A Workflow for Grant Discovery and Proposal Development Using Large Language Models: Development and Formative Evaluation

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

These rules align with published recommendations that large language models can assist grant development but require fact-checking, privacy awareness, and researcher ownership of scientific and strategic content.

“These rules align with published recommendations that large language models can assist grant development but require fact-checking, privacy awareness, and researcher ownership of scientific and strategic content [8].”

• cites [8] Seckel, E., Stephens, B. Y. & Rodriguez, F. Ten simple rules to leverage large language models for getting grants. PLOS Comput. Biol. 20 (3), e1011863. 10.1371/journal.pcbi.1011863 (2024). sourceSupported

Classified as probably not a substantial claim.

Published guidance warns against hallucinated references, privacy leaks, overreliance, and funder-policy violations.

“Published guidance warns against hallucinated references, privacy leaks, overreliance, and funder-policy violations [8,19].”

• cites [8] Seckel, E., Stephens, B. Y. & Rodriguez, F. Ten simple rules to leverage large language models for getting grants. PLOS Comput. Biol. 20 (3), e1011863. 10.1371/journal.pcbi.1011863 (2024). sourceSupported
• cites [19] Resnik, D. B. & Hosseini, M. Hallucinated citations produced by generative artificial intelligence may constitute research misconduct when citations function as data in scholarly papers. Acc. Res. 15 10.1080/08989621.2026.2645390 (2026 Mar). sourcePartly supported

  The sections “The problem of GenAI hallucinated citations” and “Research misconduct involving GenAI use” warn about fabricated references, unchecked AI output, and possible NIH action over improper AI use in grants. Our copy does not address privacy leaks, and it does not establish the broader claim about funder-policy violations. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S2 (checkable).

Evidence generation must coexist with product development cycles, limited staff, and institutional pressure to demonstrate clinical or commercial relevance before evidence is available.

“Evidence generation must coexist with product development cycles, limited staff, and institutional pressure to demonstrate clinical or commercial relevance before evidence is available [22].”

• cites [22] Sutton, R. T. et al. An overview of clinical decision support systems: benefits, risks, and strategies for success. NPJ Digit. Med. 3 (1), 17. 10.1038/s41746-020-0221-y (2020). sourceCould not verify

  The Financial challenges section says CDSS cost analyses are sparse and that implementation and staff-training costs can be substantial. It does not address evidence generation alongside product development cycles, limited staffing, or pressure to demonstrate relevance before evidence is available. Searches for those terms and close paraphrases found no passage addressing that point; page numbers were not available in our copy. (LLM, conf=0.72)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

This is consistent with published grant-writing guidance: large language models may support clarity, feedback, and iteration, but no model can replace funder-specific judgment, scientific ownership, or expert review.

“This is consistent with published grant-writing guidance: large language models may support clarity, feedback, and iteration, but no model can replace funder-specific judgment, scientific ownership, or expert review [8].”

• cites [8] Seckel, E., Stephens, B. Y. & Rodriguez, F. Ten simple rules to leverage large language models for getting grants. PLOS Comput. Biol. 20 (3), e1011863. 10.1371/journal.pcbi.1011863 (2024). sourceSupported

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

This is likely relevant to other low-and middle-income country health-AI teams whose funding fit depends on geography, entity type, partner relationships, and diseasespecific evidence.

“This is likely relevant to other low-and middle-income country health-AI teams whose funding fit depends on geography, entity type, partner relationships, and diseasespecific evidence [1,5].”

• cites [1] World Health Organization. Global strategy on digital health 2020–2025. Geneva: World Health Organization. (2021). Available from: https://www.who.int/docs/default-source/documents/gs4dhdaa2a9f352b0445bafbc79ca799dce4d.pdfNot checked yet

  Not checked in the free run.

• cites [5] Irihamye, E. et al. Sustainable by Design: Digital Health Business Models for Equitable Global Health Impact in Low-Income and Low-Middle-Income Countries. Mayo Clin. Proc. Digit. Health. 3 (4), 100261. 10.1016/j.mcpdig.2025.100261 (2025). sourcePartly supported

  The Results section (External Market-Shaping Interventions; page unavailable) discusses funding constraints for LMIC digital-health companies, strategic partnerships, and financing tied to clinical or operational performance metrics. It does not establish that a health-AI team’s funding fit depends on geography, entity type, or disease-specific evidence. The citing authors’ suggestion that these issues may be relevant to other teams is their own inference. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Digital health sustainability literature emphasizes financing barriers, fragmented markets, and the need for strategic partnerships.

“Digital health sustainability literature emphasizes financing barriers, fragmented markets, and the need for strategic partnerships [1,5,20,21].”

