Pick one under Drawings or upload your own. When you ask a question, counted items are boxed and numbered here and measured rooms are filled with their area.
Reading the drawing
Loading the model geometry
Drawing
Compare
Same drawing, same question: a large vision model alone, and Blueprint Evidence. Prepared examples, no live model calls.
The same test questions sent straight to a general vision model with the drawing attached, scored the same way, with cost per question. It appears here when out/baseline_nemotron.json exists.
Against other hosted models
small samples, direction not precision
By source and question type
Fine print: every test set, cost per route, witnesses, by object
Results by test set
Each number names the questions it was measured on and the file it was read from. The other tables show one eval run at a time.
Cost and latency per route
Loading
Witness agreement and correctness
Loading
By object
Accuracy counts a question it cannot answer (no number given) as wrong. Coverage is the share of questions answered. Precision is correct answers divided by answered ones. Counts must be exact; areas, lengths and volumes within 2%.