WEBVTT

00:00:00.400 --> 00:00:05.448
In one of the Department of Agriculture's grant
programs, an application over its page limit,

00:00:05.519 --> 00:00:10.386
or outside its rules for margins
and type size, is returned without review.

00:00:10.827 --> 00:00:15.211
N S F says plainly that its submission
system doesn't check every rule.

00:00:15.646 --> 00:00:19.851
So we planted eight problems in a proposal,
and asked an agent to check it.

00:00:20.181 --> 00:00:22.023
It found eight of the eight.

00:00:24.400 --> 00:00:30.058
AI for Research Efficiency, In Practice:
check your proposal against the rules.

00:00:34.000 --> 00:00:34.762
In this film,

00:00:34.853 --> 00:00:39.180
an agent checks your own proposal against
the call and the funder's current guide,

00:00:39.230 --> 00:00:40.634
before you submit.

00:00:40.973 --> 00:00:43.320
It measures the files and quotes the rules.

00:00:43.671 --> 00:00:47.140
It never changes a file
or writes a word of the proposal.

00:00:47.511 --> 00:00:52.096
It builds on episodes four and nine,
on small habits and on your data.

00:00:55.600 --> 00:00:57.805
The proposal is invented for this film:

00:00:58.076 --> 00:01:02.405
an N S F CAREER proposal about rice,
water and greenhouse gases.

00:01:02.795 --> 00:01:06.199
In Codex, in plan mode,
the request is a question.

00:01:06.610 --> 00:01:10.385
What is the best way to check this proposal
against everything the solicitation

00:01:10.435 --> 00:01:13.378
and the funder require? Do not change it.

00:01:13.784 --> 00:01:20.228
It asks no questions. It reads the folder,
opens N S F's pages online, and proposes a plan.

00:01:24.400 --> 00:01:27.669
Then three helpers research
the options in parallel.

00:01:28.081 --> 00:01:31.625
One reads N S F's own list of
what its submission system stops,

00:01:31.796 --> 00:01:36.341
like a project description over fifteen
pages, or a web address inside it.

00:01:36.757 --> 00:01:39.964
Another looks for what that list
leaves out, like the letters,

00:01:40.095 --> 00:01:42.991
and whether the budget justification
agrees with the budget.

00:01:43.442 --> 00:01:47.590
A third works out how to measure the type,
the lines and the margins.

00:01:48.021 --> 00:01:50.668
That second list is
where the agent helps most.

00:01:50.979 --> 00:01:53.927
The first, it catches a week
before the deadline.

00:01:58.000 --> 00:02:00.648
A reviewer agent then attacks the plan.

00:02:01.023 --> 00:02:02.224
The rules move.

00:02:02.654 --> 00:02:06.447
N S F's guide dates from twenty twenty-four
and has been amended twice,

00:02:06.578 --> 00:02:08.189
and a replacement is planned.

00:02:08.579 --> 00:02:12.923
Its published list of automatic checks dates
from January twenty twenty-three.

00:02:13.313 --> 00:02:18.099
So the process reads the current documents
every time, and names their versions.

00:02:18.570 --> 00:02:20.893
A made-up rule is worse than a missed one,

00:02:21.064 --> 00:02:24.708
so every finding has
to quote its rule, word for word.

00:02:25.109 --> 00:02:27.352
And the agent writes none of the proposal.

00:02:27.903 --> 00:02:30.878
N I H won't count applications
substantially developed

00:02:30.928 --> 00:02:33.892
by A I as the applicant's original ideas.

00:02:38.800 --> 00:02:41.887
You choose the plan,
and the agent writes it down as a skill:

00:02:42.178 --> 00:02:45.746
a file it reads,
so the same steps run every time.

00:02:46.090 --> 00:02:49.923
Then you read it beside your own
checklist and ask for eight changes,

00:02:50.134 --> 00:02:52.119
and the agent rewrites the skill.

00:02:52.510 --> 00:02:55.179
The skill builds a table of
rules from the documents,

00:02:55.370 --> 00:02:58.019
each with its section and its exact words.

00:02:58.410 --> 00:03:01.615
A script measures every P D F: pages,

00:03:01.685 --> 00:03:05.771
type sizes, lines per inch,
margins and web addresses.

00:03:06.215 --> 00:03:11.088
The agent compares the budget with its
justification, and the people with their time.

