Same test.
Different box.

The source has two machine panels, each with response time, observed response and notes. Two times three is six editable fields. Equal bars show three fields for each machine at one hundred pixels per field. Nothing connects to either computer, runs LM Studio or measures latency. The names and forty eight GPU offload context are source labels to accompany your manual observations, not verified hardware results. Suppose you manually enter one point eight seconds and two point one seconds. On a common illustrative scale of two hundred pixels per second, these would be three hundred sixty and four hundred twenty pixels. Their difference would be zero point three seconds. The source does not parse either text field, subtract times or rank machines. You can enter arbitrary text and the exported time preserves that text. This hypothetical drawing is not an actual measured comparison. Filling all six fields saves six manual values locally. Reset deletes those six keys, leaving an empty local object and blank fields. Bars share eighty pixels per field to show six entries before reset and zero after. Export creates two machine objects with three text values each, a current creation timestamp and the fixed context and observation only scope. That timestamp marks export time rather than the measurement time. No hardware performance conclusion is generated.