# Pre-registration: earthshine vs CERES, change in Earth's reflectance

Written 2026-11-08 23:52 UTC, before any albedo value from either series was opened in this session (prior knowledge only: Goode et al. 2021 report a decline).

## Question
Do the Big Bear Solar Observatory earthshine record (Goode et al. 2021, GRL, 1998–2017) and the CERES satellite record agree on the sign and rough size of the change in Earth's reflected sunlight?

## Data
- Earthshine: annual mean albedo / reflected flux anomalies from Goode et al. 2021 (table, supplement, or digitised from the figure if no table exists; digitising error stated).
- CERES: EBAF-TOA Ed4.x global monthly mean reflected shortwave flux (or albedo), annual means, from NASA LaRC (or the values plotted in the same paper if the raw series cannot be reached; that fallback is weaker and is labelled).

## Primary test (fixed now)
Overlap years = calendar years present in both annual series (expected 2001–2017).
Change = ordinary least-squares slope of annual means over the overlap years, converted to W/m² per decade of reflected shortwave (albedo × 340.2 W/m² if only albedo is given).
- PASS if (a) both slopes have the same sign, and (b) the ratio earthshine slope / CERES slope lies between 1/3 and 3, and (c) both series are reachable or reconstructable to within 0.1 W/m² per annual point.
- FAIL otherwise.

## Secondary (reported whichever way)
1. Full-record change: earthshine 1998–2017 first-3-years mean vs last-3-years mean.
2. Correlation of annual anomalies over the overlap.
3. Share of the earthshine 1998–2017 drop that falls in 2015–2017.
4. CERES change after 2017 (to today) for context.

## Pivots allowed if the primary fails (test before any REJECT)
1. Nightly prediction: model earthshine from a daily cloud map + geometry vs the paper's published nightly/seasonal values.
2. Evening (waxing) vs morning (waning) hemisphere difference in earthshine.
3. The 2015–2017 drop alone: same sign in both series.

## Amendment 1 — 2026-11-09 00:05 UTC, before any CERES value was opened (file still downloading)
- Earthshine source found: BBSO `Earthshine.txt` (550 nightly fractional albedo anomalies, Δp*/p*, 1998.90–2017.99), linked from the paper's data statement. It is already culled, so the paper's "≥5 nights per month, ≥6 months" rule leaves no year; that rule is dropped.
- Annual earthshine value = mean of all nights in the calendar year; years with ≥ 10 nights only (drops 1998 with 7 and 2006 with 1). Sensitivity: mean of monthly means.
- Conversion to flux: ΔF = (Δp*/p*) × 0.30 × 340.2 W/m² (fractional anomaly × earthshine Bond albedo 0.30 × mean TOA insolation). The paper uses 342; difference 0.6 %.
- CERES: EBAF-TOA Ed4.2.1, global monthly `gtoa_sw_all_mon` (reflected SW) anomalies vs its own monthly climatology; annual mean of calendar years with 12 months (2001 onward; 2000 has 10 months and is dropped).
- Overlap years therefore: 2001–2005 and 2007–2017 (16 years).

## Amendment 2 — secondary test: does the day's satellite picture predict the night's earthshine? (2026-11-09 00:30 UTC, before any image was scored)
Not part of the PASS/FAIL verdict above (already PASS). Prior art: Pallé et al. 2003 (JGR 108, D22) simulated 1999–2001 nights from cloud data; this is a recount with true-colour images, and it checks the page's own picture.
- Nights: BBSO `Earthshine.txt`, 2000-03-01 onward (MODIS Terra). Observer Big Bear (34.258 N, 116.921 W). Time = the file's fractional year (≈1 h resolution).
- Image: NASA GIBS MODIS Terra corrected-reflectance true colour, 0.5°, gaps filled with Aqua (from 2002-07-04). Per pixel the mosaic of that pixel's local solar date at the observation time. Linearised with the inverse of the GIBS stretch (0,30,60,120,190,255 → 0,110,160,210,240,255), mean of R,G,B. No-data pixels (all channels < 12) excluded; nights with > 25 % of the lit-and-seen weight missing are dropped.
- Model apparent albedo = Σ R·μ₀·μ·cos(lat) / Σ μ₀·μ·cos(lat) over pixels lit by the Sun (μ₀>0) and facing the Moon (μ>0); Sun and Moon from JPL DE422 via skyfield.
- Model residual = model minus a cubic fit in |lunar phase angle| over all kept nights (the paper's method for the observed anomalies).
- Test: Pearson r between model residual and the file's anomaly. Reported as "the picture tracks the nights" if r > 0 with one-sided p < 0.01; waxing and waning reported separately; Spearman as sensitivity.

## Amendment 3 — 2026-11-09 00:45 UTC: the file's timestamps are dates, not times
Run 1 (times taken from the fractional year) gave r = −0.06 (n in `nightly_result_run1_wrongtimes.json`). Its UTC-hour histogram was flat across all 24 hours for waxing and waning nights alike, which is impossible for Big Bear evenings/mornings; the fractional parts × 366 cluster at .0 and .5, so the file stores a day number (÷366) with a half-day flag that does not follow the phase sign. Run 1 is therefore void as a test and is reported only as this note.
Re-run rule, fixed before scoring: civil date D = round((t − year) × 366 − flag·0.5) days after 1 January (day 0 = 1 January). Waxing night (negative phase at 12 UT of D): observation at local sunset + 2 h at Big Bear on D. Waning night: local sunrise − 2 h on D. Everything else as Amendment 2. Sensitivity: D − 1 and D + 1, and ±1 h.


Correction note (2026-11-09, after review): in Amendment 3, "r = −0.06" for the void run 1 is its waning subgroup; its overall r was +0.016. Run 1 remains void.
Amendment 2 says "already PASS": that refers to conditions (a) and (b); condition (c) could be checked only at the endpoints (VERDICT.md, Corrections).