Why similarity search fails at risk prediction

Every risk team that touches machine learning starts in the same place: encode your incident database, embed a new situation, retrieve the nearest neighbors. We did exactly that: a multilingual E5 model, fine-tuned on 60,000 structured industrial event records. Retrieval quality was genuinely good. If something similar had happened before, we found it.

The failure mode is structural, not a matter of scale. Embedding similarity measures resemblance between descriptions. Risk propagation is governed by mechanism: what energy, what pathway, what target. Two situations can read as near-duplicates in text space and diverge completely in consequence space. A leaking flange at 3 bar next to a drainage culvert is not the same event as the same flange over sealed concrete, but their descriptions differ by four words.

Worse, the events that matter most are precisely the ones with no neighbors. A retrieval system is silent, by construction, about the scenario that has never been recorded. Senior EHS engineers are not: they reason over the mechanism (this source can reach that target through this flux) and extrapolate to configurations they have never seen.

That observation set the bar for what came next: an encoder that learns the logic of propagation rather than the geometry of descriptions. Judged not on retrieval metrics, but on whether it produces the scenarios a mechanism-literate expert would, including the unprecedented ones.

What we can say publicly about how it works, and where we draw the line, is on the Technology page. What we can demonstrate is more interesting: give us a configuration outside anyone's event history and judge the scenario tree yourself.