A custom AI reviewer that reads like a grant study section

PerfectPaper’s study section agent reads your manuscript the way a grant reviewer would: asking whether the problem matters, whether the approach is genuinely new or merely unreported, and whether the work moves a field rather than adding to it. Peer review asks a different question — whether what you did was done correctly. Copy the brief below into a custom agent slot.

Use it when the manuscript is part of a larger case — a paper written to support a renewal, a first paper from a new direction, or a preprint that will be cited in a Specific Aims page. Peer review asks whether the work is sound. A study section asks whether it was worth doing.

What a grant study section actually scores

A grant study section scores importance and rigour separately, on a 9-point scale, and since NIH’s simplified review framework took effect for application due dates on or after 25 January 2025 it does so under three factors rather than the five criteria most authors learned. Factor 1, Importance of the Research, carries Significance and Innovation and receives a criterion score of 1 to 9. Factor 2, Rigor and Feasibility, carries Approach and receives its own 1 to 9 score. Factor 3, Expertise and Resources, carries Investigators and Environment and is no longer scored numerically at all; reviewers record whether it is appropriate or whether gaps were identified. NIH gives one of the stated goals of the change as mitigating the effect of reputational bias by “refocusing the evaluation of investigator/environment to within the context of the proposed research.”

NIH compresses the whole exercise into two questions for reviewers: “Should the proposed research project be conducted?” and “Can the proposed research project be conducted?” Journal peer review lives entirely inside the second question, asked in the past tense — was this conducted properly. The first question is the one a manuscript is almost never read for, and it is the one that carries a scored factor of its own.

The scoring mechanics matter for how the agent reports. Criterion scores run 1 to 9, with 1 exceptional and 9 poor. NIH’s scoring system and procedure states that after the meeting “individual reviewer scores will be averaged and the result multiplied by 10 to determine the final impact score”, and that “the range of the final application scores is 10 through 90”. The same document defines a Not Discussed outcome: “applications unanimously judged by the peer review committee to be less competitive are not discussed at the peer review meeting”. Those applications still receive individual criterion scores; they receive no numerical impact score, and NIH notes that not every meeting uses the option.

The consequence for writing is direct. Where a panel does use it, an application can reach its outcome without anyone arguing its case aloud: two or three assigned reviewers read it alone, and the significance argument either works on the page or it does not work.

The same scoring document carries an instruction easy to miss: it states that “the impact score should reflect the reviewer’s overall evaluation, not a numerical average of individual criterion scores.” A manuscript that reads as rigorous on Factor 2 cannot assume that strength averages out a weak Factor 1 argument, and the reverse holds too — a reviewer may let one criterion carry the whole score when that criterion is the one that matters for the work proposed. The same document names a second, harsher outcome distinct from Not Discussed: a majority vote — not unanimity — can rule an application Not Recommended for Further Consideration when it “lacks significant and substantial merit” or raises serious human-subjects or biohazard concerns, and that finding blocks the application from a second level of review entirely, which a routine Not Discussed does not.

Common questions about NIH study section scoring

What does “not discussed” mean on an NIH grant application?

It means the panel judged the application, unanimously and before the meeting, to be “less competitive” — the wording NIH’s scoring guidance uses for the outcome it designates Not Discussed (ND). Those applications are never debated aloud, and NIH notes that not every meeting uses the option. An ND application still receives individual criterion scores for Significance, Innovation and Approach, and still receives written critiques. What it does not receive is a numerical impact score, and since the percentile is calculated from that score, no percentile either. ND is a triage outcome, not the harsher Not Recommended for Further Consideration, which takes a majority vote and stops an application from reaching the second level of review entirely.

Is a low NIH grant score good or bad?

Low is good, which is the reverse of a school grading scale and the reason the question gets asked at all. The criterion scale runs 1 to 9, with 1 labeled Exceptional and 9 labeled Poor; 5 is labeled Good, which NIH’s guidance describes as “a good, medium-impact application,” not a mediocre one. NIH also expects 1 and 9 to be used less often than the other scores, so most applications sit in the middle band. The overall impact score is a separate number running the same direction: reviewer scores are averaged and multiplied by ten, giving a range of 10 through 90, where 10 is the strongest.

What exactly does NIH score on a grant application now?

Since 25 January 2025, NIH scores Factor 1 (Importance: Significance and Innovation) and Factor 2 (Rigor and Feasibility: Approach) each on the 1-to-9 scale. Factor 3 (Expertise and Resources) gets no number at all — reviewers record only whether it is appropriate or whether gaps were identified, a design NIH says exists to keep reputation from substituting for an argument about the specific proposed work.

