JournalDAY 02 / X

FIELD NOTE / X

A screenshot is not an assertion.

The short film, the complete written thought, and the evidence behind it.

Journal September 25, 2026 · X target September 29, 2026
Watch the verified YouTube copy ↗

The X edition will be linked here after its public post is verified.

A screenshot is not an assertion.

Day 02 · 2026-09-29 · X

Short video caption

A computer-use screenshot proves pixels. For consequential changes, define the target account and state before the run, then verify a fresh session and independent readback. Illustrative save example. #EricFieldNotes

Full written post / accessible read

Computer-use agents can explore a UI faster than a manual test script. But a screenshot of the final screen proves pixels, not that the intended record persisted, the correct account changed, or another user can see the result.

In an illustrative settings change, the app shows Saved. A queued write then fails. The agent reports success from the toast while a cold reload restores the old value. No model defect is required; the test used the wrong oracle.

Give the actor the UI task. Then a separate observer opens a fresh session, checks the exact account and persisted value, and compares it with a read-only system record where available. Keep the expected state outside the actor's transcript.

For every consequential UI test, state the account, target value and persistence boundary first. Require cold-session and independent state checks before PASS. Do this because a plausible final frame is only evidence of what the agent saw.

#EricFieldNotes

X thread draft

Attach the video to the first post. The subsequent text adds detail; the full read above is also the accessible transcript. Review the thread in the live composer before sending.

A computer-use screenshot proves pixels. For consequential changes, define the target account and state before the run, then verify a fresh session and independent readback. Illustrative save example. #EricFieldNotes

In an illustrative settings change, the app shows Saved. A queued write then fails. The agent reports success from the toast while a cold reload restores the old value. No model defect is required; the test used the wrong oracle.

Give the actor the UI task. Then a separate observer opens a fresh session, checks the exact account and persisted value, and compares it with a read-only system record where available. Keep the expected state outside the actor's transcript.

For every consequential UI test, state the account, target value and persistence boundary first. Require cold-session and independent state checks before PASS. Do this because a plausible final frame is only evidence of what the agent saw.

Evidence and boundary

On-screen boundary: ILLUSTRATIVE UI TEST. The sources below support documented mechanisms and specifications; illustrative scenarios are not presented as measured incidents.

More notes from the work ↗