Topological Void Analysis by Kris Pan: 2,128 invention candidates

Kris Pan submitted Topological Void Analysis on 20 Apr 2026, presenting a mathematical framework that locates unexplored, relevant regions inside high-dimensional knowledge spaces and outputs concrete invention leads. Applied to roughly 140,000 indexed documents, the method produced 2,128 invention candidates across 96 targets, with automated and adversarial filtering results reported in the paper.

What is Topological Void Analysis and how does it define a void?

Topological Void Analysis, or TVA, defines a void as a "triads (A, B, C)" structure in a dense-sparse hybrid embedding space where three formal conditions must hold. First, both concepts A and B must be semantically cohesive with a domain anchor C; second, the pairwise similarity of A and B must fall within a calibrated marginality band that avoids both obvious combinations and unrelated noise; third, A and B must share a sparse lexical bridge while the geodesic midpoint on the embedding hypersphere is unoccupied.

Pan frames TVA as a way to formalize the intuition of an "unexplored region": the framework combines dense semantic embeddings with sparse lexical signals to flag candidate connections that are relevant to a target goal yet absent from prior art. The paper situates TVA against existing approaches such as keyword search, citation proximity, and human intuition, arguing TVA supplies a formal criterion for what counts as an unexplored, actionable gap.

How was TVA evaluated and what did it find?

In the reported experiment TVA was applied to approximately 140,000 indexed documents and generated 2,128 invention candidates mapped to 96 targets; 90% of those candidates survived automated quality filtering, and a subsequent four-specialist adversarial review produced 191 REVISE verdicts and 1 APPROVE verdict, with an overall 0.05% end-to-end figure stated in the paper.

The evaluation pipeline therefore had three measurable stages reported in the text: candidate generation at scale (~2,128 candidates), automated quality filtering with a 90% survival rate, and a human adversarial stage yielding 191 REVISE and 1 APPROVE from four specialist reviewers. The paper also includes two case studies that, according to the author, demonstrate TVA surfaces non-obvious connective tissue rather than merely obvious related pairs. The manuscript is 11 pages long and includes three tables and two case studies as described in the submission metadata.

Why it matters

TVA supplies a formal, reproducible way to search for innovation opportunities inside dense technical domains such as operating systems or hardware/software co-design by defining what an "unexplored" region is in embedding space. That matters because it replaces informal signals like intuition or simple proximity with mathematically specifiable criteria, enabling systematic candidate generation at scale and measurable downstream filtering and review outcomes.

If the framework scales and its human-review yield improves, teams could use TVA-style signals to prioritize R&D effort, focus patent landscaping, or seed interdisciplinary connections that are hard to surface with keyword or citation methods alone.

What to watch

Watch for independent replications on other corpora and for broader expert-review campaigns that report APPROVE rates beyond the single APPROVE noted here. Another concrete signal will be whether the author or others publish the code, embeddings, or datasets used to reproduce the ~140k-document run and the two case studies.