1
00:00:00,256 --> 00:00:03,499
These two statements describe the same loan, currency, and

2
00:00:03,499 --> 00:00:06,187
reporting date. One says the interest is one hundred

3
00:00:06,187 --> 00:00:08,881
dollars. The other says 125.

4
00:00:08,951 --> 00:00:12,279
Paper Trail Agent finds that $25 contradiction and shows

5
00:00:12,279 --> 00:00:13,740
where both numbers came from.

6
00:00:13,815 --> 00:00:16,513
This is a real browser session with fictional PDFs.

7
00:00:16,588 --> 00:00:18,981
The public demo uses deterministic checks.

8
00:00:19,091 --> 00:00:22,323
Paid Meta calls and arbitrary uploads are blocked for guests.

9
00:00:22,377 --> 00:00:24,862
Open the interest finding and select each source.

10
00:00:24,979 --> 00:00:27,521
The original statement page highlights the cited line.

11
00:00:27,653 --> 00:00:30,725
Principal and interest remain separate fields, and statements from

12
00:00:30,725 --> 00:00:34,277
different periods are sent for review instead of being called contradictory.

13
00:00:34,437 --> 00:00:36,133
Now try the wrong-total case.

14
00:00:36,229 --> 00:00:39,175
The components add to $3,300.

15
00:00:39,369 --> 00:00:41,854
The statement claims 3,350.

16
00:00:41,929 --> 00:00:45,459
Python Decimal recomputes the total and confirms a $50

17
00:00:45,470 --> 00:00:48,585
difference. The model does not get to invent the answer.

18
00:00:48,778 --> 00:00:52,746
The internal trace separates extraction, verification, and an independent

19
00:00:52,746 --> 00:00:56,074
critic. Before a finding is confirmed, the critic checks

20
00:00:56,074 --> 00:00:58,787
the amount, calculation coverage, and evidence.

21
00:00:58,961 --> 00:01:01,435
After checking the source, a person can approve or

22
00:01:01,435 --> 00:01:03,323
reject the finding and save a note.

23
00:01:03,441 --> 00:01:05,819
That decision is preserved with the report version.

24
00:01:05,958 --> 00:01:07,953
The scanned case takes a different path.

25
00:01:08,006 --> 00:01:10,543
Its image-only page has no searchable text.

26
00:01:10,669 --> 00:01:13,794
The app requests review and confirms no discrepancy.

27
00:01:13,911 --> 00:01:16,098
Missing evidence is never a clean result.

28
00:01:16,173 --> 00:01:19,757
Authenticated OCR is implemented and tested, but its separate

29
00:01:19,757 --> 00:01:21,869
worker is not enabled in the public demo.

30
00:01:21,955 --> 00:01:24,387
Here is a completed local Muse Spark investigation on

31
00:01:24,387 --> 00:01:26,040
the same fictional statements.

32
00:01:26,222 --> 00:01:28,995
These are actual tool calls through a case-scoped MCP

33
00:01:28,995 --> 00:01:32,222
server: evidence retrieval and an exact calculator.

34
00:01:32,313 --> 00:01:36,025
Inputs, outputs, validation, and the bounded tool budget

35
00:01:36,025 --> 00:01:39,481
are visible. This is recorded live-model evidence, separate

36
00:01:39,481 --> 00:01:40,943
from the public guest flow.

37
00:01:41,145 --> 00:01:43,682
The evaluation panels make the limits inspectable.

38
00:01:43,809 --> 00:01:47,094
Six baseline cases and 19 varied-layout cases have known

39
00:01:47,094 --> 00:01:50,678
answers. The agent scorecard shows field accuracy, controlled critic

40
00:01:50,678 --> 00:01:53,839
faults, tool contracts, and retained failed runs.

41
00:01:53,949 --> 00:01:57,021
Retrieval precision is 50% in this small probe,

42
00:01:57,021 --> 00:01:58,760
and that weakness stays visible.

43
00:01:58,898 --> 00:02:02,824
Paper Trail Agent turns conflicting statements into cited, calculator-checked

44
00:02:02,824 --> 00:02:05,064
findings and a human review checklist.

45
00:02:05,150 --> 00:02:08,158
Try the synthetic cases and inspect the public code.