Model-agnostic by design
The models will keep changing. The map contract will not. How a thin routing layer and a strict schema let us swap providers without you noticing.
Nodemeadow talks to language models through a layer about two hundred lines long. It speaks plain HTTP to each provider, asks for JSON, and validates what comes back against one schema: clusters, nodes with cited sources, connections with reasons, tensions, open questions. The schema is the product. The model is a supplier.
That shape lets us route. A dump goes to the primary provider; if that call fails or times out, it goes to the second. Studio accounts get priority routing. When a provider ships a better or cheaper model, we change one environment variable and run our evaluation set: three hundred real dumps with hand-checked clusters. If the new model cites correctly and clusters at least as well, it ships.
It also lets us be honest about limits. The validator does not care how confident a model sounds. If a node has no source line, it is gone. If a connection points to nothing, it is gone. We would rather show you a slightly sparser map that is entirely yours.
Finally, it means Nodemeadow does not depend on any one company's roadmap. Your maps are stored as plain JSON with your fragments beside them, exportable as markdown at any time. If every model vendor disappeared tomorrow, your thinking would still be here, cited and readable.