• cites [1] World Health Organization. Global strategy on digital health 2020–2025. Geneva: World Health Organization. (2021). Available from: https://www.who.int/docs/default-source/documents/gs4dhdaa2a9f352b0445bafbc79ca799dce4d.pdfNot checked yet

  Not checked in the free run.

• cites [5] Irihamye, E. et al. Sustainable by Design: Digital Health Business Models for Equitable Global Health Impact in Low-Income and Low-Middle-Income Countries. Mayo Clin. Proc. Digit. Health. 3 (4), 100261. 10.1016/j.mcpdig.2025.100261 (2025). sourceSupported
• cites [20] Wahl, B., Nassiri-Ansari, T., Redpath, D. D., Allotey, P. & Schwalbe, N. Artificial intelligence for research capacity strengthening: Two reviews and a pathway to shift power in global health. PLOS Digit. Health. 5 (4), e0001302. 10.1371/journal.pdig.0001302 (2026). sourcePartly supported

  The Results (Review of reviews; page not provided) emphasize sustainable funding models and equitable partnerships, supporting the financing and partnership themes in the context of AI for global health research capacity. The paper does not address fragmented markets, and its focus is narrower than digital health sustainability generally. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [21] Kiosia, A. et al. Current data science capacity building initiatives for health researchers in LMICs: global and regional efforts. Front. Public. Health. 12, 1418382. 10.3389/fpubh.2024.1418382 (2024). sourcePartly supported

  The Discussion (page not provided) says that sustaining health data science capacity-building initiatives in LMICs depends on continuous funding and describes partnerships with government, academia, and the private sector. It does not discuss fragmented markets, and its focus is narrower than digital health sustainability literature generally. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

May combine two claims.

These policies shift accountability to the researcher: AI assists in preparation, but the researcher owns every scientific claim in the final product.

“These policies shift accountability to the researcher: AI assists in preparation, but the researcher owns every scientific claim in the final product.”

• cites [7] Stokel-Walker, C. ChatGPT listed as author on research papers: many scientists disapprove. Nature 18 10.1038/d41586-023-00107-z (2023 Jan). sourceNot checked — no full text of the source
• cites [16] Biswas, S. ChatGPT and the Future of Medical Writing. Radiology 307 (2), e223312. 10.1148/radiol.223312 (2023). sourceNot checked — no full text of the source
• cites [17] Ganjavi, C. et al. Publishers’ and journals’ instructions to authors on use of generative artificial intelligence in academic and scientific publishing: bibliometric analysis. BMJ 384, e077192. 10.1136/bmj-2023-077192 (2024). sourceSupported

May combine two claims.

A workflow without issue-based state tracking, deduplication memory, or citation audit checkpoints risks producing output that lacks a recoverable decision history, cannot be verified against eligibility criteria, and may include fabricated references that would only be detected by the funder.

“A workflow without issue-based state tracking, deduplication memory, or citation audit checkpoints risks producing output that lacks a recoverable decision history, cannot be verified against eligibility criteria, and may include fabricated references that would only be detected by the funder [19].”

• cites [19] Resnik, D. B. & Hosseini, M. Hallucinated citations produced by generative artificial intelligence may constitute research misconduct when citations function as data in scholarly papers. Acc. Res. 15 10.1080/08989621.2026.2645390 (2026 Mar). sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The large language model citation audit checkpoint applies to research outputs produced by the system, but grant proposals submitted to external funders require additional manual verification for every factual claim, since hallucinated citations in a grant application are a credibility risk that automated checks may not fully prevent.

“Finally, the large language model citation audit checkpoint applies to research outputs produced by the system, but grant proposals submitted to external funders require additional manual verification for every factual claim, since hallucinated citations in a grant application are a credibility risk that automated checks may not fully prevent [19].”

• cites [19] Resnik, D. B. & Hosseini, M. Hallucinated citations produced by generative artificial intelligence may constitute research misconduct when citations function as data in scholarly papers. Acc. Res. 15 10.1080/08989621.2026.2645390 (2026 Mar). sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

Recording these structural blockers helps the business-development and evidence teams coordinate and avoids repeated discovery of the same structurally closed opportunities.

“Recording these structural blockers helps the business-development and evidence teams coordinate and avoids repeated discovery of the same structurally closed opportunities [4,21].”

• cites [4] Bill and Melinda Gates Foundation. Jun 10. New Global Partnership Launches Call for Proposals to Evaluate AI Tools in Primary Health Care in Low- and Middle-Income Countries. Published 2026 Feb 20. Accessed (2026). Available from: https://www.gatesfoundation.org/ideas/media-center/press-releases/2026/02/ai-impact-healthNot checked yet

  Not checked in the free run.

