Pre-Registration Amendment A2 — Powered Nowcast Re-run (Test A only)

Locked 13 June 2026, before any Germany/Japan or France-2019 data is pulled from Athena. Amends 2026-06-13-phase2-nowcast-preregistration.md. Applies to TEST A (nowcast) ONLY.

Why this amendment, and what was known when it was made (A1-spirit disclosure)

The 13 Jun France run (2026-06-13-phase2-results.md) left Test A inconclusive at n=18 and Test B failed on its merits (two pre-registered tests now agree national music mood does not predict French demand proxies; music degrades the out-of-sample forecast). On 13 Jun, Alex confirmed Germany and Japan chart data 2019–2025 are in Athena (same 5 source tables) and France is extended to 2019. That removes the only thing that made the generalisation bar untestable: it is now gated on an engineering pull, not a licence.

What was known at the moment of this amendment (disclosed so the chase is informed-not-blind):

• The France n=18 nowcast point estimates were inspected: savings-intent (primary) −4.9% RMSE, p=0.62; headline CCI (secondary) −23.2% RMSE, p=0.15. Neither significant.
• So this re-run is NOT blind — it is informed by having seen a large-but-non-significant point gain on the secondary outcome. That is legitimate (you may power up an informed hypothesis) but it must be stated, and it is.

What was NOT known / not done as of this lock:

• No DE/JP nowcast in any form; no France-extended (pre-2021) nowcast; no pooled estimate; nothing on DE/JP outcome or macro series. These remain genuinely unseen.

A2.1 — The verdict outcome is pre-committed to savings-intent (PRIMARY)

To stop a headline-only result being promoted after the fact:

• The verdict is decided on EUROSTAT_CCI_SAVINGS_INTENT (the primary outcome, the one Phase 1 validated contemporaneously). Headline CCI remains secondary, reported, never the verdict.
• A re-run that moves only headline CCI while savings-intent stays null is reported verbatim as: "the pre-specified primary nowcast did not confirm; one secondary macro series shows a weak point gain that did not clear significance" — NOT "the nowcast validated." This sentence is locked now.

A2.2 — Hold the generalisation bar; pooling does not substitute

The extra data ~triples n (France→~40+, more if pooled), which is the point. But:

• The bar is unchanged and per-market: a pass requires France AND ≥1 held-out market (Germany or Japan) each clearing the bar SEPARATELY (savings-intent RMSE-beat + Diebold-Mariano significant at α=0.05).
• A pooled multi-market model is a secondary, reported analysis only. A pooled n=40 that "passes" while no single market passes on its own is explicitly not a pass — pooling can let one market carry the others. Per-market first; pool second; report both.

A2.3 — 2020 COVID regime control (mandatory)

The 2019 extension pulls in the 2020 dislocation (simultaneous mood + macro shock). A structural break can manufacture or mask a relationship.

• Include an explicit 2020 regime control (March 2020–[end-of-2020] indicator, plus an interaction allowance) in both baseline and full models, so a relationship is not created or hidden by that single episode.
• Report the result with and without the COVID window; if the verdict flips depending on whether 2020 is in, the result is reported as regime-dependent, not as a clean pass.

A2.4 — Freeze caveat unchanged (history, not live)

The audio-feature freeze (end Nov-2024) is unchanged. The 2019 extension adds history at the start, not currency at the end. Therefore a pass here validates the signal on history; it does not validate a live product. Live deployment remains gated on the in-house extractor + the per-feature parity gate (pre-reg §3 deployment gate), unchanged.

A2.5 — Everything else inherits unchanged

Baseline (AR(2) + public predictors), nowcast timing (day-15, data through day-14), models (ElasticNet primary, RF secondary), DM test (Harvey small-sample corrected, α=0.05), and the pre-committed null reading ("could not confirm at this n," not disproof) all carry over verbatim from the parent pre-reg. New data only changes n and market coverage, not the test.

Data dependency to flag honestly: Test A on DE/JP needs not just their charts (in Athena) but their CCI + public predictors. Germany's CCI is Eurostat (our existing fetcher extends to DE). Japan's is not Eurostat (Cabinet Office / ESRI consumer confidence) and must be collected separately before the JP arm can run. The France-extended-to-2019 arm needs no new macro (FR CCI already spans 2019-01→2026-03) and can run as soon as the FR-2019 chart parquet lands.

Companion: 2026-06-13-phase2-nowcast-preregistration.md (parent), 2026-06-13-phase2-results.md (the run this responds to). Loader/queries: analysis/athena_export.sql, analysis/phase2_test_a_powered.py.