Model
Rowe 1.0
The spatial reasoning model.
One endpoint. Send a structured scene and a task type; get typed spatial output, usage, and a request id back. Rowe reasons over geometry that has been kept intact rather than described in prose.
curl https://fourechelon.com/api/v1/responses \
-H "Authorization: Bearer $FOUR_ECHELON_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "rowe-1.0",
"input": {
"task_type": "fit",
"object_id": "case",
"secondary_id": "shelf"
}
}'- Tasks: understand, compare, fit, stability, placement, embed, predict, act, optimize
- Inputs: cad_scene, point_cloud_scene, images, sensor readings, goals
- Output: typed JSON — measurements, relations, issues, overlays
- Billed per input and output spatial token; failed requests are never billed
What makes it different
A vision-language model sees a rendered picture of your part. Rowe sees the part. Solids, faces, dimensions, and constraints arrive as structured input, so an answer about a 4 mm radius is computed from a 4 mm radius rather than inferred from pixels.
What you get back
Every response is typed and traceable. Outputs carry the geometry they were derived from, usage carries the exact token counts you are billed for, and request_id carries through to support.
- output — an array of typed spatial results for the task you asked for
- usage — input and output spatial tokens plus projected cost
- request_id — quote this and we can find the exact call
How we talk about accuracy
We publish our evaluation methodology and error bars, and we correct our own numbers in public when a harness turns out to have been unfair. We do not publish head-to-head scores against named competitors, because the ones we ran did not survive scrutiny.
Also in the product line