• cites [21] Kiosia, A. et al. Current data science capacity building initiatives for health researchers in LMICs: global and regional efforts. Front. Public. Health. 12, 1418382. 10.3389/fpubh.2024.1418382 (2024). sourceCould not verify

  The Discussion (page not provided) recommends coordination among health data science initiatives to avoid duplicating funders’ investments. The Findings also describe identifying obstacles to tools and training, but neither passage addresses recording structurally closed opportunities, coordination between business-development and evidence teams, or repeated discovery of those opportunities. Searches for those terms and for recording barriers to avoid duplication found no supporting passage. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

May combine two claims.

This reduces rework and supports future submissions when similar calls reopen.

“This reduces rework and supports future submissions when similar calls reopen [3].”

• cites [3] Alami, H. et al. Artificial Intelligence Governance in Health Systems: Systematic Review of Frameworks and Integrative Model Proposal. J. Med. Internet Res. 28 (1), e87448. 10.2196/87448 (2026). sourceCould not verify

  The Methods (Eligibility Criteria) describe how AI-governance documents were selected, and the Results (Governance Processes) discuss monitoring and reviewer feedback. Neither addresses reducing rework or preparing future submissions when similar calls reopen. Searches for that phrasing and for funding applications or reusable submissions found no supporting passage; page numbers were not available in our copy. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

A total of 220 duplicates were detected across cycles.

“A total of 220 duplicates were detected across cycles (Figure 2).”

Probably not a substantial claim: may not pass S3 (load-bearing).

The Win Readiness Score was refined as a relative ranking tool across 49 grants.

“The Win Readiness Score was refined as a relative ranking tool across 49 grants.”

May combine two claims.

Over 50 search combinations and 200 opportunities were reviewed.

“Over 50 search combinations and 200 opportunities were reviewed.”

• Also stated in: § Phase 1: Opportunity discovery · plain

  “The pipeline explicitly tested more than 50 search combinations and reviewed more than 200 opportunities.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The scoring dashboard organized opportunities into actionable categories: advanced to next stage, submitted and awaiting response with score above 5%, submitted with score 3 to 5%, low-score exploratory submissions, ineligible or discarded, business-development outreach, and historical outcomes.

“The scoring dashboard organized opportunities into actionable categories: advanced to next stage, submitted and awaiting response with score above 5%, submitted with score 3 to 5%, low-score exploratory submissions, ineligible or discarded, business-development outreach, and historical outcomes.”

May combine two claims.

The team reported identifying approximately 18 suitable grants per scouting cycle after three search strategy changes, compared to approximately 5 to 6 per day under the prior approach.

“The team reported identifying approximately 18 suitable grants per scouting cycle after those changes, compared to approximately 5 to 6 per day under the prior approach.”

May combine two claims.

The issue documentation identified three recurring unblockers: a US national partner for federal opportunities, a European partner for European Union mechanisms, and Colombian educational institutions for domestic or academic routes.

“The issue documentation identified three recurring unblockers: a US national partner for federal opportunities, a European partner for European Union mechanisms, and Colombian educational institutions for domestic or academic routes.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The largest yield occurred on 2026-06-01, when 180 sources generated 76 pre-filter candidates, 43 post-filter candidates, and 30 net-new issues.

“The largest yield occurred on 2026-06-01, when 180 sources generated 76 pre-filter candidates, 43 post-filter candidates, and 30 net-new issues.”

May combine two claims.

Across 10 recorded scouting cycles between 2026-05-29 and 2026-06-05, the pipeline tested over 50 search combinations across 20 funding scenarios, processing 741 sources and identifying 352 pre-filter candidates.

“Across 10 recorded scouting cycles between 2026-05-29 and 2026-06-05, the pipeline tested over 50 search combinations across 20 funding scenarios (Table 1), processing 741 sources and identifying 352 pre-filter candidates.”

Probably not a substantial claim: may not pass S3 (load-bearing).

At the time of manuscript preparation, five opportunities had advanced to interview or finalist stages across a pipeline of 49 scored grants.

“At the time of manuscript preparation, five opportunities had advanced to interview or finalist stages across a pipeline of 49 scored grants.”

May combine two claims.

Of the pre-filter candidates, 83 were retained as post-filter candidates after applying eligibility and fit criteria, generating 46 net-new pipeline issues.

“Of these, 83 were retained as post-filter candidates after applying eligibility and fit criteria, generating 46 net-new pipeline issues.”

May combine two claims.

At the time of the rubric, five opportunities had advanced to a next stage, including three at interview stage and two at finalist stage.

“At the time of the rubric, five opportunities had advanced to a next stage, including three at interview stage and two at finalist stage.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The state file preserved the rationale for exclusion.

