Same preview. Different target.

A 32-second, silent English replay of actual Open-Jev-27B-v1.1 outputs. No new inference or browser actions.

0–3s: Same preview. Different target. A replay of actual Open-Jev-27B-v1.1 predictions.

3–7s: Option A and Option B both satisfy the four visible requirements.

7–11s: Fewer interactions wins. Recorded selection: Option A (item_63), 99.42% model probability.

11–15s: Option A now targets record_B; its preview remains record_A. Recorded option order also differs.

15–20s: Option A no longer meets the target-match rule. All other candidate facts are unchanged.

20–24s: The second recorded choice is Option B, K11, at 95.62 percent model probability.

24–29s: These are two selected synthetic text-state cases. No browser actions were executed. Edited playback is not inference latency.

29–32s: Open-Jev-27B-v1.1. Inspect the evidence. Model, code and data are linked.

The two options arrays are retained exactly. Stable display aliases align the charts for reading and do not change model inputs. All six options and abstention are visible. Visible rule explanations come from the supplied policy; they are not model-generated reasoning.