FIELD NOTE / TIKTOK
This computer-use demo looks finished.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved TikTok edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
This computer-use demo looks finished.
Video caption
This computer-use demo looks finished. The UI and its test shared an optimistic view. Do this because a replay is not the outcome. #EricFieldNotes
Full written post / accessibility read
Watch a fictional test account. The agent changes a role, presses Save and gets the green toast. If this were the whole demo, I would believe it. Now close the session and read the authoritative role.
A delayed policy check rejects the requested role. The first browser session still rendered it briefly. This is an oracle design problem: the navigator and grader looked at the same temporary evidence surface.
The harness tags the request ID, waits for policy settlement, reads the audit state through a read-only route and logs in as the affected fictional user. Both disagree with the green screenshot, so the run is a failure.
Deliberately inject this delayed rejection in a resettable test tenant. If the suite still passes, repair the acceptance oracle before buying a stronger model. Better clicking will not fix a test that asks the wrong question.
#EricFieldNotes
Four-beat scene transcript
1. This computer-use demo looks finished.
Watch a fictional test account. The agent changes a role, presses Save and gets the green toast. If this were the whole demo, I would believe it. Now close the session and read the authoritative role.
Visual: The saved role disappears when we reopen it.
2. The agent did not necessarily click wrong.
A delayed policy check rejects the requested role. The first browser session still rendered it briefly. This is an oracle design problem: the navigator and grader looked at the same temporary evidence surface.
Visual: The UI and its test shared an optimistic view.
3. Show the correction path.
The harness tags the request ID, waits for policy settlement, reads the audit state through a read-only route and logs in as the affected fictional user. Both disagree with the green screenshot, so the run is a failure.
Visual: Retain the trace and check a separate record.
4. Make the false green turn red.
Deliberately inject this delayed rejection in a resettable test tenant. If the suite still passes, repair the acceptance oracle before buying a stronger model. Better clicking will not fix a test that asks the wrong question.
Visual: Do this because a replay is not the outcome.
Research and claim limits
- OpenAI Computer use API guide (S159)
- Playwright Tracing API (S163)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.