“The state file preserved the rationale for exclusion.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Among the submitted applications, one opportunity (MIT Solve) did not advance past the May 2026 cohort selection.

“Among the submitted applications, one opportunity (MIT Solve) did not advance past the May 2026 cohort selection; the next application cycle opens in September 2026, and the pipeline retains the funder profile and prior application artifacts for reuse.”

May combine two claims.

Time from the start-draft decision to submission ranged from approximately one to twenty-nine days, with a mean of roughly ten to twelve days.

“Time from the start-draft decision to submission ranged from approximately one to twenty-nine days, with a mean of roughly ten to twelve days, depending on form complexity and whether a partner or supporting letter was required.”

May combine two claims.

This preserved researcher judgment and made future searches more efficient.

“This preserved researcher judgment and made future searches more efficient.”

May combine two claims.

Under the manual process, a researcher scanning funder portals, newsletters, and referrals identified on the order of five to six eligible opportunities per working day and performed triage, scoring, deduplication, and record-keeping by hand.

“Under the manual process, a researcher scanning funder portals, newsletters, and referrals identified on the order of five to six eligible opportunities per working day and performed triage, scoring, deduplication, and record-keeping by hand.”

May combine two claims.

The large language model-supported components increased drafting and search capacity, but governance depended on explicit labels, issue comments, pull requests, and researcher review.

“The large language model-supported components increased drafting and search capacity, but governance depended on explicit labels, issue comments, pull requests, and researcher review.”

May combine two claims.

Under the skill-based workflow, the median time from discovery to a documented go/no-go decision was under one hour, with a mean on the order of eleven hours.

“Under the skill-based workflow, the daily scouting skill surfaced and pre-screened candidate opportunities automatically, and the median time from discovery to a documented go/no-go decision was under one hour, with a higher mean (on the order of eleven hours) reflecting a long tail of opportunities held for partner or eligibility clarification.”

Classified as probably not a substantial claim. May combine two claims.

Relevant opportunities may appear under AI, clinical decision support, digital health, health equity, antimicrobial resistance, oncology, payer innovation, women-led entrepreneurship, corporate innovation, regional development, compute access, or partner-route academic funding.

“Relevant opportunities may appear under AI, clinical decision support, digital health, health equity, antimicrobial resistance, oncology, payer innovation, women-led entrepreneurship, corporate innovation, regional development, compute access, or partner-route academic funding.”

May combine two claims.

As of the analytic cutoff date, 52 applications had been submitted during the operating period (March to June 2026), and 5 of the active opportunities had advanced to an interview or finalist stage.

“As of the analytic cutoff date, 52 applications had been submitted during the operating period (March to June 2026), and 5 of the active opportunities had advanced to an interview or finalist stage; subsequent repository activity was not included in this analysis.”

May combine two claims.

Approximately 88% of discarded opportunities were rejected for structural ineligibility rather than weak fit, and roughly half of the eligible opportunities required a partner or letter-of-intent route.

“Approximately 88% of discarded opportunities were rejected for structural ineligibility rather than weak fit, and roughly half of the eligible opportunities required a partner or letter-of-intent route.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The researcher review burden corresponded to four explicit human checkpoints per opportunity (eligibility, scoring, draft review, and submission readiness).

“The researcher review burden corresponded to four explicit human checkpoints per opportunity (eligibility, scoring, draft review, and submission readiness).”

Classified as probably not a substantial claim. May combine two claims.

Concretely, the workflow retains four human checkpoints, one in each phase: an eligibility check during discovery, a scoring and prioritization check during triage, a draft-review check before a proposal is considered ready, and a submission-readiness check before anything is sent to a funder.

“Concretely, the workflow retains four human checkpoints, one in each phase: an eligibility check during discovery, a scoring and prioritization check during triage, a draft-review check before a proposal is considered ready, and a submission-readiness check before anything is sent to a funder.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Opportunity pipeline state and Win Readiness Scores were exported from the scoring dashboard on 2026-06-10 and verified against repository records.

“Opportunity pipeline state and Win Readiness Scores were exported from the scoring dashboard on 2026-06-10 and verified against repository records.”

May combine two claims.

This study did not require informed consent or ethics committee approval.

“This study did not require informed consent or ethics committee approval.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

This structure made prioritization visible to the team and reduced reliance on subjective urgency alone.

“This structure made prioritization visible to the team and reduced reliance on subjective urgency alone.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Application history was reconstructed retrospectively from the submission tracker with a cutoff date of 2026-06-09.

“Application history was reconstructed retrospectively from the submission tracker with a cutoff date of 2026-06-09.”

Classified as probably not a substantial claim. May combine two claims.

The evidence supports a formative evaluation, not a causal claim about grant success.

