WEBVTT

00:00:00.300 --> 00:00:01.631
Does adding review make

00:00:01.631 --> 00:00:02.829
an AI worker complete

00:00:02.829 --> 00:00:04.759
more business tasks correctly?

00:00:05.059 --> 00:00:06.445
We tested Grex against

00:00:06.445 --> 00:00:07.758
a direct runner using

00:00:07.758 --> 00:00:08.853
the same model and

00:00:08.853 --> 00:00:10.093
simulated tools.

00:00:11.050 --> 00:00:12.708
The original protocol asked

00:00:12.708 --> 00:00:14.090
whether Grex adds value

00:00:14.090 --> 00:00:15.679
beyond its worker model.

00:00:15.979 --> 00:00:18.063
Our reconstructed hypothesis was

00:00:18.063 --> 00:00:19.213
that review and one

00:00:19.213 --> 00:00:21.369
correction might improve outcomes

00:00:21.369 --> 00:00:22.806
under shared limits.

00:00:23.758 --> 00:00:24.851
30 tasks covered

00:00:24.851 --> 00:00:26.126
6 business domains.

00:00:26.426 --> 00:00:27.532
Each arm received an

00:00:27.532 --> 00:00:29.157
identical starting world.

00:00:29.457 --> 00:00:31.379
An independent evaluator graded

00:00:31.379 --> 00:00:33.472
completed states after execution,

00:00:33.472 --> 00:00:34.913
without feeding answers

00:00:34.913 --> 00:00:35.942
to the workers.

00:00:36.925 --> 00:00:38.104
Both arms completed

00:00:38.104 --> 00:00:39.735
13 tasks correctly,

00:00:39.735 --> 00:00:41.746
on time and without assistance.

00:00:42.046 --> 00:00:43.122
Grex took about twice

00:00:43.122 --> 00:00:44.137
the elapsed time.

00:00:44.437 --> 00:00:45.714
No correction rescued

00:00:45.714 --> 00:00:46.656
a failed task.

00:00:47.633 --> 00:00:48.479
The audit found

00:00:48.479 --> 00:00:49.487
a crucial defect:

00:00:49.487 --> 00:00:52.284
corrections never received reviewer feedback.

00:00:52.584 --> 00:00:54.395
Reviews also exceeded limits,

00:00:54.395 --> 00:00:57.128
and corrections repeated simulated writes.

00:00:57.428 --> 00:00:58.497
This did not cleanly

00:00:58.497 --> 00:00:59.881
test the intended loop.

00:01:00.842 --> 00:01:01.786
We kept the result

00:01:01.786 --> 00:01:03.423
and repaired the mechanics.

00:01:03.723 --> 00:01:05.164
The November 4 rerun

00:01:05.164 --> 00:01:06.638
is a separate experiment,

00:01:06.638 --> 00:01:07.292
with its own

00:01:07.292 --> 00:01:08.864
outcome and limitations.

00:01:09.164 --> 00:01:10.206
Neither study used

00:01:10.206 --> 00:01:11.703
live customer accounts.