Does the rule hold up?
Three tests, including the ones it fails
A rule that sends money to crises should be checked against the people who already do it, against anyone trying to game it, and against the choices its designer made. Here are all three, run on public data. The weak spots are in the results, not hidden from them.
1. Against the UN's own choices, 2019 to 2025
| Rule | Same places | Rank agreement | Shared top 10 |
|---|
Year by year
| Year | CERF's underfunded window | This project | Gap only | Share unfunded | Most neglected 15 | Largest differences (this project) |
|---|
2. Against padding
If one crisis's appeal asked for 20% more, or reported 20% more people in need, how much more money would the rule send it? Each crisis is tested in turn on this year's real data.
| Scenario | Ask +20%: median gain | Worst case | People in need +20%: median gain | Bought its way in |
|---|
3. Across 2,000 versions of the rule
| Crisis | Severity | Data reliability | Share of the money: 80% range, default marked | Funded in | Top five in |
|---|
What these tests cannot show
Recompute everything: backtest.json (from scripts/backtest-solidarity.py) and robustness.json (from scripts/robustness-solidarity.mjs), both in the site's repository, on the rule in allocate.js.