The seven questions in the brief, and why each is separate

The brief asks seven questions rather than one because each has a different repair, and a finding that merges them into “strengthen your significance” tells an author nothing they can act on. A significance failure is repaired by supplying a consequence; an innovation failure is repaired by narrowing the claim; a framing failure is repaired by moving one paragraph. The agent reports them under separate headings for that reason.

Significance

Significance is assessed by trying to write one sentence — the problem this work addresses and why it matters — using only material in the manuscript. If the agent cannot construct that sentence, that failure is itself the finding, and it names what is missing rather than asking the author to be more compelling.

The distinction that does the work is between asserting importance and demonstrating it. “Cardiovascular disease remains the leading cause of death worldwide” asserts importance of a field and says nothing about the importance of this paper. A demonstration names the consequence of the specific gap: what a clinician, an experimentalist or a modeller currently cannot do, and what becomes possible if this result holds. The agent also reports where the argument appears. A significance case that first surfaces in the discussion’s final paragraph has been read by a reviewer who reached that paragraph already sceptical, and the reviewer comment it produces is the work reads as incremental — an objection about position and consequence, not about data.

Innovation

Innovation is assessed by determining what is genuinely new — a concept, an approach, a technique, or an application of an existing method to a new problem — and then separating novel from merely unreported. The two are routinely conflated in submitted introductions, and only one of them scores.

Combining standard methods in a sequence nobody has published is unreported. It becomes innovation when the combination does something neither component does alone, and the manuscript says what that is. The agent’s output for this question is a proposed honest innovation claim, which is usually narrower than the draft’s and considerably more defensible: “the first application of single-cell profiling to this tissue during regeneration” survives a reviewer holding a citation in a way that “a novel approach” does not. The asymmetry is worth stating plainly, because it is the same one that governs priority claims in a journal submission: an unrefuted claim gains you little, and a claim a reviewer refutes with one reference damages every other claim in the paper. Where the pushback is about positioning rather than priority, see novelty objections and the reviewer who says this is already known.

The asymmetry does not run only one way, and NIH’s own scoring guidance states the exception directly: “a project that by its nature is not innovative may be essential to advance a field.” A validation study, a replication in a new population, or a resource paper can score well on Factor 1 with a modest or absent innovation claim, provided the significance case — the consequence of the gap, not the gap itself — is stated plainly. A narrow honest-innovation finding is a correction to an overstated claim, not a verdict against a paper that never needed one.

Approach

Approach is assessed as reasoning, not execution: whether the strategy fits the question asked, whether alternatives were considered, and whether the manuscript names what would falsify its interpretation. The agent then identifies the single weakest link in the chain and says which one it is, because a list of six concerns and a named weakest link produce different author behaviour.

The falsification question is the one most manuscripts fail silently. A discussion that considers only interpretations consistent with the authors’ model reads to a grant reviewer as a paper that has not stress-tested itself, whatever the quality of the experiments. The repair is a paragraph stating the alternative explanation and the observation that would distinguish it — which is also the strongest available form of a limitations section, as set out in how to write a limitations section. Where the weakest link is the inferential step from association to mechanism, causal language discipline is the more specific agent and should be run alongside this one.

One distinction changes how the weakest link should be reported: a limitation the manuscript names and reasons through scores differently from the identical limitation left silent. The approach criterion rewards a strategy that has already been stress-tested by its own authors, so the agent should distinguish “the strategy has an unaddressed weak point” from “the strategy has a weak point the authors have already reasoned about and defended” — a summary statement responds to those two cases in different language even when the underlying method is identical.

Rigour of prior work

Rigour of prior work is assessed by asking whether the premises the manuscript rests on are established or merely assumed, and by identifying any load-bearing assumption supported by a single prior study or by work the manuscript does not cite. This is not a general literature check; it is a check on the specific claims the argument cannot survive without.

NIH made this an explicit application requirement in NOT-OD-15-103, which instructs applicants to “describe the general strengths and weaknesses of the prior research being cited by the investigator as crucial to support the application” — the scientific premise, one of the notice’s four areas alongside rigorous experimental design, consideration of relevant biological variables, and authentication of key biological and chemical resources. The clause a manuscript author should notice is “cited by the investigator as crucial”: the obligation attaches to the load-bearing citations, not to the whole reference list. A paper whose entire model rests on one 2011 finding that has not been independently reproduced has a premise problem, and the reviewer version of it is the experiment has not been replicated. Where the premise is an animal result carried into a new context, pair this with in vivo rigour.

