FIELD NOTE / TIKTOK
Rename the labels. Does the policy flip?
The short film, the complete written thought, and the evidence behind it.
The TikTok edition will be linked here after its public post is verified.
Rename the labels. Does the policy flip?
Video caption
Rename the labels. Does the policy flip? The API type stays sound while the selected business action changes. Block rollout when labels change the consequential answer. #EricFieldNotes
Full written post / accessible read
Here is a bounded test I would run before putting Jev behind a consequential decision. Keep the customer record and written policy fixed. Change only the option names, from familiar words like approve and deny to neutral identifiers tied to the same criteria.
A September preprint reports this class of option-name effect in its tested decision heads, including the hosted Jev model, while type errors remain absent. That result is about the paper's tasks; it does not tell us the flip rate in your customer workflow.
Run the original and neutral-label forms over independently adjudicated cases. Log model version, exact state, rubric, option names, selected answer and downstream effect. Compare answer flips with a same-prompt repeat baseline so normal variability is not mistaken for label sensitivity.
If a name swap changes high-stakes outcomes beyond your accepted tolerance, revise the rubric, route the case to a stronger check or keep it human-owned. Do this because a grammar can close the output space; the label-swap canary tests whether the model followed the actual decision rule.
#EricFieldNotes
Evidence and boundary
On-screen label: NEW PREPRINT · TEST PROPOSAL. Illustrative cases are not measured incidents. Research papers and vendor documents support the stated mechanism only within their studied or documented scope.