The build, measured.
Every number below is computed from the build artifacts — the full administrative target surface, the shipped weights, and the comparison runs. The gaps are shown next to the wins.
loading…
Where every layer comes from.
Each block of the population traces to a named public source. Every donor is a primary survey; the enhanced CPS appears only as the benchmark the finished dataset is scored against.
How close the population sits to the record.
distribution of target misses
fit by target family — top families by target count
the worst-fit targets, named
We publish the misses with the hits: most of the residual error is concentrated in thin demographic and tail-income cells, the known frontier of the pool. The calibration roadmap targets these directly.
No aggregate leans on a handful of records.
Calibration without a bound concentrates mass: the unconstrained run pushed one synthetic household to represent over a million real ones. The shipped weights carry a hard per-record cap — chosen by sweeping the bound and keeping the fit — so the distribution stays honest.
Every variable, accounted for.
Each variable the incumbent enhanced CPS carries, side by side with this release: how filled it is, what it totals, and whether it is live here or still a gap. Search the full set.
release —
Against the incumbent.
loading…