WEBVTT

00:00:00.300 --> 00:00:02.262
After the first extraction study,

00:00:02.262 --> 00:00:03.724
we tested whether keeping

00:00:03.724 --> 00:00:05.187
less worker history could

00:00:05.187 --> 00:00:07.315
improve efficiency without reducing

00:00:07.315 --> 00:00:07.980
quality.

00:00:08.280 --> 00:00:10.074
This was a separate comparison,

00:00:10.074 --> 00:00:12.227
with a revised extraction contract.

00:00:13.217 --> 00:00:13.904
One arm kept

00:00:13.904 --> 00:00:15.073
its full history.

00:00:15.373 --> 00:00:16.379
The other retained

00:00:16.379 --> 00:00:17.510
recent reasoning and

00:00:17.510 --> 00:00:18.957
shortened tool results.

00:00:19.257 --> 00:00:20.371
Both used the same

00:00:20.371 --> 00:00:21.634
worker model and the

00:00:21.634 --> 00:00:23.417
same 6 document families.

00:00:24.383 --> 00:00:25.820
Each family received 3

00:00:25.820 --> 00:00:26.879
slots per arm.

00:00:27.179 --> 00:00:28.313
The evaluator scored

00:00:28.313 --> 00:00:29.825
anonymous artifacts before

00:00:29.825 --> 00:00:31.211
learning their source.

00:00:31.511 --> 00:00:33.080
Slots and execution attempts

00:00:33.080 --> 00:00:34.837
were counted separately because

00:00:34.837 --> 00:00:36.845
some attempts needed replacement.

00:00:37.800 --> 00:00:39.412
Full history produced

00:00:39.412 --> 00:00:40.685
17 accepted slots

00:00:40.685 --> 00:00:41.533
out of 18.

00:00:41.833 --> 00:00:43.711
Trimmed history produced 13.

00:00:44.011 --> 00:00:45.665
The trimmed arm accumulated

00:00:45.665 --> 00:00:47.043
6 deadline failures and

00:00:47.043 --> 00:00:48.422
2 outputs with critical

00:00:48.422 --> 00:00:49.387
regressions.

00:00:50.342 --> 00:00:51.504
8 quota interruptions

00:00:51.504 --> 00:00:52.543
were excluded under

00:00:52.543 --> 00:00:54.011
the infrastructure rule.

00:00:54.311 --> 00:00:56.295
Deadline failures remained visible.

00:00:56.595 --> 00:00:58.096
Looking only at completed

00:00:58.096 --> 00:00:59.528
artifacts would make the

00:00:59.528 --> 00:01:01.097
trimmed policy appear more

00:01:01.097 --> 00:01:02.462
reliable than it was.

00:01:03.425 --> 00:01:04.524
We kept full history

00:01:04.524 --> 00:01:05.494
as the default.

00:01:05.794 --> 00:01:06.846
A later worker pool

00:01:06.846 --> 00:01:08.556
run used different conditions

00:01:08.556 --> 00:01:09.805
and stays separate.

00:01:10.105 --> 00:01:12.435
Neither result establishes sustained

00:01:12.435 --> 00:01:13.989
throughput or lower total

00:01:13.989 --> 00:01:15.118
operating cost.