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 — vague hypotheses, hidden incrementalism, missing risk logic, weak panel fit, CV–project disconnect — before workshop time is spent on them. It does not replace human review.
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 Annex 1 Criterion 1 and Criterion 2 sub-questions from the
ERC Guide for Peer Reviewers for Starting and Consolidator Grant
calls, plus source-checked notes on scoring scales, proposal
structure, eligibility windows, panel structure, and the ERC's
AI-use rules. The prompt loads the rubric, applies structural caps
that fire when an expected element is missing, and produces an
evidence-grounded critique.
This package is the ERC-specific application of two general
research-audit projects by the same author group:
tjhavranek/research-audit-duel-protocol
(manual multi-model adversarial-debate protocols) and
tjhavranek/mad-research
(Claude Code skills that automate the same audit discipline for
Claude + Codex). The Advanced pre-review here is a direct
extension of the latter; the multi-model convergence walkthrough
in docs/multi_model_walkthrough.md
is a direct application of the former.
- Open a paid chat session with a current frontier model. Confirm training opt-out is on. (Provider settings change; see Privacy below.)
- Copy the text between the
=== PROMPT BEGIN ===and=== PROMPT END ===markers inbasic/prompt.md. Paste it into the chat. - 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).
- Read the output. The model returns the top findings ranked by severity, each grounded in a verbatim quote from your draft; which structural caps fired; a short list of what the next draft stage needs to add; and a paragraph you can attach when sending the draft to a workshop reader.
Use a paid tier with training opt-out enabled, or remove names and
institutions before pasting a proposal draft into any chat product.
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.
The tool is calibrated for ERC Starting and Consolidator Grants only — not Advanced, Synergy, Proof of Concept, or ERC Plus. It is in English; the ERC working language of evaluation is English. It catches routine structural problems and does not judge novelty in a specific field, panel strategy, or interview credibility — these 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. The version in
shared/rubric_locked.md was last
verified on 2026-05-30 against ERC Work Programme 2026 and the
Guide for Peer Reviewers v6.0 (30 September 2025). If you read
this twelve months later, re-verify before use.
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.
The single-prompt pre-review above is what most applicants will use. The repository also contains:
- 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 Claude Code + Codex CLI integration in
advanced/that runs themad-researchprotocol against the same ERC rubric (experimental; not yet end-to-end pilot-verified — seeadvanced/README.md). - 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.
Synthetic test fixtures used to exercise the prompts are in
tests/fixtures/. Roadmap and development
notes are in docs/roadmap.md.
Tomáš Havránek, Charles University, Prague. 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. 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.