Hypothesis scoreboard
One rule produced every verdict: a least-squares trend on the monthly series, read through the 95% interval on its slope. Entirely in the predicted direction is supported, entirely opposite is rejected, spanning zero is inconclusive. Each card opens its evidence.
The signature metric
Why the share rose: the decomposition
The evidence, hypothesis by hypothesis
H3 — Retention drifts down, but not past the bar we set
Sensitivity — all 27 threshold combinations
The H1 decline survives every cell of the grid
Change in economic accounts, first month to last, for every pre-registered (d, c, b) combination. The outlined cell is the default. Tightening the filter makes the decline steeper — the opposite of a threshold artefact.
Exploratory (not pre-registered)
Something happened in April 2026
What would change our mind
Each is a quantity this pipeline already computes;
re-run them any month with make refresh.
Bull / base / bear — indicators, not price targets
Amendments to the pre-registration
The pre-registration is frozen; changes are appended as dated amendments, each with its reason, its cost, and an assessment of whether it could bias a result. The original text is never edited.
Don't trust us. Re-run it.
Clone the repository and run the check — no cloud account, no keys. It rebuilds every number from the committed data and fails if a single figure in the prose has drifted.
$ git clone https://github.com/sambitsargam/near-filtered
$ cd near-filtered && make verify