Observed arrival · 2026-10-11
Semantic Observatory puts model decisions under the microscope
An independent research pilot compares how language models handle meaning across seven-position conversation arcs, with measured results, input formats, and failure modes.
- For
- People comparing language-model decision behavior
- Worth noticing
- The Jev run uses the same 119 inputs, with its hosted API adapter recorded separately from local native-tree scoring.
Field notes
The evidence view is organized around a strength × clarity matrix: readers can select a cell, compare the underlying sentence readings, and inspect a seven-position arc. The October 2026 research preview says its measurements are saved rather than generated live on the site. A Jev run uses 119 inputs; its hosted API adapter is documented separately from local native-tree scoring. The homepage also links to methods, limitations, input formats, field notes, and a pilot-measurement download.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue