Engineered for the
legal outcome.
Most patent search treats retrieval as a search-engine problem and quietly fails at the legal one. We started over from the model layer up. One rule: depth lives inside the system, restraint shows on the surface.
Trained on outcomes, not surrogates.
Most retrieval systems learn from proxies — keyword overlap, raw co-citation, abstract similarity. Useful, but a shadow of the real signal. The patents that look the most similar are rarely the ones that hold up where it matters.
We took a different path. The model is shaped by the legal record itself — by what experienced decision-makers have actually treated as decisive evidence. The result is a retriever that understands what kind of prior art wins, not just what kind of prior art looks adjacent.
A second look the language model can’t do alone.
The transformer is the first read. After it, every candidate passes through a structural-reasoning layer that's learned to weigh contextual signals an experienced examiner would notice without thinking — the kind of context that flat embedding similarity throws away.
The features and the way they combine are tuned against held-out outcomes, not chosen by intuition. The result is a final ranking that consistently outperforms either signal alone.
Every change has to prove itself.
Nothing ships unless it moves the eval. We hold out a frozen slice of recent legal outcomes and measure whether each new version of the model recovers more of the references that actually mattered. No vibes-based shipping.
When you read “the model is better,” what we mean is that it surfaces measurably more of the references that expert reviewers later relied on — on a benchmark we don't train against.
The model never invents what it surfaces.
Every reference Nulity returns is a real, issued patent retrieved from a deduplicated index — never a generated citation. The same applies to passage previews: pulled verbatim, with character offsets so anyone on your team can audit them.
We don't write claim charts on your behalf. We don't draft arguments. The model surfaces evidence and explains why each result matched. The legal craft is yours to apply.