This is the run behind our NSF sample report. From upload to verdict in one continuous run, by a team of more than a dozen purpose-built AI agents working independently, then together: compliance, solicitation alignment, reference checks, a deep literature search, prior funded work, a technical analysis, three independent reviewers, three rounds of panel argument, a chaired consensus that preserves dissent, and a recommended revision for every weakness the panel raises.
Sample review
CAREER: Learning Sediment Transport from Low-Cost Coastal Sensor NetworksNSF 22-586
Sample panel review
1/1700:00
Progress Status
Step 1 is yours. Everything after it is GrantPanel.
PI · 01
PI (user) uploads
Required
Proposal: the project description, plus any other parts, in one PDFDrop a PDF here, or browse
Solicitation: the agency's call and its review criteriaDrop a PDF here, or browse
Optional
+
Supporting materialsRelated articles, prior reviews, other related materials
Name this run It labels the run in your history, on the report, and in the email about it.
e.g. NSF DMREF: autonomous catalyst discovery
The proposal and solicitation are all that is required. The panel works from your documents alone.
CAREER proposal
figures
PDFcareer-coastal-sensors.pdf 2.4 MB
PDFnsf22586.pdf 0.7 MB
DOCnotes-for-panel.docx 48 KB
ONE PROPOSAL, START TO FINISH
From upload to a stronger proposal.
Your proposal
Run 1, then deletedRuns 2-3, then deleted
Your report
Yours until you delete it
The solicitation
Kept for the 60-day window
Never used to train AI, by us or by our providers.
1 in 5proposals is funded at NSF
6–12 mobefore a real panel's critique arrives
5–15 minto hear ours
OpenAI (ChatGPT) and Anthropic (Claude) are referenced only to identify compatible AI providers and products. GrantPanel is independent and is not affiliated with, sponsored by, or endorsed by either company. No third-party logos are used.
THE REVIEW STANDARD
What a complete pre-submission review requires.
A purpose-built review workflow, not a blank chat window, and not a months-long black box.
01
Your funder's own criteria and scale
Hover for detailsTap for details
Your funder's own criteria and scale
Review criteria are read from your solicitation and the agency's policy rules rather than a generic template, and rated on the funder's own scale, with each criterion tied to its source location, so the rulebook stays visible in the review record. Solicitation-specific risks are flagged before the first reviewer reads.
Tap to close
02
Three independent reviewers, three rounds
Hover for detailsTap for details
Three independent reviewers, three rounds
Each reviewer works from separate context and forms its assessment before seeing any panel position: no majority cue, no premature consensus. Then the three argue across three panel rounds. The consensus summary records where they converged, where they differed, and why.
Tap to close
03
Every finding tied to a page and passage
Hover for detailsTap for details
Every finding tied to a page and passage
Every finding points back to the page and passage it comes from, so you can verify the panel's claim against your own document.
Tap to close
04
Your math, figures, and claims checked
Hover for detailsTap for details
Your math, figures, and claims checked
A technical analyst, a senior researcher in your field, checks your math, figures, tables, and key claims before the panel meets, and specialist reviewers evaluate methods, aims, and feasibility, not just formatting, before the panel weighs significance.
Tap to close
05
Checked against current literature and funded awards
Hover for detailsTap for details
Checked against current literature and funded awards
Novelty claims are checked against the field as it stands: a live deep search across prior work, competing approaches, and recent publications, not just what a model remembers, and the related awards NSF, NIH, and other agencies have already funded.
Tap to close
06
Finds weaknesses, but can't predict your score
Hover for detailsTap for details
Finds weaknesses, but can't predict your score
Practice, not prophecy: a simulation can expose weaknesses while you can still fix them. It cannot predict a real panel's score, and we never claim it does. Every finding comes with a transcript, so you can inspect the reasoning rather than take it on faith.
Tap to close
WHY IT IS BUILT THIS WAY
Disagreement is the feature.
Ask an AI to be critical and it will politely agree with itself. So we don't just ask: independence is built into the pipeline, and the panel's instructions hold each reviewer to its position and the chair to the record.
Independence is structural
Hover for detailsTap for details
Independence is structural
Each reviewer runs in its own context and files a complete assessment before any other position exists to be influenced by. A prompt asking a model to be impartial is a request. Separate context is a guarantee.
Tap to close
Positions must be defended
Hover for detailsTap for details
Positions must be defended
In round two each reviewer answers the others, pushing back or conceding with the evidence it means, and a turn that only repeats its review is sent back. Round three locks a final rating with the reason for it. Agreement has to be earned in the transcript.
