We keep the receipts
A foresight product is worth what its record is worth. Every collision the engine publishes is written down the day it is published and graded here against what the literature actually did — including the calls that went nowhere, and the ones that turned out to have been fused all along.
A prediction is matured once it has had 180 days to come true; younger calls are listed but excluded from the hit rate, so the number cannot be flattered by counting predictions made last week. 42 calls are marked revealed: the two fields turned out to have already co-published before we called the gap. Those count against us.
How often do pairs this good actually fuse?
The Collision Index ranks. It does not, on its own, tell you the odds. So we rebuilt the graph as it stood in 2019, 2020, 2021, scored every pair the engine would have considered, and counted what happened next across 361,581 candidate pairs.
| Strength band | Pairs | Fused | Rate | Within 3y | vs base rate | Median lead |
|---|---|---|---|---|---|---|
| Top 0.1% | 363 | 9 | 2.48% | 2.48% | 7.2× | 2y |
| Top 0.5% | 1,809 | 51 | 2.82% | 1.93% | 8.2× | 3y |
| Top 1% | 3,617 | 90 | 2.49% | 1.69% | 7.3× | 3y |
| Top 5% | 18,081 | 310 | 1.71% | 1.12% | 5.0× | 3y |
| Top 10% | 36,160 | 466 | 1.29% | 0.83% | 3.8× | 3y |
| Top 50% | 180,791 | 1,037 | 0.57% | 0.36% | 1.7× | 3y |
| All candidates | 361,581 | 1,241 | 0.34% | 0.22% | 1.0× | 3y |
Base rate: 0.34% of all candidate pairs went on to co-publish. That is the number every collision has to beat to be worth reading, and it is why the board shows a band rather than a bare score. 'First co-occurrence' means first inside this corpus, which begins at the ingest year and is a topic-seeded sample of OpenAlex — not first in the literature. That censoring inflates the absolute probabilities; it does not favour high-scoring pairs over low-scoring ones, so the lift between bands is the more robust figure.