Two-universe calibration (ONS frame + HMRC registered subset), full regeneration, and paper revision

[vahid-ahmadi]

Summary

Resolves the universe incompatibility in #37 by calibrating on two declared universes, regenerates every checked artifact, and revises the manuscript to match. This is the single regeneration María asked for; PRs #42–#45 are cherry-picked onto this branch so their text lands once.

Two decisions taken (both flagged in #39 as mine to make):

1. Population universe. The ONS VAT/PAYE enterprise frame is the modelled population; HMRC margins describe its VAT-registered subset and enter through a per-band registration propensity p_b = HMRC count / frame mass (0.89 below the threshold; 0.51–1.00 above, the remainder being PAYE-only/exempt enterprises). HMRC negative/zero traders are appended before calibration as an out-of-frame stratum. The 2023-24 employment CSV is rebuilt from Table 3 (enterprises) via scripts/etl_ons_tables.py, which also sources both vintages' turnover tables from Table 8; a test asserts the two tables agree to disclosure rounding and that the CSVs match the extract.
2. Ageing convention. Turnover and liability are aged together against nominal thresholds (the paper's own fiscal-drag logic), growth indexed to each vintage's data year; anchor_reform and threshold_sweep now use one rule.

What changed in the results

	before	after
Calibration overall (2023-24 / 2024-25)	92.9% / 93.5%	98.0% / 98.0%
ONS population / employment (2023-24)	88.3% / 92.1%	99.7% / 99.8%
Kish ESS (2023-24 / 2024-25)	93,887 / 275,186	137,054 / 192,917
Anchor 2025-26 (full dereg. / 43% retention) vs HMRC −£185m	−£364m / −£207m	−£234m / −£134m
Sweep at £70k / £120k (2025-26)	+£1,456m / −£1,553m	+£187m / −£1,190m
Raise to £100k / linear taper / 10% / 15%	−£876m / −£1,525m / −£515m / −£258m	−£742m / −£1,406m / −£446m / −£223m
Constant-50% marginal taper (new)	—	−£968m
Bunching E, 2023-24 (frame)	10,418	5,498
In-scope firms in dominated band	156,747 (all rows)	142,860

Why the big moves:

• Anchor now lands between the two conventions because a rise releases only in-scope firms below the £88k deregistration threshold (statutory gap £2,000). A fixed cross-section has no entry margin, so 2027-28 undershoots HMRC's −£50m; the fixed-preference reading of the frame's 89% below-threshold registration share (−£38m) is reported as a lower bound and explained as frame selection.
• Cut rows collapse because ~89% of below-threshold frame enterprises are already registered; the paper now calls them frame-confined lower bounds and points at Extend the population frame to all businesses (BPE) so threshold cuts and the administrative cliff are in-frame #25.
• Reform menu is priced on in-scope firms only.
• Taper: the £141,667 band top is a property of the linear design; a constant-50% marginal taper of the same width removes the notch at a third less cost with half the peak marginal rate (taper_band_top, make_schedule_taper_flat, tests).
• Marginal buncher depends on δ under formulation A (£127k → £149k at δ=0.6, e=0.17); reported for δ ∈ {0, 0.4, 0.6}.

Fresh-reviewer items folded in

Deregistration threshold (1), taper band-top frontier (2), 2024-25 coarse-band bunching readback in Section 6 (3), stale "smooth-counterfactual sweep" text (4), literature additions Keen–Mintz, Kanbur–Keen, Dharmapala–Slemrod–Wilson, Mirrlees Review, BizTax (5), δ-dependent n_H (6), retention as a bound given HMRC's ~zero below-threshold liability (7), sector targets vs negative/zero traders (8), by-construction band accuracy stated (9), dominated-band claim rewritten (10), ageing factors and the [£70k,£130k] re-optimisation window stated (11), population statement (12), pass-through and FRS conventions stated (15).

Reproduction

make reproduce runs ETL → both vintages → every artifact → figure sync (reproduce-heavy adds seed sensitivity and placebo B). All 70 tests pass; ruff clean. No quarto locally: the Paper workflow renders the PDF and now uploads it as an artifact (paper-pdf); paper/main.pdf should be replaced from that artifact before merge.

Not done here (tracked)

• results/populace_ledger_comparison.txt NOTE is pinned by a test and still cites the old scores (Reproduction and claim synchronisation: one reproduce command, claim manifest, and stale README/results notes #40).
• Claim manifest / reproduce --check (Reproduction and claim synchronisation: one reproduce command, claim manifest, and stale README/results notes #40).
• Citation checks for the reviewer's list (new issue).
• Nominal-growth factor provenance: stated as assumptions in the paper; source needed.

Closes #37. Refs #32, #38, #39, #40, #41.

🤖 Generated with Claude Code