1
00:00:00,350 --> 00:00:03,500
MirrorCheck gives familiar feeds a clearer signal.

2
00:00:03,500 --> 00:00:09,162
Our new landing page introduces the product with a moving gallery and liquid glass navigation.

3
00:00:09,162 --> 00:00:15,662
The goal is simple: turn suspected profile copies into evidence, while keeping the decision human.

4
00:00:18,350 --> 00:00:22,237
This is the original profile, and this is a new fictional copy.

5
00:00:22,237 --> 00:00:25,906
The display name, biography, avatar and photos are

6
00:00:25,906 --> 00:00:29,575
reused, while the username and profile statistics differ.

7
00:00:29,575 --> 00:00:32,100
Neither screenshot proves account ownership.

8
00:00:34,350 --> 00:00:39,637
From the suspected copy, I choose Report account, then Pretending to be someone else.

9
00:00:39,637 --> 00:00:42,287
I upload screenshots of both profiles.

10
00:00:42,287 --> 00:00:47,375
PaddleOCR extracts the visible names, biographies and counts.

11
00:00:47,375 --> 00:00:53,900
I compare those fields with the screenshots, correct anything misread, and acknowledge the review before submitting.

12
00:00:58,350 --> 00:01:00,837
Privacy comes before the model request.

13
00:01:00,837 --> 00:01:07,762
Presidio detects the example contact details locally, and the server masks those regions in the images.

14
00:01:07,762 --> 00:01:10,637
I can inspect the exact processed previews.

15
00:01:10,637 --> 00:01:15,437
The moderator later sees the same processed evidence that was prepared for Meta AI.

16
00:01:18,350 --> 00:01:22,325
Meta Muse Spark compares the processed screenshots through the backend.

17
00:01:22,325 --> 00:01:25,875
This is a real, fresh response from the recorded case.

18
00:01:25,875 --> 00:01:30,887
It returns visible similarities, differences, and a review recommendation.

19
00:01:30,887 --> 00:01:34,618
The interface separates Meta's visual score, the combined

20
00:01:34,618 --> 00:01:38,350
score, text reuse ML, and profile spam patterns.

21
00:01:38,350 --> 00:01:41,537
These are signals, not identity probabilities.

22
00:01:42,350 --> 00:01:45,900
The report remains under review until a moderator acts.

23
00:01:45,900 --> 00:01:50,194
In the protected review queue, I inspect the processed images,

24
00:01:50,194 --> 00:01:54,487
matching photo regions, the model's observations, and its limitations.

25
00:01:54,487 --> 00:01:57,162
Confirmation is a deliberate human action.

26
00:01:57,162 --> 00:02:00,937
Here, the moderator confirms this fictional copied account.

27
00:02:04,350 --> 00:02:06,587
The decision is saved on the server.

28
00:02:06,587 --> 00:02:10,025
I return to the reported profile and reload the page.

29
00:02:10,025 --> 00:02:12,937
The impersonation warning remains visible.

30
00:02:12,937 --> 00:02:18,587
An analysis score alone never creates this label; the moderator's saved decision does.

31
00:02:22,350 --> 00:02:24,475
The opposite outcome is possible too.

32
00:02:24,475 --> 00:02:28,475
This earlier fictional report concerns an ordinary profile.

33
00:02:28,475 --> 00:02:31,762
The moderator dismissed it, and the account has no warning.

34
00:02:31,762 --> 00:02:37,175
Shared stock photographs can create visual overlap without establishing impersonation.

35
00:02:40,347 --> 00:02:42,756
MirrorCheck is a working local prototype.

36
00:02:42,756 --> 00:02:46,132
Its text model uses synthetic training pairs, and

37
00:02:46,132 --> 00:02:49,508
performance on real account copies still needs independent evaluation.

38
00:02:49,508 --> 00:02:51,651
Human review remains essential.

39
00:02:52,350 --> 00:02:54,912
Explore the prototype and both outcomes.

40
00:02:54,912 --> 00:02:57,987
MirrorCheck: a second look, a human decision.