1
00:00:00,300 --> 00:00:02,698
This tiny plant was labeled crop.

2
00:00:03,001 --> 00:00:05,008
The annotation says weed.

3
00:00:05,500 --> 00:00:08,112
GreenPoint makes that disagreement
inspectable.

4
00:00:10,275 --> 00:00:14,920
For growers reviewing scattered fields, a
label alone isn't enough.

5
00:00:15,285 --> 00:00:20,168
Our online flow uses Sam three point one
for plant regions, and Muse

6
00:00:20,207 --> 00:00:22,820
Spark for structured visual suggestions.

7
00:00:24,300 --> 00:00:28,995
Select a candidate. Its box connects to
pixels from that exact source

8
00:00:29,046 --> 00:00:33,500
image. Zoom and pan: the evidence stays
with the plant.

9
00:00:33,866 --> 00:00:36,264
Here, the model suggests weed.

10
00:00:36,629 --> 00:00:38,598
I keep it for review and save.

11
00:00:39,090 --> 00:00:42,246
The original suggestion and my decision
remain separate.

12
00:00:48,274 --> 00:00:51,132
Switch to the video inventory, then
export.

13
00:00:51,541 --> 00:00:53,426
The same review follows the session.

14
00:00:53,679 --> 00:00:58,494
The CSV preserves the original label,
human status, source,

15
00:00:58,590 --> 00:01:00,031
and run identity.

16
00:01:01,275 --> 00:01:03,484
Muse can request evidence too.

17
00:01:03,913 --> 00:01:08,734
In this separate recorded run, it
discovers our MCP tools and requests

18
00:01:08,809 --> 00:01:10,815
authenticated frame perception.

19
00:01:11,308 --> 00:01:12,986
Seven candidates come back.

20
00:01:13,352 --> 00:01:16,999
Muse cites evidence IDs and chooses human
review.

21
00:01:22,263 --> 00:01:27,033
We then tested a local agricultural
expert, using published PhenoBench

22
00:01:27,096 --> 00:01:31,817
weights. This is real CPU inference on the
same RGB frame,

23
00:01:32,003 --> 00:01:34,457
with no annotations sent to the model.

24
00:01:34,940 --> 00:01:38,806
It is a research prototype, separate from
the online Meta flow.

25
00:01:40,263 --> 00:01:44,970
On this selected example, the previous
pipeline missed a fifty-one-pixel

26
00:01:45,046 --> 00:01:48,996
weed. The agricultural detector recovers
its box.

27
00:01:49,298 --> 00:01:54,283
Both views show the same source pixels,
checked against the public annotation.

28
00:01:57,286 --> 00:02:01,636
We chose settings on twenty calibration
images, then evaluated forty different

29
00:02:01,683 --> 00:02:06,614
images. On that same comparison, weed
precision rose from seventy-six

30
00:02:06,697 --> 00:02:09,803
point one to ninety-six point seven
percent.

31
00:02:10,431 --> 00:02:13,004
Recall stayed near sixty-one percent.

32
00:02:13,287 --> 00:02:16,203
False weed flags fell from thirty-seven to
four.

33
00:02:17,288 --> 00:02:21,958
Tiles add fragments. Containment cleanup
raised tile precision from

34
00:02:22,033 --> 00:02:26,450
eighty-seven point two to ninety-two point
five percent, with sixty-three

35
00:02:26,500 --> 00:02:28,241
point three percent recall.

36
00:02:31,262 --> 00:02:33,191
The smallest weeds remain difficult:

37
00:02:33,524 --> 00:02:37,635
nineteen of seventy targets at thirty-two
pixels or less were recovered.

38
00:02:38,025 --> 00:02:39,747
Human review stays essential.

39
00:02:39,989 --> 00:02:44,617
This is public-data validation; movement
tracking and independent farm trials

40
00:02:44,697 --> 00:02:46,144
remain unvalidated.

41
00:02:47,262 --> 00:02:48,839
Inspectable Meta evidence.

42
00:02:49,046 --> 00:02:50,727
Traceable human decisions.

43
00:02:50,899 --> 00:02:53,329
A tested local recognition prototype.

44
00:02:53,721 --> 00:02:56,622
Next: cloud integration and farmer trials.

45
00:02:56,795 --> 00:02:59,258
GreenPoint. Every judgment has a frame.