Correct the paper's corporation-tax finding, and state three limits it omits (main.pdf needs rebuilding)

[vahid-ahmadi]

The correction

The paper reports that the corporation-tax deviation does not converge — compounding 1.21–1.27× per quarter to £43,387m by quarter 25 "with no sign of settling".

That was a bug in the closure's anchoring, not a property of the published equation, and it was fixed in PolicyEngine/obr-macroeconomic-model#37: holding the add-factors in log space (the EViews convention for a dlog equation) lets the published equation's own error-correction term give the deviation a steady state.

Measured on the corrected closure, same sustained +5pp rise:

horizon	|ΔIF|
q8	£0.24bn
q12	£0.38bn
q25	£0.68bn
plateau	~£0.95bn/qtr

The root is slow (~0.958/quarter), so the twelve-quarter horizon every published result uses reaches only ~40% of it. Corrected in experiments.tex, limitations.tex and results.tex, with the superseded figures kept and labelled rather than deleted.

Three limits the paper does not state

• No quarter of a reform solve reaches tolerance. The appendix describes the stall-break mechanism but not how often it fires: on a scored reform all 20 quarters of both solves exit stall, none on tol. A delta is the difference of two non-converged solves, and under a costing path that is flat across the first four quarters the GDP response can still change sign within them.
• The reform path is configuration-dependent. The 12-quarter 2025Q1–2027Q4 experiment is smooth and monotone; the same reform over the adapter's default 20-quarter window is neither, reaching −0.085% by 2027Q4 against Figure 4's −0.057%. Same equations, same vintage, different solve window.
• Whole blocks never respond. Under every lever tested, unemployment, employment, retail prices, exports and imports move by exactly zero. No Okun channel, no Phillips channel, no trade channel.

What I checked and got wrong first

I initially reported two further findings that did not survive checking, and they're worth stating so the record is clean:

• I said the −0.057% figure was an error. It isn't. figures/fig_reform_data.csv gives that exact path, smooth and monotone, and the text matches its own artifact. The −0.085% I measured is the adapter's default 20-quarter window — a different configuration. That's now recorded as a limitation, not a correction.
• I said the paper doesn't discuss solver non-convergence. It does — appendix.tex documents stall detection and calibration.tex documents de-seeding because of it. What was missing is the frequency, which the new item supplies.

⚠️ main.pdf is NOT rebuilt

There's no LaTeX toolchain on this machine and no CI job builds the paper, so main.pdf still reflects the pre-correction source. It must be rebuilt before the PDF is cited or re-published:

cd papers/obr-macro && latexmk -pdf main.tex

Testing

LaTeX not compiled (no toolchain) — brace and math-mode balance checked by hand on all three edited files. Site suite 1564 passed; contract clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Lcj9DDqam9KmVCfhEdnJcJ

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
macromod	Ready	Preview	Aug 26, 2026 12:47pm