“The evidence supports a formative evaluation, not a causal claim about grant success.”

Probably not a substantial claim: may not pass S3 (load-bearing).

All generated artifacts were version-controlled using Git with the repository hosted on GitHub (arkangelai/grants, private).

“All generated artifacts were version-controlled using Git with the repository hosted on GitHub (arkangelai/grants, private).”

May combine two claims.

Later runs showed lower net-new yield and higher duplication, consistent with saturation of the immediate search space.

“Later runs showed lower net-new yield and higher duplication, consistent with saturation of the immediate search space.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Scouting throughput was verified directly from the structured log across 10 recorded runs.

“Scouting throughput was verified directly from the structured log across 10 recorded runs.”

May combine two claims.

No patient or participant data were collected or analyzed.

“No patient or participant data were collected or analyzed.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The workflow uses repository commits and pull requests to preserve artifacts.

“The workflow uses repository commits and pull requests to preserve artifacts.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Version control makes recovery possible without relying on institutional memory.

“Version control makes this recovery possible without relying on institutional memory.”

Classified as probably not a substantial claim. May combine two claims.

Researcher review was required before any proposal was submitted or any scoring decision was acted upon.

“Researcher review was required before any proposal was submitted or any scoring decision was acted upon.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The Win Readiness Score introduced a transparent ranking mechanism.

“The Win Readiness Score introduced a transparent ranking mechanism.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The workflow description was reconstructed from internal and public sources.

“The workflow description was reconstructed from internal and public sources.”

Classified as probably not a substantial claim.

The curated scouting state file functioned as a memory layer for the pipeline.

“The curated scouting state file functioned as a memory layer for the pipeline.”

Classified as probably not a substantial claim. May combine two claims.

The score served as a disciplined way to allocate drafting effort rather than a guarantee of funding success.

“The score therefore served as a disciplined way to allocate drafting effort rather than a guarantee of funding success.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Drafting contains the fourth and final researcher checkpoint, a submission-readiness review, at which the research team verifies factual claims and citations, confirms that no confidential material is exposed, and approves the package before it is sent to the funder.

“Drafting therefore contains the fourth and final researcher checkpoint, a submission-readiness review, at which the research team verifies factual claims and citations, confirms that no confidential material is exposed, and approves the package before it is sent to the funder.”

Classified as probably not a substantial claim. May combine two claims.

The disposition label assigned to each opportunity is a human decision: although the triage skill proposes a label, the research team confirms or overrides it before the opportunity advances.

“The disposition label assigned to each opportunity is a human decision: although the triage skill proposes a label, the research team confirms or overrides it before the opportunity advances, so that eligibility interpretation and partnerroute judgments remain under human control.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Subsequent repository activity was not included in this analysis.

“Subsequent repository activity was not included in this analysis.”

May combine two claims.

Because these constraints are embedded in the skill modules rather than applied manually, they are enforced automatically on every invocation: each drafting skill verifies citations, blocks disclosure of confidential funder communications and unpublished strategy, and flags unresolved eligibility or partner dependencies for the researcher rather than fabricating content.

“Because these constraints are embedded in the skill modules rather than applied manually, they are enforced automatically on every invocation: each drafting skill verifies citations, blocks disclosure of confidential funder communications and unpublished strategy, and flags unresolved eligibility or partner dependencies for the researcher rather than fabricating content.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Teams adopting a similar workflow should expect to recalibrate scoring weights, partner routes, and triage criteria to their own organizational context.

“Teams adopting a similar workflow should expect to recalibrate scoring weights, partner routes, and triage criteria to their own organizational context.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

In this sense, grant development is part of implementation capacity, not merely fundraising administration.

“In this sense, grant development is part of implementation capacity, not merely fundraising administration.”

Probably not a substantial claim: may not pass S4 (based).

A governance layer spans all phases: researcher review checkpoints are required at the discovery, triage, scoring, and submission readiness stages.

“A governance layer spans all phases: researcher review checkpoints are required at the discovery, triage, scoring, and submission readiness stages.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Scoring constitutes the third researcher checkpoint.

“Scoring constitutes the third researcher checkpoint.”

Classified as probably not a substantial claim.

The workflow keeps researcher review as the decision point for scientific claims, confidential information, partner strategy, and submission readiness.

“Large language models support summarization, section drafting, narrative adaptation, and consistency checks, but the workflow keeps researcher review as the decision point for scientific claims, confidential information, partner strategy, and submission readiness.”

Classified as probably not a substantial claim. May combine two claims.

Interactive review, eligibility interpretation, partner strategy, and submission readiness decisions were performed by the research team in a human-in-the-loop role.

