WEBVTT

00:00:00.300 --> 00:00:01.763
Grex runs AI assignments

00:00:01.763 --> 00:00:03.436
and keeps their evidence.

00:00:03.736 --> 00:00:05.600
This research experiment asked

00:00:05.600 --> 00:00:06.842
whether at least 4

00:00:06.842 --> 00:00:07.947
of 5 tasks could

00:00:07.947 --> 00:00:09.397
reach acceptance with no

00:00:09.397 --> 00:00:11.053
more than one correction.

00:00:12.008 --> 00:00:13.231
The plan used MiniMax

00:00:13.231 --> 00:00:14.453
to gather sources and

00:00:14.453 --> 00:00:16.083
write a structured brief.

00:00:16.383 --> 00:00:17.670
A separate Claude reviewer

00:00:17.670 --> 00:00:18.900
would judge the result.

00:00:19.200 --> 00:00:21.051
Every launched assignment allowed

00:00:21.051 --> 00:00:22.592
one automatic correction.

00:00:23.550 --> 00:00:24.757
The team repeated the

00:00:24.757 --> 00:00:26.232
panel across 3 rounds

00:00:26.232 --> 00:00:28.042
while changing platform builds

00:00:28.042 --> 00:00:29.182
and input limits.

00:00:29.482 --> 00:00:31.893
These repetitions diagnosed failures.

00:00:32.193 --> 00:00:33.626
They were not independent

00:00:33.626 --> 00:00:35.777
trials under identical conditions.

00:00:36.758 --> 00:00:38.762
9 worker executions produced

00:00:38.762 --> 00:00:40.193
no reviewable brief.

00:00:40.493 --> 00:00:41.727
2 more launches failed

00:00:41.727 --> 00:00:43.108
before the worker started,

00:00:43.108 --> 00:00:44.224
and 4 planned slots

00:00:44.224 --> 00:00:45.400
were never launched.

00:00:45.700 --> 00:00:47.003
No independent review

00:00:47.003 --> 00:00:48.238
or correction ran.

00:00:49.217 --> 00:00:50.466
Some outputs violated

00:00:50.466 --> 00:00:51.585
the brief schema.

00:00:51.885 --> 00:00:53.570
Others hit execution bounds.

00:00:53.870 --> 00:00:55.144
These failures explain

00:00:55.144 --> 00:00:56.386
why delivery stopped,

00:00:56.386 --> 00:00:57.150
but they do not

00:00:57.150 --> 00:00:58.551
establish that a reviewer

00:00:58.551 --> 00:01:00.462
judged the research inaccurate.

00:01:01.425 --> 00:01:02.567
The decision was to

00:01:02.567 --> 00:01:04.280
repair execution and output

00:01:04.280 --> 00:01:06.493
contracts before claiming research

00:01:06.493 --> 00:01:07.206
quality.

00:01:07.506 --> 00:01:09.981
Later successes need separate evidence.

00:01:10.281 --> 00:01:11.161
They cannot erase

00:01:11.161 --> 00:01:11.982
what happened in

00:01:11.982 --> 00:01:13.566
this historical experiment.