What Is RVGN
At Iridae, we build models for real environments — environments where the underlying network of interactions, dependencies, or influence is often incomplete, noisy, or still emerging. RVGN, our Recursive Variational Graph Network, is a core research architecture designed for exactly that setting.
Most graph models assume the graph is known. RVGN starts from a different premise: in many important problems, the graph itself is uncertain. Instead of forcing a brittle yes-or-no view of structure, RVGN maintains a calibrated belief over possible connections and refines that belief as more evidence arrives.
The result is an architecture that can reason through uncertainty, not just around it.
Why It Matters
When relationships are only partially observed, conventional graph models tend to collapse ambiguity too early. That can make downstream predictions look confident even when the underlying structure is unclear. RVGN is built to preserve that uncertainty explicitly, so decisions can reflect what is known, what is likely, and what remains unresolved.
That matters in settings where structure is noisy, delayed, or contested — and where acting on the wrong structure can be expensive.
What Makes RVGN Different
RVGN is not just a graph neural network with an uncertainty score attached. It is a probabilistic graph reasoning architecture that treats connectivity as something to be inferred, updated, and inspected.
At a high level, RVGN:
- Maintains explicit probabilistic beliefs over candidate relationships
- Iteratively refines those beliefs using learned message passing
- Exposes uncertainty in a form downstream models can actually use
- Supports reasoning and planning over multiple plausible graph realizations rather than a single guessed structure
This makes RVGN useful not only for prediction, but for scenario analysis, prioritization, and decision support where the relational picture is incomplete.
Why We’re Excited About It
RVGN represents the kind of work we believe matters in applied AI: not another wrapper around a foundation model, but a new modeling layer for problems where uncertainty is part of the environment itself.
It reflects a broader Iridae view that intelligence architectures should not just output answers — they should maintain structured beliefs, update them as evidence changes, and make their uncertainty legible to the models and people around them.
How It Fits Into Iridae
Inside Iridae, RVGN is becoming a reusable uncertainty-aware graph module that other architectures can build on top of. It helps turn partial evidence into actionable structural beliefs, which can then feed forecasting, planning, prioritization, and other downstream reasoning workflows.
For customers, that means decisions informed by the uncertainty in the world, rather than obscured by it. For researchers and builders, it means a serious modeling substrate for uncertain relational inference — and a chance to work on graph learning that is both principled and product-relevant.