“Interactive review, eligibility interpretation, partner strategy, and submission readiness decisions were performed by the research team in a human-in-the-loop role.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Large language models can support parts of this work, but published guidance emphasizes caution, fact-checking, and human ownership of grant strategy.

“Large language models can support parts of this work, but published guidance emphasizes caution, fact-checking, and human ownership of grant strategy.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Skills are stored and versioned alongside the repository, which means that changes to the workflow are reviewed and traceable in the same way as any other artifact.

“Skills are stored and versioned alongside the repository, which means that changes to the workflow are reviewed and traceable in the same way as any other artifact.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Triage is the second researcher checkpoint.

“Triage is the second researcher checkpoint.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

A proposal that uses large language models for discovery or drafting but cannot trace its factual claims to verified sources, or that was submitted without explicit researcher review, creates both integrity risk and reputational risk for the applying organization.

“A proposal that uses large language models for discovery or drafting but cannot trace its factual claims to verified sources, or that was submitted without explicit researcher review, creates both integrity risk and reputational risk for the applying organization.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

It helps identify which funders support evaluation, which require local or international partners, which opportunities are too low-fit, and which narratives need more evidence before submission.

“It helps identify which funders support evaluation, which require local or international partners, which opportunities are too low-fit, and which narratives need more evidence before submission.”

Classified as probably not a substantial claim. May combine two claims.

The score serves as a relative ranking and kill-or-go tool.

“Instead, it serves as a relative ranking and kill-or-go tool.”

Probably not a substantial claim: may not pass S2 (checkable).

A grant pipeline is one way to manage those barriers.

“A grant pipeline is one way to manage those barriers.”

Classified as probably not a substantial claim.

A governed workflow that uses large language models for search expansion and drafting while preserving researcher review and traceability can partially substitute for the research development operations resources that many health-AI teams in low-and middleincome countries cannot maintain.

“A governed workflow that uses large language models for search expansion and drafting while preserving researcher review and traceability can partially substitute for the research development operations resources that many health-AI teams in low-and middleincome countries cannot maintain.”

• AssumptionRests on an assumption: Health-AI organizations frequently operate without dedicated research development offices or grant operations infrastructure.

Probably not a substantial claim: may not pass S4 (based).

Each skill embeds the governance rules described below, so that constraints such as citation verification and confidentiality are enforced at every invocation rather than left to the operator's memory.

“Each skill embeds the governance rules described below, so that constraints such as citation verification and confidentiality are enforced at every invocation rather than left to the operator's memory.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Metrics are reported using predefined cutoffs: structured scouting metrics cover ten cycles from 2026-05-29 through 2026-06-05; application-history and advancement metrics are reported through 2026-06-09; and repository state labels were exported on 2026-06-10.

“Metrics are reported using predefined cutoffs: structured scouting metrics cover ten cycles from 2026-05-29 through 2026-06-05; application-history and advancement metrics are reported through 2026-06-09; and repository state labels were exported on 2026-06-10.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

For health-AI teams in low-and middle-income countries, the workflow has broader relevance.

“For health-AI teams in low-and middle-income countries, the workflow has broader relevance.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

No patient-level data, protected health information, funder-confidential peer review information, or private applicant records were used for this manuscript.

“No patient-level data, protected health information, funder-confidential peer review information, or private applicant records were used for this manuscript.”

Probably not a substantial claim: may not pass S3 (load-bearing).

It also separates structural discards from paused opportunities.

“It also separates structural discards from paused opportunities.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The curated scouting state file preserves institutional memory.

“A curated scouting state file preserves institutional memory.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

These estimates are derived from internal workflow records and were not independently validated; they should be interpreted as retrospective operational estimates, not as controlled measurements.

“These estimates are derived from internal workflow records (issue timestamps, commit histories, and submission logs) and were not independently validated; they should be interpreted as retrospective operational estimates, not as controlled measurements.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The decision to commit drafting effort to a given opportunity, or to set it aside, is taken by the research team rather than by the score alone.

“The Win Readiness Score orders opportunities, but the decision to commit drafting effort to a given opportunity, or to set it aside, is taken by the research team rather than by the score alone.”

Classified as probably not a substantial claim.

This prevents the same ineligible or strategically weak opportunities from being repeatedly rediscovered.

“This prevents the same ineligible or strategically weak opportunities from being repeatedly rediscovered.”

Classified as probably not a substantial claim.

This breadth was necessary because the relevant opportunity set for a Colombian health-AI company is fragmented across scientific funding, innovation funding, digital health programs, corporate partnerships, public-sector mechanisms, and regional initiatives.

