1
00:00:00,077 --> 00:00:02,123
An AI refund assistant can learn

2
00:00:02,143 --> 00:00:04,763
the wrong lesson chase good ratings

3
00:00:04,783 --> 00:00:07,603
and quietly weaken the refund rules

4
00:00:08,177 --> 00:00:10,529
We built SAAF Circuit Breaker a

5
00:00:10,549 --> 00:00:13,157
safety layer outside the assistant so

6
00:00:13,177 --> 00:00:16,163
suggestions cannot override business limits

7
00:00:16,807 --> 00:00:19,147
Forge inspects the agent and produces

8
00:00:19,167 --> 00:00:21,592
a risk report and rule files

9
00:00:22,236 --> 00:00:25,673
Rehearsal plays thirty-six sample tickets With

10
00:00:25,693 --> 00:00:28,504
protection off the scripted scenario shows

11
00:00:28,524 --> 00:00:31,919
rules eroding and unauthorized refunds

12
00:00:32,536 --> 00:00:34,506
Switch protection on and the same

13
00:00:34,526 --> 00:00:37,095
scenario shows the guard containing those

14
00:00:37,115 --> 00:00:38,137
violations

15
00:00:38,757 --> 00:00:41,148
Now a live test twenty-one euros

16
00:00:41,168 --> 00:00:43,941
thirty-five no photo The hard guard

17
00:00:43,961 --> 00:00:46,467
blocks it Real Jev checks the

18
00:00:46,487 --> 00:00:48,682
instructions and saved lessons

19
00:00:49,291 --> 00:00:51,529
When unsafe rules are detected the

20
00:00:51,549 --> 00:00:54,692
local runtime restores safe instructions and

21
00:00:54,712 --> 00:00:57,128
memory and saves the decision for

22
00:00:57,148 --> 00:00:57,992
human review

23
00:00:58,604 --> 00:01:01,377
Verify rechecks the saved decision and

24
00:01:01,397 --> 00:01:04,030
rules without another model call Change

25
00:01:04,050 --> 00:01:06,582
the recorded verdict and the consistency

26
00:01:06,602 --> 00:01:07,802
check fails

27
00:01:08,404 --> 00:01:10,795
This is a working local prototype

28
00:01:10,815 --> 00:01:13,971
real Jev and containment scripted rehearsal

29
00:01:13,991 --> 00:01:15,654
and no real payments Let the

30
00:01:15,674 --> 00:01:18,801
AI propose Keep policy in charge