Raw data stays home.
Only conclusions cross.
Keeping bees, you want to understand the hive — but it only tells you anything if you listen to its signals, on its terms. Instrument it, and the question generalizes: how does any system come to understand its own domain, from its own data, without surrendering that data to anyone? That's the work.
The hive is the architecture.
This program didn't start with a model; it started with a colony. A hive reasons locally — no bee sees the whole, nothing reports upward, and the intelligence lives exactly where the work happens. The aquaponics loop taught the same lesson a second way: a living system tells you what it needs, if the instruments are where the life is.
The research is the software version of that pattern: small models close to the signal, memory the operation owns, and conclusions that can travel without dragging the raw data with them.
Local reasoning.
The research is a federated, locally-run reasoning system: each node runs its own local model with retrieval over its own data, so the intelligence stays where the data lives. The knowledge corpus — the Brain — is the substrate, and the open problem is fidelity: getting a small local model to retrieve the right thing and cite it honestly instead of confidently making something up.
One measured result anchors the research — a margin-conservation law: across the tested range, the answer margin narrows on a clean, predictable curve, from 0.090 to 0.011.
The work is being developed toward federal research funding — a USDA SARE resubmission and an NSF SBIR proposal are both in progress.
Live nodes
Working systems instrumented to read their own domain and write to memory they own.
An open lab
The build is shown in the open — the methods, the messy parts, and what each node learns.
The experiments
Where the pattern gets pushed: new domains, new signals, the same principle of kept data.
What we build with this lives on the lab side — the live systems, the working notes, and the parts that break.
Visit Killercatfish