“This breadth was necessary because the relevant opportunity set for a Colombian health-AI company is fragmented across scientific funding, innovation funding, digital health programs, corporate partnerships, public-sector mechanisms, and regional initiatives.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The score is deliberately not described as a calibrated probability because the organization does not yet have enough outcome data to estimate true win probability.

“The score is deliberately not described as a calibrated probability because the organization does not yet have enough outcome data to estimate true win probability.”

Classified as probably not a substantial claim.

Governance components address these risks not by restricting model use but by adding structured accountability at each step.

“Governance components address these risks not by restricting model use but by adding structured accountability at each step.”

Classified as probably not a substantial claim.

The workflow converts grant discovery from an ad hoc search process into a traceable pipeline with structured triage, decision-aware deduplication, relative scoring, and repository-based artifact preservation.

“The workflow converts grant discovery from an ad hoc search process into a traceable pipeline with structured triage, decisionaware deduplication, relative scoring, and repository-based artifact preservation.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

Large language models can then support search expansion, summarization, first-pass drafting after human framing, and consistency checks.

“Large language models can then support search expansion, summarization, first-pass drafting after human framing, and consistency checks.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

An organization also needs operating infrastructure: opportunity state, source tracking, scoring, versioned drafts, partner-route memory, and submission records.

“An organization also needs operating infrastructure: opportunity state, source tracking, scoring, versioned drafts, partner-route memory, and submission records.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Artifacts generated at each phase are preserved in version-controlled repositories.

“Artifacts generated at each phase are preserved in version-controlled repositories.”

May combine two claims.

Researcher review remained necessary for eligibility interpretation, partner strategy, scientific claims, privacy constraints, final submission readiness, and decisions to discard or monitor an opportunity.

“Researcher review remained necessary for eligibility interpretation, partner strategy, scientific claims, privacy constraints, final submission readiness, and decisions to discard or monitor an opportunity.”

Classified as probably not a substantial claim. May combine two claims.

The analysis draws on internal operational artifacts from the organization's grants repository and associated issue records, plus public literature on large language model-assisted grant writing and digital health sustainability.

“The analysis draws on internal operational artifacts from the organization's grants repository and associated issue records, plus public literature on large language model-assisted grant writing and digital health sustainability.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

This design keeps the human researcher in control of state transitions while delegating the repetitive drafting and bookkeeping work to the skills.

“This design keeps the human researcher in control of state transitions while delegating the repetitive drafting and bookkeeping work to the skills.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A researcher review is required before any proposal is submitted.

“A researcher review is required before any proposal is submitted.”

Probably not a substantial claim: may not pass S3 (load-bearing).

These are treated as structural pipeline constraints rather than one-off narrative edits.

“These are treated as structural pipeline constraints rather than one-off narrative edits.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

That guidance is necessary but incomplete for organizations.

“That guidance is necessary but incomplete for organizations.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

The workflow is not inherently limited to low-resource contexts; the same governance, triage, scoring, and artifact-preservation logic applies to any organization managing a fragmented funding pipeline.

“The workflow is not inherently limited to low-resource contexts; the same governance, triage, scoring, and artifact-preservation logic applies to any organization managing a fragmented funding pipeline.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Privacy and confidentiality constraints also limit how much actual proposal text and partner strategy can be shared in a journal manuscript.

“Privacy and confidentiality constraints also limit how much actual proposal text and partner strategy can be shared in a journal manuscript.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Not all portfolio outcomes are attributable to these 10 cycles; earlier discovery and triage activity also contributed to the current pipeline state.

“Not all portfolio outcomes are attributable to these 10 cycles; earlier discovery and triage activity also contributed to the current pipeline state.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The operational scouting logs and pipeline state data that support the findings of this study are available from the corresponding author upon reasonable request. The data are not publicly available due to their proprietary nature and the inclusion of unpublished grant application materials.

“The operational scouting logs and pipeline state data that support the findings of this study are available from the corresponding author upon reasonable request.The data are not publicly available due to their proprietary nature and the inclusion of unpublished grant application materials.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

Version-controlled repositories provide the infrastructure; the governance rules provide the discipline.

“Version-controlled repositories provide the infrastructure; the governance rules provide the discipline.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Other teams would need to adapt the rubric, and results should not be generalized to teams with different institutional contexts or funding environments.

“Other teams would need to adapt the rubric, and results should not be generalized to teams with different institutional contexts or funding environments.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

This research received no external funding. The workflow described was developed as part of the internal research and evidence development operations of Arkangel AI.

“This research received no external funding.The workflow described was developed as part of the internal research and evidence development operations of Arkangel AI.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The structured scouting log offers granular, run-level data on search yield, filtering, and deduplication.

“The structured scouting log offers granular, run-level data on search yield, filtering, and deduplication.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

These baseline values are retrospective operational estimates derived from internal workflow records and were not independently validated.

