WEBVTT

00:00:00.350 --> 00:00:07.790
Did the lorebook really trigger? Run a controlled test

00:00:08.120 --> 00:00:21.610
Result: three entries; row #9 is the 21-character question

00:00:21.990 --> 00:00:36.430
Prerequisite: one unique keyword; diagnostic messages may cost money

00:00:36.810 --> 00:00:47.250
Test card: two constant entries and two keyword entries

00:00:47.580 --> 00:00:56.570
Confirm 4 of 4 enabled; enabled is not triggered

00:00:56.950 --> 00:01:14.390
Design: omit the keyword, then add only Blue Umbrella

00:01:14.770 --> 00:01:27.210
Negative control: the prompt omits Blue Umbrella

00:01:27.590 --> 00:01:41.030
Negative evidence: Prompt Inspector shows two entries

00:01:41.410 --> 00:01:52.850
Negative control holds: no keyword, no target injection

00:01:53.230 --> 00:02:05.670
Positive test: add only Blue Umbrella

00:02:06.050 --> 00:02:19.490
Positive evidence: two to three; row #9 is that question

00:02:19.870 --> 00:02:37.310
Successful injection does not guarantee exact model compliance

00:02:37.690 --> 00:02:57.130
Troubleshoot: enabled → keyword → gates → Inspector; redact before sharing

00:02:57.510 --> 00:03:10.950
Unique keyword → negative → positive → inspect the request

00:03:11.280 --> 00:03:16.750
Foreverse · Guide and Android download in the description