Citation count is not the same test as independence, and the agent should keep the two separate. Ten downstream papers that each cite the same original finding without attempting to reproduce it are ten citations of one data point, not ten confirmations of it. The distinction NOT-OD-15-103 is actually drawing is between a premise that has been tested again and one that has only been repeated in prose, and the finding worth reporting is which of the two a manuscript’s load-bearing citation represents.

The obvious objection

The obvious objection is the first substantive criticism a sceptical reviewer will raise, and the agent is instructed to state it specifically enough that the author could write a paragraph in response. Vagueness here is the failure mode: “reviewers may question the model system” is not usable, while “reviewers will ask why the effect was not tested in primary cells given that the transformed line carries the mutation under study” is.

The second half of this question is whether the manuscript anticipates the objection, and the answer is usually partial — the objection appears in a limitations list without an answer. A named objection with a named next experiment reads as command of the problem; the same objection listed and unanswered reads as an admission. The commonest objection this agent returns across mechanistic work is the mechanism is not shown, and the commonest across observational work is that the comparison group is not the one the claim requires.

Framing

Framing is assessed by asking whether the introduction positions the work inside a question a field cares about or inside a gap in the literature. A gap is not a reason. “X has not been characterised in Y” is true of an unbounded number of things, and an introduction that argues only for absence gives the reviewer no basis for caring about the presence.

The repair is mechanical and usually short. Take the gap sentence and add the consequence clause: not “the role of X in Y has not been examined”, but “without knowing whether X operates in Y, the dosing strategy used in Z rests on an assumption no one has tested.” The same edit improves the abstract, since the abstract’s first two sentences carry the framing to anyone who reads no further — see how to write an abstract. A framing failure is also the most common route to a rejection that cites fit, because an editor cannot place a paper whose question is unstated; out of scope is frequently a framing verdict wearing an editorial label.

One exception is worth naming before an author over-corrects into inventing a consequence that is not really there. A methods or resource paper can legitimately frame its significance as an absence — “no existing tool measures X” is a reason, not a gap, once it is paired with a quantified need: how many groups currently work around the absence, and how much time each workaround takes. The failure mode the agent flags is a gap stated with no downstream quantification at all, not the presence of the word “gap” itself.

Trajectory

Trajectory is assessed by determining what this work sets up, and by saying which of two things the paper reads as: a step in a programme, or an isolated result. The agent is required to pick one and say so, because the hedged version of this finding is worthless.

Trajectory is the question with no journal peer-review equivalent at all. A referee evaluates one manuscript against one journal’s standard and has no reason to ask what comes next. A study section is deciding whether to support a line of work for several years, and reads a publication record as a direction. The practical output is a sentence the author can put into a Specific Aims page or a cover letter: what this paper establishes, and what it makes it now possible to ask. Where a paper genuinely is an isolated result, saying so is useful — it tells the author to build the programme argument outside the manuscript rather than to insert an unsupported claim into the discussion.

Questions about significance and innovation

How do I tell whether my significance argument is strong enough?

Write one sentence containing the problem and the consequence of not solving it, using only words that appear in your manuscript. If the sentence needs knowledge you hold and the paper does not state, the argument is not on the page. Then note the first page on which a reader encounters it; an argument that first appears in the discussion has arrived after the reviewer formed a view.

Reviewer says my approach isn’t novel enough — is combining two existing methods innovation or not?

Only if the combination does something neither method does alone, and your manuscript states what that is. Applying an established technique to a new tissue or population without a novel mechanistic claim is unreported, not innovation — the honest version, “first application of X to Y,” is narrower but survives a reviewer holding a citation, which the broader claim usually does not.

What is the difference between significance and innovation in an NIH application?

They are scored together under Factor 1 and have to be argued separately, because they fail separately. Significance is about consequence: what a field currently cannot do, and what becomes possible if this work holds. Innovation is about newness: whether the concept, the approach or the technique departs from what is already available. A paper can be highly significant with almost no innovation — NIH’s scoring guidance says outright that “a project that by its nature is not innovative may be essential to advance a field” — and a genuinely new method aimed at a question nobody needed answered is the mirror failure. What decides the innovation half is novel versus merely unreported: a sequence of standard methods nobody has published is unreported, and becomes innovation only when the combination yields something neither component gives alone and the manuscript says what that is.

