Four Preprints and a Patent Stack: What VIDRAFT Is Actually Publishing

Four Preprints and a Patent Stack: What VIDRAFT Is Actually Publishing

Research on a small-lab budget

VIDRAFT publishes. For a team competing with much larger labs, writing things down -- preprints and patent filings -- is how we turn "method over scale" into something checkable instead of rhetorical. Here is what is actually on the record, with identifiers so you can go read the primary sources.

Four preprints

ID	Topic
arXiv 2605.14386	Darwin: model merging
arXiv 2607.18340	Quantum
arXiv 2607.24830	A numerical realization related to the Riemann/Weil line
SSRN 6280258	AI metacognition

The Darwin paper is the backbone of how we build models: evolutionary merging, where you search over combinations of existing checkpoints to produce a stronger model instead of pretraining a new one from scratch. It is the most direct expression of the capital-efficient thesis -- compute spent on search rather than on raw pretraining.

The quantum preprint and the Riemann/Weil-line numerical work sit further from product and closer to basic research. The metacognition paper, posted on SSRN, studies whether a model can estimate its own reliability -- knowing when it does not know -- which feeds the AI-safety benchmark work we run under the FINAL-Bench org.

The patent portfolio

Alongside the papers there is a patent portfolio: 12 filings with 317 claims in total. Patents and preprints do different jobs. Preprints invite scrutiny and citation; filings stake defensible ground around the methods. A small company needs both -- open enough to be credible, protected enough to be fundable.

How to read all of this

Treat the identifiers as invitations, not trophies. Pull up arXiv 2605.14386 and check whether the merging method does what the abstract claims. Look at the metacognition setup on SSRN and decide whether the evaluation convinces you. That is the whole point of putting a number on the record: it lets a reader replace our summary with their own judgment.

We publish because a challenger earns credibility by exposing its reasoning, not by asking for trust. The same discipline that keeps our compute budget honest -- measure it, write it down, let others check -- applies to the research.

Honest scope

These are recent preprints. Citation counts and social traction are early-stage -- low, in plain terms -- because the work is new and, in some cases, narrow. A preprint is not peer-reviewed: an arXiv or SSRN ID means the work is public and timestamped, not that a community has validated it. The Riemann/Weil-line entry is a numerical realization, not a claimed proof of anything. Patent filings are filings -- claim counts describe scope on paper, not granted, litigated, or commercially proven rights. We list the identifiers precisely so you can judge for yourself, and if a result does not replicate, tell us. That feedback is why we publish.

More: https://vidraft.net