OrgLens: readable reports and the limits of their evidence
Doug Belshaw’s second OrgLens development account describes moving from an agent-driven experiment to a fixed reporting pipeline.
The reported design separates structural checks from model-generated writing. It checks quotations against notes, tests for contradictions, and moves specialist terminology out of the main report. These are claims about the experimental design and its tests.
The cases were fictional and the readers were simulated. The same AI setup contributed to construction and evaluation, with small samples. Report clarity under those conditions does not establish whether diagnoses are fair or useful to people in a real organisation.
Inspect OrgLens and its claimed systems lenses · Read the maintained source repository
Editorial application: inspect the evidence chain
- Can a reader trace each claim to the submitted account?
- Does a structural check test the claim, or only the consistency of its representation?
- Can the people described dispute the interpretation and decide what happens next?
A readable report can still encode a weak or inappropriate model. Human review must address the interpretation as well as the quotations.