# Deviations and added checks

Everything below was added after the locked plan was run (28 October 2026). The locked results stand as reported; these are follow-ups, labelled exploratory.

## Why
A first diagnostic (`analysis/grid_diagnostic.json`) showed the ANM grid copies the gauges at their own cells (median grid/gauge ratio 1.000) and that anomalies at cells 10 km apart correlate at 0.989. The grid is a smooth surface between about 160 gauges. Consequence: design A cannot see inside a 10 km circle that holds no gauge, and its placebo spread measures that smoothness, not a limit on real rain. Design A and the 2025 reversal are therefore not shown as findings; the gauges are.

## Predictions written before the added checks
Design B (downwind minus upwind, −5.0 %, placebo p = 0.031) is the one result that touches the "drought in the next village" claim. Two explanations compete:
- Terrain: east of the Carpathian belt the plain is drier than the hills to the west, and 2019–2024 were dry in the east. Prediction: B for the Timiș unit (mountains lie east of its sites, so "downwind" points uphill) is near zero or positive; B shrinks when post-years stop at 2022; B shrinks in placebo draws whose 1995–2016 downwind/upwind ratio matches the real one.
- A rocket effect: B stays negative in Timiș, with post-years to 2022 and in matched placebo.

## Added checks
1. Leave-one-unit-out for A and B (run).
2. Gauge-only check at R = 10, 15, 20 km (run).
3. B without the 2018 cohort (run): −1.2 %.
4. B for each operating unit on its own.
5. B with post-years ≤ 2022 (before the 2023 grid change).
6. Placebo B restricted to draws whose 1995–2016 downwind/upwind ratio is within 0.03 (log) of the real one.
7. Count of the 86 dated circles that hold a rain gauge.