When to use this agent

• The paper is preliminary data for a grant application
• You are establishing a new research direction and need to know how it reads
• The work is technically sound and you suspect the significance is undersold
• You are preparing a renewal and this paper carries the productivity argument
• You want the objection a reviewer will raise before a reviewer raises it
• A previous submission was rejected on importance or interest rather than on method
• Co-authors disagree about what the paper’s central claim should be

The agent brief

Paste this into a custom agent. Suggested settings: work type general_review, skill level expert, tools scholar search.

Name: Study section lens

You are reading this manuscript as a grant reviewer would, not as a journal
peer reviewer. Do not check whether the analysis is correct; other reviewers
cover that. Assess whether the work matters and how it reads to someone
deciding what to fund.

Address each of the following, and be direct rather than encouraging:

1. Significance. State, in one sentence, the problem this work addresses and
   why it matters. If you cannot construct that sentence from the manuscript,
   say so plainly and identify what is missing, because that failure is the
   finding. Determine whether the manuscript asserts importance or
   demonstrates it. Identify where the significance argument actually appears,
   and report if it appears only in the final paragraph.

2. Innovation. Determine what is genuinely new: a concept, an approach, a
   technique, or an application of an existing method to a new problem.
   Distinguish novel from merely unreported. Report claims of novelty that
   rest on a combination of standard methods, and say what the honest
   innovation claim would be.

3. Approach. Assess whether the strategy is well reasoned for the question
   asked, whether alternatives were considered, and whether the manuscript
   acknowledges what would falsify its interpretation. Identify the weakest
   link in the chain of reasoning and name it.

4. Rigour of prior work. Determine whether the premises the work rests on are
   established or assumed. Identify any load-bearing assumption supported by a
   single prior study or by work the manuscript does not cite.

5. The obvious objection. State the first substantive criticism a sceptical
   reviewer will raise, and determine whether the manuscript anticipates it.
   Be specific enough that the author could write a paragraph in response.

6. Framing. Assess whether the introduction positions the work within a
   question a field cares about, or within a gap in the literature. A gap is
   not a reason; report introductions that argue only that something has not
   been done before.

7. Trajectory. Determine what this work sets up. Assess whether it reads as a
   step in a programme or as an isolated result, and say which.

Where scholarly search is enabled, check whether the positioning claims are
consistent with the recent literature, and report claims of unexplored
territory in areas with recent activity. Cite what you rely on.

Do not soften findings. An encouraging review of a weak significance argument
is worse than no review, because it leaves the author confident.

How to adapt the brief to your funder

The brief is written against the NIH factor structure, and three edits fit it to other funders without disturbing the seven-question shape. Each edit changes what a question asks, never whether it is asked separately.

For an NSF proposal, replace question 1 with the two NSF merit review criteria: intellectual merit and broader impacts. Broader impacts has no NIH analogue and is routinely the criterion a manuscript supplies least material for, so ask the agent to determine whether the manuscript contains any material a broader-impacts statement could be built from — data or software released for reuse, a resource other groups can use, a training or dissemination component already in place.

For an ERC or a national-agency fellowship, weight questions 2 and 7 above the others. Those schemes fund a person and a direction rather than a project, so the trajectory question stops being a supplement and becomes the primary output. Ask explicitly for the two-sentence programme statement the paper supports.

For a renewal or a competing continuation, add a productivity determination to question 7: does this manuscript, read alone, demonstrate that the previous award produced something, and which specific aim of the prior application does it close? A renewal reviewer is doing exactly that mapping, usually with the prior Specific Aims page open beside the reprint.

Two limits are worth keeping as written. The brief forbids checking whether the analysis is correct, and an agent that starts auditing statistics has been given too broad a scope — the negative instructions are the load-bearing ones, as how to write a custom reviewer agent sets out. And the brief assumes an original research article; a review or a meta-analysis has a different claim structure and needs a different brief.

Questions about adapting this for other funders

Does this work for a manuscript going to a non-NIH funder?

Yes, with one substitution. The seven questions are funder-neutral; the NIH factor names are not. For an NSF proposal, replace the significance question with intellectual merit and broader impacts. For a fellowship or an ERC scheme, weight innovation and trajectory above the others, since those schemes fund a direction rather than a project.

I’m writing an ERC fellowship application, not an R01 — should I still answer all seven questions the same way?

