# Pre-registration: the dark-evening test on regions not yet examined

Written 2026-09-26, committed and pushed before any of the tests below was run.
The git history of this file is the evidence of that order.

## What had already been seen

On the full Coral Spawning Database (window of nights -2 to 12 around full
moon), `analysis.mjs` gives:

- all ecoregions pooled: beta = +0.066, 95% [-0.370, +0.476], 454 spawning events
- six ecoregions examined one at a time (listed as `EXPLORED` in
  `analysis.mjs`): Central/northern GBR +0.726 [+0.219, +1.232]; the other
  five scattered around zero, North Ryukyu and Taiwan negative, all with wide
  intervals.

The GBR result was found by breaking the pooled test down by region. That is
exploratory, and it is the reason for this file: a subgroup that stands out
after looking at six is exactly what chance produces.

The held-out ecoregions contributed to the pooled figure above but have not
been looked at on their own.

## The tests, fixed now

Same code, same definitions (`analysis.mjs`, unchanged): beta is the
within-(taxon x ecoregion) slope of the night count from the full-moon date
on delta (first dark evening a day late), 2,000 event-level bootstrap draws,
seed 20260926.

1. **Primary.** `node analysis.mjs --set heldout`: every ecoregion not in
   `EXPLORED`, pooled.
2. **Secondary.** `node analysis.mjs --set "Southern Great Barrier Reef
   (including Keppel Islands and Capricorn Bunker Reefs)"`, the nearest thing
   to a replication of the GBR result.
3. **Secondary.** `node analysis.mjs --set Palau`, where de la Torre Cerro et
   al. (2025) found that darkening the post-sunset sky advanced spawning in
   *Acropora* aff. *hyacinthus* by one night.

## How each will be read

- 95% interval excludes 0 and includes 1: supports the dark-evening timer.
- excludes 1 and includes 0: the spawning night does not follow the dark
  evening.
- includes both: this region cannot tell.
- excludes both: something in between (a partial or mixed clock).

## My expectation, written down so it can be wrong

The pooled figure already contains these regions and sits near zero, so I
expect the primary test to come out near zero, "excludes 1". I have no
confident expectation for the two small secondaries and expect "includes
both" from at least one of them for lack of data. Whatever comes out is
reported on the page with the same prominence as the exploratory results.

## Erratum, added after the tests were run (2026-09-26, same session)

The summary of "what had already been seen" above is wrong in one detail.
It says the five explored regions other than the central/northern GBR
"scattered around zero". Fiji did not: the exploratory run gave Fiji
beta = +0.675 with an interval excluding 0, on 16 spawning events (the
first bootstrap, in a scratch script, printed its upper bound as NaN, and
I read past it). The final figures for all six are in
`public/strata/before-the-moon-comes-up/data/results.json`. Nothing about
the held-out set, the tests or the decision rules changed; this corrects
only the description of the exploratory results, and the text above is left
as it was committed.