00:03:11.359 --> 00:03:13.461
Each finding gets one of two labels:

00:03:13.831 --> 00:03:18.694
the system would also stop this,
or only a person catches this today.

00:03:22.000 --> 00:03:24.164
The run, replayed from its log:

00:03:24.615 --> 00:03:29.904
forty-three minutes of work in a copy of
the folder outside any repository, sped up.

00:03:30.304 --> 00:03:35.034
It fetches N S F's guide, its two supplements
and the CAREER solicitation again,

00:03:35.365 --> 00:03:39.132
checks they match the saved copies,
and writes down each version.

00:03:39.476 --> 00:03:41.859
It builds two hundred and seventeen rules,

00:03:42.010 --> 00:03:46.976
and the quote check finds all two hundred
and seventeen quotations, word for word.

00:03:47.354 --> 00:03:49.797
The script measures all eleven files.

00:03:50.328 --> 00:03:54.855
Then the agent reads the description,
the budget and the letters side by side.

00:03:55.165 --> 00:03:58.835
Then it writes the report:
each finding with its file and page,

00:03:58.966 --> 00:04:02.134
the rule's own words,
and one of the two labels.

00:04:02.585 --> 00:04:05.768
Midway through, it notes
that the budget's totals add up,

00:04:06.018 --> 00:04:10.493
but the justification asks for two graduate
assistants where the form budgets one,

00:04:10.763 --> 00:04:12.705
and for more travel than the form.

00:04:17.200 --> 00:04:21.065
Now the check that doesn't rest on
the agent's word: the answer key,

00:04:21.275 --> 00:04:24.539
written before the run
and kept out of the agent's folder.

00:04:24.943 --> 00:04:28.215
It found eight of the eight. It missed none.

00:04:28.546 --> 00:04:32.656
Three of the eight are on N S F's
published list of checks its system makes:

00:04:32.948 --> 00:04:36.175
the sixteen pages, the heading,
and the web address.

00:04:36.606 --> 00:04:40.628
The other five aren't on that list, so you
can't count on the system to catch them:

00:04:40.978 --> 00:04:44.050
ten-point type,
the missing headings in Prior Support,

00:04:44.380 --> 00:04:46.932
two budget lines the
justification doesn't match,

00:04:47.242 --> 00:04:52.125
and a letter that praises the proposal where
the rules ask for one sentence of intent.

00:04:52.495 --> 00:04:55.530
And its labels agree
with the key on all eight.

00:04:56.000 --> 00:05:01.493
A second agent, in a fresh session, re-checked
every quoted rule and the report's measurements.

00:05:01.965 --> 00:05:03.107
Nothing failed.

00:05:03.494 --> 00:05:07.280
And our own script found every
quoted rule in N S F's documents.

00:05:07.671 --> 00:05:13.100
The report flagged none of the eleven decoys:
things that look wrong, but the rules allow.

00:05:17.200 --> 00:05:19.242
The same skill works for other funders.

00:05:19.673 --> 00:05:22.065
For the Agriculture
and Food Research Initiative,

00:05:22.215 --> 00:05:25.859
the narrative can run eighteen pages,
in twelve-point type.

00:05:26.189 --> 00:05:30.572
Checked against those rules, the same
project description fits the eighteen pages,

00:05:30.842 --> 00:05:36.566
but its ten-point type breaks the twelve-point
rule, and it lacks the header N I F A asks for.

00:05:41.200 --> 00:05:43.323
Use it on your own proposal only.

00:05:43.794 --> 00:05:46.367
N I H bars A I from review critiques,

00:05:46.577 --> 00:05:51.463
and N S F bars reviewers from putting
proposals into unapproved A I tools.

00:05:51.953 --> 00:05:57.508
Unpublished ideas can be data your campus
protects, so use a tool your campus approves.

00:05:57.900 --> 00:06:00.927
And the grants office
still has the last word.

00:06:05.200 --> 00:06:07.224
Try it on the practice proposal first.

00:06:07.475 --> 00:06:12.785
The folder and the skill are on the page,
for Codex, Antigravity and Claude Code.

00:06:19.600 --> 00:06:22.529
Ask for the process,
check every finding's quote,

00:06:22.620 --> 00:06:26.191
and run the skill on your own
proposal a week before the deadline.

00:06:29.200 --> 00:06:31.628
More films for this shelf are on the way.