Lead with the useful action.
Put the important next step near the beginning. When a recommendation is justified, explain which approach fits the situation and why.
THE OFFGRID AI INTELLIGENCE LAYER
Knowledge you can use.
A practical next step.
An equipment problem. A plan for your property. A question far from a connection. Useful guidance needs to account for your situation, the resources you have, and what matters first.
The OffGrid AI Intelligence Layer guides the included models toward clear priorities, practical steps, and honest uncertainty. It is built into Core Edition and Map Edition, ready to work offline on your compatible computer.
Hear the thinking behind the Intelligence Layer.
BUILT AROUND YOUR SITUATION
What matters first.
What you can do with what you have.
The layer provides persistent instructions that shape each response. These are the behaviors we design for and evaluate.
Work with the real constraints.
Consider the tools, supplies, equipment, time, and conditions you describe. Organize complex tasks into steps you can follow.
Make the unknowns visible.
Identify missing details, relevant risks, and facts that need checking. Explain when a different condition could change the recommendation.
The knowledge is already included. The AI arrives with broad learned knowledge. You do not need to build a library or write special instructions before asking a question. Your saved answers and personal references are an optional way to make the ToolKit more useful to you.
HOW THE LAYER WORKS
The model brings knowledge.
The layer guides the response.
The Intelligence Layer is a set of instructions supplied to the model with your question. It guides behavior rather than retraining the model or adding new facts to its learned knowledge.
In the offline ToolKit, both the model and these instructions run locally. Your offline questions and photos are processed on your computer. No subscription or cloud login is required for that use.
A self-check instruction is not an independent fact check. AI can still make mistakes, and a confidence statement is not a measured probability of correctness. Check consequential guidance against trusted references and your actual conditions.
Give the answer something to work with.
-
Describe the problem.
Include the goal, surroundings, equipment, and what has already happened.
-
Name the constraints.
Mention available tools or supplies, time limits, and the details that could change the answer.
-
Review and keep what helps.
Check the recommendation, verify important details, and save useful guidance to your personal Knowledge Base.
WHERE THE IDEA TOOK SHAPE
An earlier experiment changed our direction.
During Gemma 3 development, an online comparison tested responses with and without a self-review instruction added to the question. The April 2026 development work helped us see the value of shaping how an answer is delivered, and informed the integrated Intelligence Layer that followed.
~14.25 / 20
Estimated average composite score reported without the added self-check guidance.
~18.25 / 20
Estimated average composite score reported with the additional guidance.
+4 points
About a 28% relative increase in the reported rubric score across five historical batches.
Historical Gemma 3 development context. These are approximate summary scores from an earlier online instruction comparison, not scores measured from the current offline ToolKit. They helped shape the Intelligence Layer. The reported score increase is not a percentage increase in factual accuracy.
THE TESTING CONTINUES
A stronger foundation.
More questions worth testing.
As the ToolKit moved to newer Gemma 4 models, the research continued. We built arenas to compare OffGrid answers with other models, then introduced matched comparisons using the same Gemma foundation with and without the added OffGrid instructions.
The later results are more nuanced than a single improvement number. The layer received higher rankings on some scenarios, while baseline responses won others. Results also varied by model size, question, and evaluator.
That is useful information for development. We retain original answers, wins, ties, losses, and unsuccessful experiments so the next change can be judged on evidence. Experimental instruction sets are kept separate from the version supplied to customers.
The Arena runs online. Its hosted comparisons are research evidence, not a guarantee that every offline answer or computer will produce the same result.
INCLUDED IN BOTH EDITIONS
Your own intelligence system.
Ready when a question comes up.
Core Edition includes offline AI, photo analysis, ready-made prompts, and your personal Knowledge Base. Map Edition adds preloaded offline maps covering the United States.
The Intelligence Layer is part of both editions and both AI modes. Choose the separate Windows or Mac build that matches your compatible computer.