Two counterfactual estimators · one physics-gated engine
Property prediction breaks at the discontinuities. That's where discovery lives.
A small structural change, a large and sudden jump in property — an activity cliff in a small molecule, a phase boundary in a material. A smooth regressor interpolates across the jump and underestimates it. Our analogy-based counterfactual estimators with a physics-gated engine instead transfer the jump from a case that has already crossed the same boundary.
The problem
However rich the features, a smooth model can't represent a discontinuity.
Near the boundary the structure-to-property map is many-to-one, so no amount of capacity resolves it. The lever is which example you reason from — the most chemically similar one, or the one that crossed the same boundary.
Two counterfactual estimators
Counterfactual prediction by analogy.
Every discovery decision is a counterfactual — what would the property be under a different structure? We make it computable by casting the question as an analogy to cases that have already resolved it empirically.
If A has this property, what does an analogous B have?
Use when — you have entities, but no measured transformations.
If A→B did this, what does C→D do?
Use when — you have the transformations: measured changes to reason from.
The result is a counterfactual constrained by physical law — a transfer between cases the same mechanism governs, not an unconstrained extrapolation.
The product
It doesn't replace your model. It wraps it.
Two estimators, one physics-gated engine: the dual-entity estimator conditions on a single measured anchor; the quadruple-entity estimator transfers a measured change. Same wrapper — the data you have picks which.
Any point predictor drops into one slot — a GNN, a random forest, a transformer, an in-house model. Both are validated on out-of-distribution kinase potency and perovskite oxide bandgaps — domains that share no chemistry, only the same failure structure.
The flywheel
It isn't just an improved predictor. It's a loop.
The analogy engine is native to active learning. Every experiment you run adds on to the pool of analogies for the next proposal — so the system doesn't just predict, it compounds.
The ask
Become the counterfactual layer under every discovery pipeline.
Every experiment a customer runs deepens a proprietary library of analogies that have crossed real discontinuities, and that library is what makes the next prediction better. We don't sell a formula. We sell a system that gets sharper the more it's used.