Project summary

I am seeking USD 15,000 for a six-month first phase of AI-assisted scientific criticism. My goal is to use advanced AI models, computational verification, and invited subject experts to revisit and improve human-authored scientific papers, particularly papers claiming to solve longstanding open problems or conjectures.

The project has a general remit: constructive scientific criticism across disciplines, covering theoretical arguments, methods, calculations, and empirical claims where suitable data and expertise are available. It is open to ordinary research papers as well as major contributions. Papers published before widespread use of generative AI, and papers claiming to resolve open problems or conjectures, will be priorities rather than eligibility restrictions. This first grant will support a selected set of cases and establish a method that can expand with further funding and collaborators.

Relationship to MathHelm: this project is complementary in scientific purpose and reverses the main direction of review. MathHelm emphasizes discussion and expert assessment of AI-generated mathematical papers. Our project brings advanced AI and computational tools into expert-led assessment of human-authored scientific papers. Both approaches contribute to reliable research through scrutiny of arguments and evidence. Here, AI proposes questions and possible issues, while human researchers and invited specialists verify the findings. The aim is to improve papers by clarifying assumptions, strengthening proofs, reproducing calculations, correcting demonstrated errors, and identifying useful extensions. Complementarity describes these shared objectives; this is an independent initiative without a claimed formal partnership or endorsement from MathHelm.

What are this project's goals? How will you achieve them?

I will lead an initial program centered on mathematical analysis and dynamical systems, extending to physics, engineering, or other scientific areas when appropriate domain experts agree to participate. I will invite experts for paid, bounded verification assignments. No expert participation is currently confirmed. Conclusions outside my expertise will require qualified domain review.

The initial target is 10 detailed case reports over six months, each focused on specified claims rather than certification of an entire paper. Selection will consider scientific value, opportunities for clarification or improvement, availability of the paper and necessary data, tractability, and reviewer availability. Claims to resolve open problems or conjectures will receive particular attention, but are not required for inclusion. Cases will be chosen for scientific value, not an expectation that their authors made mistakes. Previously investigated correction cases will be clearly separated from newly selected cases.

Advanced AI models will assist with reading proofs, checking assumptions, proposing counterexamples, comparing arguments, and designing reproducibility checks. Funding will cover access to existing advanced models and API services, not training a new foundation model. Models will be selected using task-specific trials and cost limits; their names, versions, prompts, dates, and relevant outputs will be recorded. Agreement between models will not count as independent proof.

Candidate findings will be checked using mathematical reasoning, symbolic computation, numerical experiments, and expert assessment. Lean formalization will be attempted for selected tractable mathematical statements where useful, not promised for whole papers. For empirical work, reanalysis will depend on adequate lawful access to data and methods. Numerical evidence, formal proof, expert judgment, and unresolved questions will be distinguished.

Month 1: establish the review protocol, shortlist papers, invite experts, and trial model workflows. Months 2-4: conduct focused audits and computational checks. Months 5-6: complete expert checks where available, invite author responses, develop constructive corrections or clarifications, and release case reports and a synthesis. Alongside the reviews, we will configure and test a small forum during months 2-4 and assess a moderated pilot during months 5-6, subject to recruiting suitable contributors.

Proposed forum: SciRevia (working name)

We also propose developing a public discussion forum for AI-assisted scientific criticism, with the long-term ambition of a large international community across scientific disciplines. SciRevia is a proposed short name combining science, review, and AI; it is subject to final naming and availability checks. Inspired by the value of a dedicated discussion space such as MathHelm, the forum would provide a complementary setting for expert-led, AI-assisted examination and improvement of human-authored research. No affiliation with MathHelm is implied.

Paper-specific discussion threads would bring together authors, reviewers, and invited specialists to examine proofs, calculations, methods, corrections, and new ideas. Contributions would include evidence, reproducible code where applicable, clear disclosure of AI assistance, and explicit labels distinguishing open questions from verified findings. Authors would have opportunities to respond, and corrections to forum claims would remain visible.

Operating the forum requires funding for setup and maintenance, hosting, moderators, and specialist verification. Moderators would manage discussion quality, conflicts of interest, and reports of problematic content; subject experts would assess scientific arguments. Moderation would not itself constitute scientific validation.

Within the USD 15,000 first phase, we propose a modest pilot using established forum software, with limited paid moderation and expert-review hours. A large, continuously staffed international forum would require additional funding. The pilot would test demand, participation, moderation workload, review quality, and operating costs before expansion. No forum or moderator appointments currently exist.

How will you know it worked?

Outputs will include a public protocol, a target of 10 detailed case reports, reproducible scripts and formal proof files where produced, and a final report on costs, time, useful findings, false alarms, and limitations. The forum pilot will also report participation, substantive discussions, moderation hours, and operating costs to guide any expansion. Every case report will record the scope of checking and distinguish confirmed findings, rejected AI criticisms, unresolved issues, and findings awaiting independent review.

