WEBVTT

00:00:00.300 --> 00:00:02.170
Could Grex find conversations

00:00:02.170 --> 00:00:03.608
worth joining and draft

00:00:03.608 --> 00:00:04.759
useful replies?

00:00:05.059 --> 00:00:06.996
The experiment proposed comparing

00:00:06.996 --> 00:00:08.093
it with Hermes using

00:00:08.093 --> 00:00:08.965
the same model,

00:00:08.965 --> 00:00:11.160
through controlled and field tracks.

00:00:12.133 --> 00:00:13.462
This study did record

00:00:13.462 --> 00:00:15.306
a directional hypothesis and

00:00:15.306 --> 00:00:17.446
numerical pass conditions before

00:00:17.446 --> 00:00:18.405
scored runs.

00:00:18.705 --> 00:00:20.831
It expected better recommendation

00:00:20.831 --> 00:00:22.887
precision and reply usability.

00:00:23.187 --> 00:00:25.428
Calibration was explicitly excluded

00:00:25.428 --> 00:00:26.899
from that formal test.

00:00:27.883 --> 00:00:29.161
The retained evidence

00:00:29.161 --> 00:00:30.977
shows Grex-side calibration.

00:00:31.277 --> 00:00:32.554
Searches fed relevance

00:00:32.554 --> 00:00:33.799
scoring and drafting,

00:00:33.799 --> 00:00:35.714
followed by a separate reviewer.

00:00:36.014 --> 00:00:37.570
The intended 7 matched

00:00:37.570 --> 00:00:38.991
comparison pairs did not

00:00:38.991 --> 00:00:40.546
produce a completed scored

00:00:40.546 --> 00:00:41.155
result.

00:00:42.133 --> 00:00:43.692
3 slots rediscovered the

00:00:43.692 --> 00:00:45.376
same 53 candidates.

00:00:45.676 --> 00:00:47.433
8 repeatedly reached drafting

00:00:47.433 --> 00:00:48.428
and were refused.

00:00:48.728 --> 00:00:50.292
A later brief mistakenly

00:00:50.292 --> 00:00:51.782
carried refusal text as

00:00:51.782 --> 00:00:52.490
its reply;

00:00:52.490 --> 00:00:54.278
independent review found it

00:00:54.278 --> 00:00:55.619
not worth reading.

00:00:56.592 --> 00:00:57.799
The source and evidence

00:00:57.799 --> 00:00:58.977
link could be checked,

00:00:58.977 --> 00:01:00.184
yet the recommendation

00:01:00.184 --> 00:01:01.392
was still unhelpful.

00:01:01.692 --> 00:01:04.183
These calibration observations cannot

00:01:04.183 --> 00:01:06.418
establish comparative precision,

00:01:06.418 --> 00:01:07.701
engagement uplift,

00:01:07.701 --> 00:01:09.093
customer demand or the

00:01:09.093 --> 00:01:11.292
cost of useful recommendations.

00:01:12.258 --> 00:01:13.466
We retain the selection

00:01:13.466 --> 00:01:15.036
and delivery failures without

00:01:15.036 --> 00:01:16.184
inventing a winner.

00:01:16.484 --> 00:01:18.006
A resumed comparison would

00:01:18.006 --> 00:01:19.727
need fresh bounded conditions

00:01:19.727 --> 00:01:21.646
and honest failure accounting.

00:01:21.946 --> 00:01:23.661
This historical record does

00:01:23.661 --> 00:01:25.018
not imply a restart.