erc-ai-feedback

Independent and unofficial. Not produced, endorsed, or approved by the European Research Council or ERCEA.

This is a small package for ERC Starting and Consolidator Grant applicants. It provides a rubric-based pre-review of a draft proposal that the applicant runs in one chat session with a current frontier model. The intent is to clear routine structural problems before workshop time is spent on them: vague hypotheses, no risk plan, weak panel fit, a CV that does not connect to the project. It does not replace human review.

Web page: https://tjhavranek.github.io/erc-ai-feedback/. Its Copy buttons load each prompt directly from this repository, so the copied text cannot drift from the files.

It is for applicants, and for workshop preparation the applicant has authorised. Nothing else. If you are evaluating a proposal for the ERC, do not use this or any AI tool on it. The ERC's guidelines for panel members and remote reviewers rest on two principles, confidentiality and non-delegation. The second is not discharged by keeping the data safe: the guidelines answer "no" to a model run locally (Q2), to a tool that discloses nothing to a third party (Q3), and to asking an AI whether your review is missing points (Q8). Running it offline is not a workaround.

Applicants are in a different position, and this package is built to sit inside it. The ERC Scientific Council "recognises that researchers regularly seek input from AI technologies or human third parties, for example to brainstorm or generate ideas, to search the literature, and to revise, translate or summarise text", and requires only that this "does not relieve the author from taking full and sole authorship responsibilities with regard to acknowledgements, plagiarism and the practice of good scientific and professional conduct". The ERC sets no disclosure wording of its own; a host institution may. The Basic pre-review stays deliberately inside that permission and does less than it allows: it identifies problems in a draft and never writes proposal content, so the words submitted to the ERC remain the applicant's own.

Two files do the work: the locked evaluation rubric in shared/rubric_locked.md and the prompt in basic/prompt.md. The rubric contains the verbatim Criterion 1 and Criterion 2 sub-questions applied to the 2027 Starting and Consolidator Grant calls, taken from ERC Work Programme 2027 and Information for Applicants v11.0, plus source-checked notes on scoring scales, proposal structure, eligibility windows, resubmission restrictions, panel structure, and the ERC's AI-use rules. The prompt embeds a condensed extract of that rubric, applies structural caps that fire when an expected element is missing, and produces an evidence-grounded critique. The caps are this package's own heuristics, not ERC rules.

The Advanced pre-review extends tjhavranek/mad-research, a general research-audit project by the same authors. A sibling project, tjhavranek/paper-workshop, does the same kind of adversarial review for manuscripts, but it writes edits, so it is not suitable for an ERC proposal. Related tools and the reasoning behind them are listed in docs/roadmap.md.

Use

1. Open a paid chat session with a current frontier model, and select the provider's strongest reasoning setting rather than the fast default. The prompt asks the model to check several conditions at once and to tie every finding to a quote from a long document; a fast model does that noticeably less reliably. Confirm training opt-out is on. (Provider settings change; see Privacy below.)
2. Copy the text between the === PROMPT BEGIN === and === PROMPT END === markers in basic/prompt.md, or use the Copy button on the web page, which copies the same text. Paste it into the chat.
3. Paste your current draft after the prompt. State which stage you are at:
   • Pitch: idea sketch, or abstract plus draft CV.
   • Part I + CV: Extended Synopsis (Part I of the Scientific Proposal) in progress, with a draft CV.
   • Full proposal: complete B1 (cover-page abstract, Part I, CV and track record), plus any available B2 draft (Part II; required for final ERC submission).
4. Read the output. It opens with a plain-language summary meant to stand on its own. Below that: the top findings ranked by severity, each grounded in a verbatim quote from your draft; which structural score caps fired; what the next draft stage needs to add; and a paragraph you can attach when sending the draft to a workshop reader. The risk level and the provisional marks are the model's reading of structural completeness, not a prediction of how the ERC will score you.

Privacy

Use a paid tier with training opt-out enabled. That is a floor rather than a substitute for judgement about what you paste. Payment is not a confidentiality control: the provider still receives the text, and opt-out is not the same as zero retention or deletion. Remove names and institutions you do not need, including other people's. Provider UI labels change every few months, so check the provider's current data-controls page rather than relying on setting names copied here: ChatGPT data controls, Anthropic Claude privacy, and Gemini activity and privacy. Pasting an unpublished draft into any chat product carries residual risk even when training is off. If the proposal contains embargoed, patent-sensitive, or otherwise non-public material, sanitise it or do not use the tool. If someone other than the applicant runs the pre-review, follow the consent rule in docs/for_workshop_leaders.md.

Limitations

The tool is calibrated for ERC Starting and Consolidator Grants only, not Advanced, Synergy, Proof of Concept, or ERC Plus. It is in English. That is a limit of this tool rather than an ERC rule: an application may formally be submitted in any official EU language, but English is the working language of the evaluation panels. It is designed to flag routine structural problems and does not judge novelty in a specific field, panel strategy, or interview credibility. Those remain workshop, mentor, and research-office work. Frontier models confidently miss field-specific failure modes; the applicant should push back on findings they disagree with on merit.

The ERC rubric changes annually, and the 2027 changes are already in force. The version in shared/rubric_locked.md was last verified on 2026-09-17 against ERC Work Programme 2027 (adopted 20 July 2026) and Information for Applicants v11.0 (22 July 2026). Note that sub-question 1.2 gained a third clause for the 2027 calls and the eligibility windows widened. The Guide for Peer Reviewers is still v6.0, stamped for the 2026 calls; a v7.0 is expected before the Starting Grant deadline. Re-verify before use if you are reading this after the current call cycle.

For workshop leaders

If you are running an ERC workshop and want to recommend the tool to applicants, the operational note at docs/for_workshop_leaders.md covers the hand-off template and the consent rule that applies when a reader runs the tool on someone else's draft.

Other ways to use the package

The single-prompt pre-review above is what most applicants should use. The alternatives below cost more of your time, and Advanced also sends the draft to a second provider. Run them as cross-checks on a draft that has already been through the pre-review, not instead of it.

• A three-prompt sequence in standard/ for applicants who want a more thorough single-model pass (Panel Reviewer, then Devil's Advocate, then Writing Coach).
• A question-bank and answer stress-test prompt pair in mock_interview/ for Step-2 applicants preparing for the panel interview.
• A resubmission audit prompt in resubmission/ that checks whether a new draft engages with a previous Evaluation Report.
• A Claude Code + Codex CLI integration in advanced/ that runs the mad-research protocol against the same ERC rubric. Experimental, not yet end-to-end pilot-verified, and the only mode that sends the draft to two providers. A pre-registered study by this package's authors found a single pass preferred over that protocol on a different kind of document. See advanced/README.md.

Synthetic test fixtures used to exercise the prompts are in tests/fixtures/. Roadmap and development notes are in docs/roadmap.md.

Citation, licence, maintainer

Developed by Tomáš Havránek and Zuzana Iršová, Charles University, Prague. Further tools and materials from the same research group are collected at meta-analysis.cz. MIT licence; see LICENSE. The ERC materials cited in the rubric remain the property of the European Research Council Executive Agency and are referenced under fair use with source URLs in shared/rubric_locked.md. Archived on Zenodo with a DOI that resolves to the latest release: https://doi.org/10.5281/zenodo.20829165. For citation metadata see CITATION.cff.

Bug reports and factual corrections to the rubric: GitHub Issues. Please do not paste real proposal content in public issues. ERC-panel-veteran feedback on the prompt is particularly welcome.

Roadmap and development notes are in docs/roadmap.md. The audit history of each version is in CHANGELOG.md.