Anywhere.

Treat the support — the record set — as a frozen representation, and the weights as a task head fit on top of it. Recalibrating that frozen support to a new geography's published totals is head-only adaptation: every household stays exactly as observed; only its influence on the aggregate changes. Importance weighting of a fixed sample is the zero-parameter-update case of adapting a generative model — the same operation populace already runs to hit a congressional district's targets, run again at a larger domain distance.

A capacity ladder sits above that zero-parameter case: frozen support plus new weights, then frozen support plus imputed extensions for variables the target geography needs and the US support doesn't carry, then a fully adapted generative model that replaces the support itself. Higher rungs spend more capacity and can close gaps reweighting cannot. The mechanism at each rung is swappable — the question this page and its paper ask is whether the adaptation framing survives the climb, not which rung is correct.

populace's local-area method already reweights one national file to a US congressional district's published totals — same country, same institutions, the smallest domain shift the operator runs on. Moving the same operator to the UK and to Belgium does not change what it does; it changes how far the source domain sits from the target. The diagram below places three actual runs on that axis.

The protocol borrows directly from held-out benchmark practice: adapt to a target domain without touching its held-out ground truth, then score against that ground truth once. The UK supplies the ground-truth leg because native survey microdata — the Family Resources Survey — exists to hold out and score against. Belgium supplies the deployment leg precisely because that check is unavailable: no comparably granular Belgian household microdata sits behind the recalibration, so deployment is checked the only way available, cross-engine agreement on an identical population, rather than against native microdata.

Calibration forces target-geography totals to match by construction. Reweighting is fit to hit published aggregates, so once it runs, the aggregates match almost by definition. Every residual gap that remains against the FRS microdata — not the totals, the joint distribution underneath them — measures something reweighting cannot touch: the joint structure the recalibrated file inherited from the US support it started from. That gap is the quantitative test of the populace thesis stated on the support strategy page: everything downstream can reweight the support; nothing downstream can repair it. The UK benchmark turns that sentence into a number.

The UK leg works because native household microdata exists and can be held out. Most geographies fail that condition — not because no one holds the data, but because the data cannot leave the building it lives in. National statistical institutes and research centres routinely hold exactly the restricted microdata this page has no access to. The protocol above does not need that data to travel; it needs a referee that can.

It requires an open referee. The referee travels even where the data cannot. Instead of asking a data custodian for microdata, ship the custodian a sealed, pre-registered evaluation procedure: the recalibrated support, the population-view harness, and a frozen scoring configuration. The custodian runs it locally, against their own restricted ground truth, inside their own environment. Nothing sensitive leaves the building — only a low-dimensional scorecard comes back: energy distance, coverage, classifier AUC, and the tail block, the same four numbers this page already reports for the UK leg. A scorecard contains no micro records, so publishing it is safe by construction, not by review.

The dress rehearsal already earned its keep. Run against the UK's own held-out file — the one geography where this page can score against native microdata — the pack did not just confirm it executes. It caught a concrete, mechanical defect in the transferred file: a missing income stream that zeroed out state pensions and cascaded into the means-tested benefits that depend on them (section 07). A referee that surfaces a specific data error on its first real run is a referee worth shipping to a custodian who holds data this page will never see. The pre-registered pack, run as a custodian would run it, did exactly the job it exists to do.

Reweighting can only redistribute mass across the households already in the support; it cannot manufacture a joint structure the US population never exhibits. If the target geography has a household-composition pattern, an income-tax-and-benefit interaction, or a demographic combination absent from the source support, no amount of reweighting recovers it — that is a limit of this rung of the capacity ladder, not a bug in the reweighting method. Higher rungs (imputed extensions, or a fully adapted generative model) exist precisely to spend more capacity against this limit. This page and its paper report where that limit binds, not just where the operator succeeds.

The US support, recalibrated to Belgian published totals, is the first deployment-leg result this page can report in full. Both numbers below come from a completed populace-be run, not a benchmark still executing.

The UK leg has now run three times. Its subject is the prior US-to-UK transfer: an existing recalibrated file built from open US microdata, reweighted to UK national, regional, and country targets — the same target set the native Family Resources Survey file calibrates to. The pre-registered pack scored the as-found v1 file against the held-out native FRS file before that file was touched, exactly as section 03 requires; scored a repaired-feed v1.1 rebuild under the identical frozen scoring configuration; then scored a v1.2 rebuild applying one declared mapping transform on top of that repair, isolated against four hypotheses committed to the record before the run. All three are diagnostic passes, not a finished benchmark: v1 caught a concrete feed defect, the v1.1 repair moved the headline verdict and a continuous proper score in opposite directions on the same view, and v1.2 closed that same view's gap to its floor — reported together below, not as a single improved number.

Two defects the referee caught, one repaired feed, one declared fix. Running the pack as a custodian surfaced specific, mechanical data errors, not vague quality complaints, twice over. The v1 transferred file's state-pension total came out at zero against a native total of 112.5 billion pounds — the feed carried survivor, dependant, and disability social-security streams but no retirement stream at all, so every household's state pension defaulted to nothing. The cascade was fully traced: no state pension raised means-tested pension-credit entitlement nearly tenfold, which in turn drove reported household benefits far below the native file. A repaired feed and rerun — with the pre-registered scoring configuration left unchanged — completed as v1.1: state pension recovered to 91.5 billion pounds against native's 112.5 billion (an 18.6% gap, down from a complete absence), and universal credit landed within 0.6% of native. But the same corrected feed carried a far larger US realised-capital-gains distribution into the transfer's existing one-to-one income mapping, which the UK market-income surface admits almost none of — income tax rose to 34.2% above native (was 2.6% below at v1) and household market income to 872% above native (was 156% above at v1). A repaired feed cannot repair a wrong mapping; it can only show more clearly what the mapping gets wrong.

The mapping needed a mapping decision, not another feed repair, so v1.2 declares one: US realised capital gains are excluded from the transferred UK file entirely — an explicit, documented reclassification, not a silent patch. Four hypotheses about what that exclusion would do were committed to the record before the run. Three were confirmed: household market income flipped sign, landing 31.6% below native rather than 872% above it — the transfer now understates market income once the phantom gains are gone, which is itself informative about what the transfer is still missing. The net-income view's energy ratio to its sampling-noise floor fell from 12.5× to 1.5×, essentially closing that view's gap. And the tax-and-benefit view, whose failure was never about capital gains, stayed exactly where it was: 90.1× its floor, clearly distinct, because the reported-benefit surface it needs was never written to the transfer and this fix doesn't touch it. The fourth hypothesis was refuted: income tax's 34.2% overshoot, which the v1.1 repair blamed on phantom gains entering taxable income, did not move at all — UK income tax doesn't tax chargeable gains, and gains entered no calibration target, so removing them changed nothing on this aggregate. That refutation opens a new, still-open question: what actually explains the income-tax gap, if not capital gains. A repaired feed cannot repair a wrong mapping, and a repaired mapping doesn't explain every remaining gap — but naming which mechanism did and didn't move each number, and testing that in the open before looking at the result, is what a referee is for.