Register UK State Pension gates and run the one-shot scored projection #504

Part of #501 (WP9). The plan is at docs/plans/uk-state-pension-dynamics.md (merged in #502), section "Acceptance benchmarks". WP0 (ADR 0005, 1 agent-day and half a human day) has no dependencies and lands first.

Scope

• ADR 0005, recording the domain decision. ADR 0002 stays Proposed for general UK panel work.
• A UK gate file, separate from gates.yaml, as ADR 0002 requires ("No change to gates.yaml").
  • G1: DWP Spring 2026 caseloads by type, 2030-31.
  • G2: spending by type, 2030-31, on the forecast's own path.
  • G3: Stat-Xplore full-rate bands for cohorts held out of estimation.
  • G4: Stat-Xplore abroad backcast, May 2022 to March 2026. The window spans documented series breaks, disclosed in the registration.
• Floors. The plan's tolerances are proposals. This repo's half-versus-half floor measures sampling noise in panel moments and does not apply to comparisons with administrative aggregates. The amendment defines a floor for each gate, with a machine-checked derivation:
  • base-sample noise from replicate bases, for every gate;
  • DWP's revision between forecast vintages, for G1 and G2;
  • Stat-Xplore's disclosure noise and quarter-to-quarter variation, for G3 and G4.
• Disclosure. G1-G4's targets are published and were read in scoping. They are held out from estimation, and the registration says they are not blind.
• Registered comparisons, configured in advance and reported, not gated:
  • R1-R2: the OBR's July 2026 chart 3.11 path, and triple lock minus earnings, with rates set directly.
  • R3: DWP's £15bn and £50bn as a single-path comparator, against single-path runs only.
  • R4-R7: protected-payment inflows, Pension Credit, the cost of unfreezing, Northern Ireland.
  • R8: a blind prediction of the September 2027 Stat-Xplore release, registered before it exists.
  • R9: DWP Spring 2026 caseloads and spending by type, each year 2025-26 to 2029-30.
• Registration on Candidate 2 design: latent-permanent conditioned chained QRF (research memo) #42, then the one-shot run, published whatever the result.

Acceptance

• The gate is locked before the run.
• The run is published whatever the result.
• The scorecard holds aggregates only.

Depends on

• WP8 bundle.
• An engine release carrying WP1-WP4.

Estimate: 1-2 agent-days plus 1-3 weeks elapsed; 1-2 human review days.