System Index
Plain-English explanations of the terms used across EcoKure and Jarvi3. No jargon left unexplained — if a word on this site isn't obvious, it's here.
Taxonomy lanes (taxonomy lines)
The "lanes" Jarvi3 uses to send each request to the right place. Instead of one giant model answering everything, a request is sorted by type and routed down the lane built for it. Every new lane teaches Jarvi3 to handle another kind of task.
Deterministic Taxonomy GLM
Jarvi3's core architecture. It classifies a task, then deterministically (predictably, the same way every time) routes it to the best specialist — rather than firing a huge general model at every query.
Specialist (specialist brain)
A small, focused model or tool that's excellent at one kind of task. Routing to a lean specialist instead of a giant generalist is what makes Jarvi3 fast, accurate and low-energy.
Global taxonomy
The living, growing map of every task-type Jarvi3 can route. It's already in use and expands as more lanes are added — the long-term goal being one shared taxonomy that covers the world's work.
Routing
Choosing the right specialist for a request, instead of sending everything to one enormous model. Routing is why the same answer can cost a fraction of the energy.
Model-agnostic
Not tied to any single AI model. A lane can call on potentially any model — whichever is the best tool for that job — which keeps Jarvi3 flexible and future-proof.
Offline-capable
Able to run without an internet connection or a remote data centre — on the computer already in front of you.
Edge / on-device
Running the AI locally, "at the edge" (your device), rather than on distant servers. Lower latency, no data leaving your machine, far less energy.
Decentralised
Not dependent on one central data centre or company. Work is spread out, which is more resilient, more private, and lighter on the grid.
Frontier model
The largest, most general AI models at the cutting edge (e.g., GPT-5.5). Powerful, but they burn a lot of energy because they fire the whole model at every task.
SWE-bench (Verified)
A respected benchmark of real-world software-engineering tasks pulled from open-source projects. "Verified" is the human-checked subset. Jarvi3 scored 500/500 (independent verification in submission).
Per-query
Measured for a single request. "0.11 kWh per 1,000 queries" means the energy used across one thousand individual questions.
kWh / Wh
Kilowatt-hours and watt-hours — units of energy. A typical kettle boil is roughly 0.1 kWh. It's how we measure the electricity an AI task consumes.
CO₂ / CO₂e
Carbon dioxide (and "equivalent" greenhouse gases) — the emissions tied to the electricity and water an AI task uses. Less energy means less CO₂.
Zero data exposure
Your data isn't sent to or stored on outside servers. Because Jarvi3 can run locally, what you put in stays with you.
Superalignment
Industry term for the research challenge of keeping AI that's smarter than humans safe and under control. (See our Safety page.)
Right to Warn
The principle (and a 2024 open letter) that AI staff should be free to raise safety concerns publicly without retaliation. EcoKure backs it.
Spotted a term we haven't explained? Tell us and we'll add it.