reform (OBR, default window) has not finished within 10 minutes since the 30 Oct deploy #243). The three-configurations line says every quarter converges.
• "No quarter converges" is removed from /obr, /obr/methodology, /obr/validation and /models/validation; the /obr verdict says every quarter converges.
• capabilities.py OBR runtime: "about one to five minutes for a raw shock; over ten minutes for a translated reform scenario on the default window" (was "seconds"). The reform topic page is regenerated.

Decision log (Fable 5.2, asked because Vahid asked for no questions)

• Earlier today, question: bump each hosted pin now or wait? Fable: OBR should wait for obr Bring back the ambient field, tuned calmer #61, Point canonical URLs at policyengine-macro.vercel.app #65 and Fix audit findings; add evidence charts and cross-shop comparisons #66 and for a check of whether main fixes the stall.
• Re-asked with new evidence (main converges everywhere; hosted e4d7faa doesn't converge in any quarter). Fable's answer: prepare the bump PR now, as a draft. If obr Point canonical URLs at policyengine-macro.vercel.app #65, Fix audit findings; add evidence charts and cross-shop comparisons #66 or Bring back the ambient field, tuned calmer #61 merge before Vahid merges this, re-pin and regenerate once inside this PR.
• Reasoning: "the stall check was the open item, and it came back showing main fixes it". The hosted pin is not demo-safe, and preparing a PR is not publishing. Fable also said: take figures from the same SHA as the pin; rewrite the 1.0000 caveat; re-measure the canaries one heavy job at a time; and say plainly that this fixes correctness, not runtime.

Re-pinned to 6fbbe06 (obr #65 and #66 merged)

The OBR session's final figures are in (obr-for-macro-243.md, "FINAL figures", Actions run 43720935064), and obr #65 (EU payments held at zero) and #66 (household ratio on both bases) are merged to main 6fbbe06. This PR now pins 6fbbe06. I re-measured locally at 6fbbe06 under the heavy-job lock:

	6fbbe06 (mine)	OBR session final
Bank Rate +100bp, GDP £m q1/q4/q8/q12	−128.7 / −405.4 / −480.6 / −522.2, every quarter converged	identical; cut +133.9..+541.5
CGG +1,250 multiplier	1.014 → 1.049	1.01 → 1.05
Household 1p basic rate, q12	real 0.330, current-price 0.372 (df.attrs)	0.33 / 0.37
Corporation tax +5pp, FYCPR q12	site already says −£404m	−403.9

The canary expectations are unchanged from 94f08b9. The worked example now quotes the like-for-like current-price ratio, 0.16 → 0.37×, against the OBR's 0.3. That is the basis obr #66 says matches the OBR. The real-GDP figures obr_shock returns (0.145 → 0.33×) sit beside it, and the canary still checks 0.33, since that is what the tool reports.

Still pending in obr, neither of which changes the hosted deltas (per the OBR session): #61 (IBUSX bracket) and #67 (demo and result labels). Fable's condition was "re-pin and regenerate inside this PR if #65/#66/#61 merge first". #65 and #66 are in. Marked ready for review.

Scope and what's still open

• Correctness, not speed. The OBR solve is slightly slower at main (216 s vs 178 s cold, OBR session). score_reform's runtime is mostly the microsim; see Run each population simulation once, not twice, and say how long it takes #260 (the double run) and Report stage timings in the OBR score_reform result #253 (timings).
• Not re-measured here, and left unchanged: the CGIPS sweep figures (/obr/validation, OBR session: now proportional, −1,387/−2,773/−5,547), the free-running MAPE in capabilities.py (4.32% / 7.22%; the OBR session's 4.12% / 6.90% needs Point canonical URLs at policyengine-macro.vercel.app #65), and the /models/compare OBR rows (hosted runs from 6 Nov at e4d7faa). The compare battery should be rerun against the hosted server after this deploys.
• Free-running scorecard not refreshed here: the /obr/validation tile (4.32%), the consumption, household-income, profits (82.77%) and current-account figures, the free-running chart (papers/obr-macro/figures/fig_free_running_data.csv), and capabilities.py's "4.32% and 7.22%". The OBR session's final note gives the new default-pin scorecard (GDP 4.12%, consumption 6.90%, household income 4.50%, profits 12.45%, current account 1.19%, 8 of 11 in band), but not the chart series or business investment. Refreshing it properly needs the scorecard CSV copied from obr, as a follow-up.

Merging redeploys Modal

Tests

pytest tests 3257 passed; integration (excluding remote) 453 passed; site_contract.py and topics.py --check OK.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
macromod	Ready	Preview	Nov 15, 2026 2:21pm UTC