UK dataset evaluation reports

Static evaluation surfaces for microcosm UK dataset candidates: calibration quality, support, downstream instruments and their sources. Internal to PolicyEngine; nothing here is a release verdict.

• K=25 local candidate at 60,000 households: inspection of a gate-blocked build2026-10-06

  The first graph-path local build from main at 25 clones: the dense 1.6m-household solve and the exact-k 60,000 refit side by side, with the gates that refused the artifact. Inspection only; nothing here is a release.

• The v21c national candidate against the national incumbent: the income anchors, the reserved top tail, every gate signed2026-09-23

  microcosm uk-income-anchors-280 @ f6d33a26 (on main b78f5a88): spine-u, the spine with the reserved SPI income band donors, calibrated to the 1,062-target national scoring surface through the rowwise national role with the HMRC income anchors at the calendar-2025 window, all six terminal gates green with two signed deferrals, and scored head-to-head against the enhanced FRS 1.57.3 on the common resolvable surface at one engine release (policyengine-uk 2.99.1 for every leg). Assessment only: nothing here is a publication decision.

• Two 55,000-household draws from a K=20 pool: the tolerant certainties-only draw beside the exact one, the dense solve and the enhanced FRS2026-09-20

  microcosm main ce76b358 + the signed v20 deferral and the tolerant-draw lane (3f18dd42): the spine-s K=20 pool solved dense (1,056,920 rows), drawn to 55,000 exactly (X50, pi >= 0.5, baseline floor) and to 55,361 as a certainties-only draw within +/-1 % at pi >= 0.95 (T95); both evaluated with the 9 September instruments on policyengine-uk 2.99.1 beside the exported dense solve (D20) and the enhanced FRS 1.57.3, every expectation sourced.

• The v20 national candidate against the national incumbent: one surface, one engine, every gate signed2026-09-20

  microcosm main ce76b358 + the signed self-employment deferral (90bdb809): spine-s calibrated to the 638-target national scoring surface through the production seam, all six terminal gates green, and scored head-to-head against the enhanced FRS 1.57.3 on the common resolvable surface — the candidate and the incumbent measured on the same targets, the same engine release (policyengine-uk 2.99.1 for every leg), with the rows the incumbent cannot express pruned loudly and receipted.

• A 55,000-household candidate rebuilt on the 10 September spine: the refit fixed, beside its dense version and the enhanced FRS2026-09-10

  microcosm main 0afb1235 + the baseline floor (ca435e20): the 55,000-household request rebuilt after the spine and target PRs merged on 8–10 September, first with the untrimmed Horvitz–Thompson baseline (Q50, mass-starved), then re-refitted from the same checkpoint with the baseline's probabilities floored (Q50f); evaluated with the 9 September instruments beside the exported dense solve (D2) and the enhanced FRS 1.57.3, every expectation sourced.

• Two 55,000-household candidates against their dense solve, R17 and the enhanced FRS2026-09-09

  microcosm #355 / PR #877: the first two sized UK candidates (P50 and P95b, 55,000 households each from the K=15 pool) evaluated for calibration quality, area support, downstream reforms, admin anchors and dashboard replays, with every expectation sourced.