For newly selected claims, I will record an initial manual assessment before AI assistance and track what the AI adds. This is an exploratory comparison, not a controlled demonstration that AI outperforms experts. Publicly available older papers may have been included in model training, so results will not be presented as a clean test on unseen material.

Success includes clearer arguments, reproducible calculations, justified confidence in checked claims, and corrections where warranted. The project does not promise to discover errors, resolve new conjectures, or secure author agreement or journal acceptance. If expert availability or case difficulty reduces the number of completed reports, I will explain the change and document unresolved work.

How will this funding be used?

The funding request and minimum required for this proposed phase are both USD 15,000. The proposed allocation is:

Research lead time and coordination: USD 4,500, budgeted as 225 hours at USD 20/hour.

Invited domain-expert verification: USD 3,500, budgeted as approximately 70 hours at an illustrative average of USD 50/hour; actual assignments and rates will be agreed with reviewers.

Advanced AI subscriptions and API usage: USD 2,000, with spending caps and recorded usage.

Computational resources and reproducibility: USD 750.

Forum setup, hosting, backups, and technical maintenance: USD 1,750, using established software.

Forum moderation and community coordination: USD 1,500, budgeted as approximately 75 hours at USD 20/hour; this provides limited coverage rather than continuous staffing.

Documentation, repository hosting, and dissemination: USD 500.

Contingency for additional verification or compute: USD 500.

Total: USD 15,000. Work will begin after funding is received. No journal publication fees are included. Material changes to the budget, expert participation, or scope will be reported. Any unspent funds will be handled under the grant agreement.

Who is on your team? What's your track record on similar projects?

I am Rafik Zeraoulia, a mathematics researcher based in Algeria with teaching experience at Abbes Laghrour University, Khenchela. My research interests include mathematical analysis and dynamical systems, and my work includes examining mathematical arguments through counterexamples and computational checks. I will coordinate the project and conduct reviews within my expertise.

This proposal builds on my experience using AI to examine 200 research papers. In this work, I have used AI to assist with proof correction, computational checks, and the analysis of theorems and their assumptions. I have found this approach successful in supporting constructive scientific criticism and improvements to research arguments. Through correspondence about this work, I have received very positive responses from mathematicians and physicists around the world, including agreement with corrections and observations I shared. This experience motivates me to develop a more systematic, reproducible initiative supported by advanced AI models and invited domain experts.

I plan to invite mathematicians and other subject specialists to check selected findings and help improve the papers. The budget includes expert honoraria, but no named expert or institution is currently committed to the project. The longer-term objective is a collaborative initiative across science, developed as expertise and funding become available.

What are the most likely causes and outcomes if this project fails?

AI can invent objections, misunderstand assumptions, repeat existing commentary, or miss serious gaps. Verification may take longer than expected, data may be unavailable, and suitable experts may decline invitations. To manage these risks, the project will use bounded claims, spending limits, explicit evidence levels, and human checks. Unresolved concerns will remain labeled unresolved.

The purpose is to improve papers and scientific understanding. Authors will have an opportunity to respond to substantiated concerns before public release, and relevant responses and subsequent corrections will accompany the reports. Criticism will address evidence and arguments without inferring misconduct or motives. Only public or appropriately authorized materials will be used; confidential review manuscripts and sensitive personal data will not be uploaded to AI services.

How much money have you raised in the last 12 months, and from where?

I have not received grant funding. I have submitted other applications for related AI and mathematics research. Funding received for overlapping activities will be disclosed, and budgets or scope will be adjusted to prevent double funding of the same costs.

AI assistance disclosure

AI assisted with drafting this proposal and will support the proposed research. I remain responsible for the final proposal and reports. AI-generated criticisms will be treated as hypotheses to verify, not established findings.

All uses of AI assistance in the project will be disclosed in the relevant reports and publications, including assistance with drafting, proof analysis, theorem development, coding, computation, and generation of research ideas. Disclosures will identify the tools and model versions where available, explain their roles, and describe the human verification performed. Human researchers retain responsibility for the accuracy and integrity of the final work.

Conclusion

Our goal is to improve scientific papers, help researchers strengthen the correctness and clarity of their reasoning, and develop new ideas through constructive examination of existing research. By combining advanced AI, reproducible computation, and human expertise, we aim to turn critical review into a collaborative contribution to scientific progress. Complementary to MathHelm's emphasis on expert assessment of AI-generated mathematics, this initiative applies AI-assisted human review to human-authored scientific research. Transparent disclosure of AI assistance and accountable human verification will be central to every stage of the project.