Tap to close
The chair keeps dissent
Hover for detailsTap for details
The chair keeps dissent
The chair writes from the reviews and the discussion: consensus points are the ones reviewers agreed on, and where the panel did not converge it says so under Points of disagreement instead of averaging the readings into one.
Tap to close
WHAT THE DESIGN RULES OUT
×
The agreeable readerHover for detailsTap for details
The agreeable reader
A single model, however well prompted, tends to validate the framing it was handed. Three separate readings do not share that framing.
Tap to close
×
Manufactured consensusHover for detailsTap for details
Manufactured consensus
The panel gives one verdict, but a disagreement that survives round three is reported as a disagreement, not averaged away.
Tap to close
×
Criticism you cannot traceHover for detailsTap for details
Criticism you cannot trace
Each reviewer's findings sit in its own review and its turns in the discussion, and the compliance and technical checks in their own sections, so you can read the argument behind a conclusion rather than take it on faith.
Tap to close
Every one of these is checkable after the fact: the deliberation transcript ships with the report. Read a sample one →
RECOMMENDED REVISIONS & REVISION REVIEW
Every weakness comes with a recommended fix. Every revision gets a second look.
A revision should answer a critique, not merely sound different. Every weakness and gap the panel raises leaves with a specific recommended revision and how to fix it: the steps, the sentence as it stands, and suggested wording, or for research content, what to write. Revisions are grouped by criterion, most urgent first, and tied to the section they belong to. Your proposal comes with three runs, so when you revise, a full panel reads the new version and returns its own verdict, then checks every concern the earlier runs raised against it: addressed, partly addressed or not addressed, with the passage that shows it. Rewording a passage does not count as addressing the concern.
01 Every weakness paired with a specific revision02 Grouped by criterion, tied to the section it applies to03Three runs per proposal, so start with your draft: each revision gets its own verdict and a check of every earlier concern
RECOMMENDED REVISIONSFrom your panel report
01
Intellectual Merit
Fix before submission
Reconcile probe span with reported erosion
How to fixExplain what produced the erosion record: what the probe itself measured, and what came from another measurement.
Now“…recorded 0.42 m of erosion at the innermost surf-zone node…”
What to write[how the 16-electrode, 5 mm-spaced probe recorded this range]
Support the alongshore exclusion rule
02
Broader Impacts
Address next
Assess education outcomes
State and justify release licenses
A recommendation, not an edit. Your document is never changed by GrantPanel.
SINCE PRIOR RUNSRun 3 · revision
3 of 5 addressed
Raised in run 1
Reconcile probe span with reported erosionAddressed
Support the alongshore exclusion rulePartly addressed
Assess education outcomesNot addressed
State and justify release licensesAddressed
Raised in run 2
Justify the storm events held out for validationAddressed
WHAT THE PANEL CATCHES
The kind of thing you'd rather hear early.
Illustrative findings drawn from real review patterns: the sort of note a panel writes when a strong proposal has a fixable seam.
Compliance review · stage 2
A missing sub-heading that triggers a return
Hover for detailsTap for details
A missing sub-heading that triggers a return
The Project Summary is missing the Intellectual Merit sub-heading required by PAPPG II.D.2.b. This is a return-without-review trigger, not a stylistic note.
HOW TO FIX
Add the two required sub-headings verbatim. Five minutes of work; the alternative is an automatic return.
Tap to close
Independent reviews · stage 8
Aim 3 has no fallback if Aim 1 falls short
Hover for detailsTap for details
Aim 3 has no fallback if Aim 1 falls short
Aim 3 depends entirely on Aim 1 reaching its performance target, and no fallback is described. If Aim 1 underdelivers, two thirds of the project has nothing to run on.
HOW TO FIX
Name the existing MPC baseline as an explicit contingency path for Aims 2–3, roughly half a page.
Tap to close
Technical analysis · stage 7
The link budget can't support the measurement
Hover for detailsTap for details
The link budget can't support the measurement
The detection threshold needs a 20 dB signal-to-noise ratio, but the link budget in Section 3 gives 11 dB at the proposed range. As designed, the central measurement may not be possible.
HOW TO FIX
Rework the link budget with the measured noise floor, or narrow the range claim to what 11 dB supports.
Tap to close
Literature review · stage 5
Three recent overlapping papers left uncited
Hover for detailsTap for details
Three recent overlapping papers left uncited
Three 2025–26 papers overlap the Aim 2 formulation and none are cited. A panelist who knows that work will read the omission as unawareness or avoidance.
HOW TO FIX
Cite the three, then add one sentence in Aim 2 stating what your approach does that theirs does not.
Tap to close
Put your proposal in front of our panel first.
Upload your proposal and the agency's solicitation, and get a full simulated panel review in about 5–15 minutes.