Ask all seven, but not with equal weight. ERC and comparable fellowship schemes fund a person and a direction rather than a single project, so Innovation and Trajectory carry more of the argument than Significance alone does under NIH’s factors. Write the two-sentence programme statement the Trajectory question produces; it belongs in the proposal itself, not only the cover letter.

What this agent surfaces

PerfectPaper’s study section agent surfaces eight questions a standard journal review never asks, listed here beside the reason each one falls outside a referee’s remit.

Worked example: an introduction that argues a gap

Consider a constructed but ordinary case: an introduction whose third paragraph ends “however, the role of transporter T in tissue Y has not been examined”, and whose results are clean — a knockout, a rescue, a dose response, three biological replicates, everything a referee would ask for.

A journal referee reads that paper and finds nothing to object to, because nothing in it is wrong. The study section reading returns four findings from the same eight pages. Significance: the sentence “the problem this work addresses and why it matters” cannot be constructed, because the manuscript states an absence and never states a consequence of the absence. Innovation: the claim rests on applying an established knockout strategy to a tissue where it has not been applied, which is unreported rather than novel, and the honest claim is the narrower one. Framing: the introduction argues a gap, and the gap is one of many. Trajectory: the paper closes a question and opens none, so it reads as isolated.

The repair touches two sentences and no experiments. The gap sentence gains its consequence — “transporter T sets the intracellular concentration of drug D in liver, and dosing in tissue Y currently assumes the same relationship holds there, an assumption no one has tested” — and the discussion gains a trajectory sentence naming what the result makes it possible to ask next. The data are unchanged. What changed is that a reviewer now has a reason to care, stated before the methods rather than after the figures. This is the mirror image of the overclaim check, which finds claims stronger than the evidence; here the evidence was stronger than the claim.

How to run the study section questions yourself

The study section questions are answered by reading only the introduction, the abstract and the discussion, with the figures closed. Closing the figures is the method, not a shortcut: the data are what a referee checks, and looking at them makes it impossible to see whether the argument stands without them.

Write the significance sentence on a blank line, using only the manuscript. One sentence, problem plus consequence. If you find yourself importing knowledge you have and the paper does not state, the paper does not state it. Note where in the text you found each half.

Mark the first page on which a reader learns why this matters. If that page is in the discussion, move the material. Reviewers form a disposition early and spend the rest of the reading confirming it.

Underline every claim of novelty and write beside each one what kind it is — concept, approach, technique, or application. A claim that resists classification is usually the unreported kind, and rewriting it as an application claim strengthens rather than weakens it.

List the premises the argument would collapse without, and put a citation count beside each. Any premise supported by one paper, or by no paper, is the finding. This is the NOT-OD-15-103 exercise applied to a manuscript rather than an application.

Write the objection you would raise if the paper were someone else’s. Then find the paragraph that answers it. If there is no such paragraph, write it — this single step converts more study-section findings into text than any other.

Say out loud which the paper is: a step, or a result. Both are legitimate. Only one of them supports a renewal argument, and knowing which one you have determines whether the programme case belongs in the paper or in the Specific Aims page.

What reviewers say when significance is the problem

Study section reviewers and journal referees write different sentences about the same manuscript, and the study-section versions are the ones this agent predicts. The sentences below are composites in the register summary statements use, not quotations from any particular critique. They rarely contain the word significance.

“The application is technically sound but the significance of the proposed work is not clearly articulated.” “The premise rests on a single prior report that has not been independently confirmed.” “The innovation is incremental; the methods proposed are established and the combination, while not previously reported, does not by itself constitute a conceptual advance.” “The introduction establishes that this has not been done rather than why it should be done.” “The investigators do not consider alternative interpretations of their preliminary data.” “It is unclear what the next stage of this work would be.” “The proposed work would confirm an expectation rather than test it.”

The last three are the expensive ones, and none of them is answerable with more data. A referee’s objection normally names an experiment; a study section’s objection frequently names an argument, and the author who receives it after submission has to rebuild framing under a deadline. That asymmetry is the reason to run these questions before a manuscript is submitted rather than after a summary statement arrives. The general pattern of importance-based rejection is covered in why papers get rejected.

When the obvious remedy is unavailable

The obvious remedy — add the consequence clause, narrow the novelty claim, write the alternative-interpretation paragraph — is unavailable more often than review guidance admits, and each case has a specific substitute.

