01
A multi-graph from OpenAlex
We ingest works, concepts, authors, institutions, and funders — 8.6k works across 6.8k fields — and connect fields by how they co-occur across the literature.
Method · and how we grade ourselves
Confident, because it's testable
A prediction is only worth funding if the method that made it can be checked. Here is the whole pipeline, and the score it earns when we hide the future and ask it to rediscover the past.
02
Cross-community link-prediction
We score pairs of fields that live in different research communities on shared neighbors (Adamic–Adar), author bridges, and recent momentum. The highest-scoring non-edges are the collisions: connections that don't exist yet but the structure says should.
03
A Frontier Brief, written by Claude
For the top collisions, Claude reads the evidence and drafts a short brief — thesis, why now, who is positioned, what to fund — always with an explicit block on what would disconfirm it. The model writes prose; it does not invent the ranking.
The backtest
Does it beat the obvious baseline?
We hide everything after a cutoff year, rank collisions from the graph as it stood then, and measure the ROC-AUCagainst the truth — the pairs that actually fused afterward — versus preferential attachment, the “rich get richer” null model.
Held out from 2,021 · 288 field pairs actually fused afterward · 91 within reach of the engine
0.82
ROC-AUC — our engine
0.71
ROC-AUC — popularity baseline
1.5×
Skill lift over baseline
91
Real future collisions graded
ROC-AUC of cross-community Adamic-Adar vs preferential-attachment; ground truth = concept pairs whose first co-occurrence fell after the cutoff