“These baseline values are retrospective operational estimates derived from internal workflow records and were not independently validated.”

May combine two claims.

Repository metrics reflect operational records maintained for workflow management rather than prospectively collected research data, and denominators and time windows vary across metrics as specified in the Methods.

“Repository metrics reflect operational records maintained for workflow management rather than prospectively collected research data, and denominators and time windows vary across metrics as specified in the Methods.”

Classified as probably not a substantial claim. May combine two claims.

The Win Readiness Score includes heuristic inputs, since public acceptance rates are not always available and adjustments for for-profit or foreign entities can only approximate funder behavior; the score should remain a prioritization tool until sufficient outcome data allow calibration.

“The Win Readiness Score includes heuristic inputs, since public acceptance rates are not always available and adjustments for for-profit or foreign entities can only approximate funder behavior; the score should remain a prioritization tool until sufficient outcome data allow calibration.”

May combine two claims.

The workflow converts broad grant searches into a managed pipeline with structured triage, decision-aware deduplication, relative scoring, artifact preservation, and explicit researcher review checkpoints at every decision stage.

“The workflow converts broad grant searches into a managed pipeline with structured triage, decision-aware deduplication, relative scoring, artifact preservation, and explicit researcher review checkpoints at every decision stage.”

Classified as probably not a substantial claim.

It describes a single-company workflow whose categories, scoring weights, and partner routes reflect one organization's products, geography, and evidence portfolio.

“It describes a single-company workflow whose categories, scoring weights, and partner routes reflect one organization's products, geography, and evidence portfolio.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Some source material was missing from the repository at the time of drafting, and this analysis relied on the structured metrics available rather than expanding unsupported transcript details.

“Some source material was missing from the repository at the time of drafting, and this analysis relied on the structured metrics available rather than expanding unsupported transcript details.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

AI-assisted drafting tools were used to prepare initial text under human review; the corresponding author takes full responsibility for the integrity of the content.

“AI-assisted drafting tools were used to prepare initial text under human review; the corresponding author takes full responsibility for the integrity of the content.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

This study did not require ethics committee approval or participant consent. No patient or participant data were collected or analyzed. The workflow description draws exclusively on organizational process artifacts, internal repository records, version histories, and issue-level workflow logs, with no identifiable personal or clinical information. This type of methodological process description is

“This study did not require ethics committee approval or participant consent.No patient or participant data were collected or analyzed.The workflow description draws exclusively on organizational process artifacts, internal repository records, version histories, and issue-level workflow logs, with no identifiable personal or clinical information.This type of methodological process description is”

• cites [15] Colombia Ministerio de Salud. Resolution 8430 of 1993, establishing scientific, technical and administrative standards for health research (Ministerio de Salud, 1993).Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim. May combine two claims.

The formative evaluation was conducted by members of the development and writing team, which limits independence; however, internal formative evaluation is appropriate at this stage because the objective was to describe system design and identify governance gaps.

“The formative evaluation was conducted by members of the development and writing team, which limits independence; however, internal formative evaluation is appropriate at this stage because the objective was to describe system design and identify governance gaps.”

May combine two claims.

Early operational data support the feasibility of this approach for a Colombian health-AI company, but do not establish that automation improves grant success rates.

“Early operational data support the feasibility of this approach for a Colombian health-AI company, but do not establish that automation improves grant success rates.”

May combine two claims.

The throughput improvement from approximately 5 to 6 grants per day to 18 per cycle reflects a change in search strategy rather than a controlled comparison and should not be interpreted as a causal estimate of workflow impact.

“The throughput improvement from approximately 5 to 6 grants per day to 18 per cycle reflects a change in search strategy rather than a controlled comparison and should not be interpreted as a causal estimate of workflow impact.”

May combine two claims.

The current evaluation is operational, not outcome-based: throughput, issue creation, and duplicate detection do not prove higher proposal quality or funding success.

“The current evaluation is operational, not outcome-based: throughput, issue creation, and duplicate detection do not prove higher proposal quality or funding success.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The analyzed cases represent completed writing programs and may not capture framework behavior during projects that were paused, canceled, or substantially redirected by human decision-making.

“The analyzed cases represent completed writing programs and may not capture framework behavior during projects that were paused, canceled, or substantially redirected by human decision-making.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Katherine Monsalve, Natalia Castaño, Jose Zea, and Laura Velásquez are employees or co-founders of Arkangel AI, the organization described in this paper. This constitutes a potential competing interest and is declared here.

“Katherine Monsalve, Natalia Castaño, Jose Zea, and Laura Velásquez are employees or co-founders of Arkangel AI, the organization described in this paper.This constitutes a potential competing interest and is declared here.”