FIELD NOTE / LINKEDIN
Computer use needs a second witness.
The short film, the complete written thought, and the evidence behind it.
The LinkedIn edition will be linked here after its public post is verified.
Computer use needs a second witness.
Day 02 · 2026-09-29 · LinkedIn
Short video caption
Computer-use QA should record the target identity, postcondition and allowed delay before execution. Separate the UI actor from a fresh-session grader; report PASS, FAIL and UNVERIFIED. Benchmark evidence does not replace your app state. #EricFieldNotes
Full written post / accessible read
Browser and desktop models make end-to-end QA possible at a useful scale. The hardest part is not producing the click trace. It is deciding whether the real user outcome happened, in the right account, after the run ends.
A model submits a form and shows the confirmation page. The backend may still reject the queued job, attach it to another identity, or leave the result invisible to the recipient. A transcript alone cannot settle those states.
Before the run, record the exact user, requested transition, allowed delay and evidence sources. Give the actor the UI. Give the grader a fresh session and read-only system observation. Count PASS, FAIL and UNVERIFIED separately.
Set a postcondition for every consequential UI action and require a fresh readback after the allowed delay. Block a success claim when the probe never ran. This turns computer-use QA from a screen recording into an accountable result.
#EricFieldNotes
Evidence and boundary
On-screen boundary: EVALUATION DESIGN. The sources below support documented mechanisms and specifications; illustrative scenarios are not presented as measured incidents.