The journal’s introduction is capped and there is no room. Spend the words at the front. One consequence clause attached to the gap sentence outperforms three scattered explanations later, and a background sentence describing something the audience already knows is nearly always the sentence to delete for it.

The result really is confirmatory. Say so and state what the confirmation licenses. “This result was expected under model M, and its confirmation removes the main alternative explanation for the finding in [prior work]” is a legitimate significance argument. What fails is a confirmatory result described as a discovery, which invites the reviewer to make the point for you.

Your load-bearing premise has one supporting paper and you cannot replicate it. State the dependency explicitly and give the observation that would break it. A manuscript that names its single point of failure is read as rigorous; the same manuscript with the dependency unstated is read as unaware, and a reviewer who finds it will treat everything downstream as provisional.

The trajectory argument does not fit inside the paper. Put it in the cover letter and in the Specific Aims page instead. A discussion paragraph asserting programme-level ambition that the manuscript does not support is a framing overclaim, and the same reviewer who wanted a direction will penalise the unsupported version of it.

A co-author objects that the narrower innovation claim undersells the work. Compare the two claims against what a reviewer can refute. The narrow claim is checkable and survives; the broad claim is a bet placed against everyone who has read the field’s literature more recently than you have.

What this agent does not do

The study section agent evaluates the argument and takes the data as given, which sets four hard boundaries worth stating before you run it.

It does not check whether the analysis is correct, whether the statistics fit the design, or whether the figures support the results. Those are peer-review questions, and the standing team covers them; running this agent in place of that review leaves the soundness question unasked. It does not predict a score, a percentile or a funding outcome, none of which is a property of a manuscript. It does not verify priority claims from its own knowledge — with scholar search enabled it checks positioning against the recent literature and cites what it relies on, and without it the innovation finding is reported as unverifiable within the manuscript. And it reads original research articles; a narrative review, a methods protocol or a meta-analysis has no significance-versus-execution split of the kind this brief is built on.

The failure mode to watch for is agreeable output. An agent that returns “the significance argument is strong” for a paper whose introduction argues a gap has been over-instructed toward encouragement, and the brief’s closing line exists to prevent exactly that.

How this differs from the built-in review

The standing team is built to evaluate whether a manuscript is correct, complete and well supported — the peer review question. Every specialist in it takes the research question as given and examines execution. This agent takes execution as given and examines the question, which is the axis a study section scores on and the one a technically flawless paper can still fail.

It is the one agent in this collection that will sometimes tell you the work is fine and the framing is the problem. Pairs with the overclaim check for opposite reasons: that agent finds claims stronger than the evidence, this one finds evidence stronger than the claims. PerfectPaper’s study section agent reports where a manuscript’s significance argument first appears, whether the introduction gives a reason or a gap, and what the honest innovation claim would be — three things a grant reviewer scores and a journal referee has no reason to ask about. It is a supplement to AI peer review, never a replacement for it.

Review my manuscript

Related agents

The agents below share a review with this one; you can run three at a time. See Custom AI reviewer agents for the full library, organised by research area.

• How to write your own — use it when you are writing your own brief rather than pasting one.
• Ethics, consent and data sharing — use it for approval statements and consent scope for deposition.
• Overclaim check — use it for the distance between what was shown and what the abstract says.

Frequently asked questions

How is grant review different from peer review?

Peer review asks whether the work was done correctly and whether the conclusions follow. Grant review asks whether the problem is important, whether the approach is innovative, and whether the work is worth supporting. A paper can be flawless by the first standard and unpersuasive by the second.

Can this agent predict my grant score?

No. The agent surfaces the questions a study section asks and reports how your manuscript answers each one — it does not output a number. Scoring depends on the panel’s composition, the funding line’s competition that cycle, and reviewer disagreement resolved in a room the manuscript never enters, none of which is knowable from the text alone.

Does a resubmission get judged on how well I addressed the reviewers’ last critique, or as a whole new application?

As a whole application. NIH’s additional review criteria instruct reviewers, for a resubmission, to “evaluate the full application as now presented,” not to grade the introduction-of-changes page against the prior summary statement. Answering every point from the last review is necessary but not sufficient — the significance and innovation arguments still have to stand on their own in the resubmitted text.

Should I use this before or after peer review feedback?

Before submission, and again before writing a Specific Aims page that cites the work. Peer-review feedback addresses execution, not the case for funding it, so running this agent afterward does not duplicate that ground — it is most useful precisely when the science is already settled and the argument for why it matters is not.