Methodology

Tool-agnostic where possible; the reference implementation uses Claude/Cowork, GPT, REDCap, and Python/React.

Design principles

• Human-in-the-loop. Every AI output is reviewed and approved by an expert before it affects the study.
• Separation of roles. Drafting and review are kept distinct — investigators (with Cowork assistance) author; an independent model (GPT-5.5) reviews — reducing self-confirmation bias.
• Reproducibility by construction. Prompts, instrument definitions, code and analysis are version-controlled and archived with a DOI.
• Privacy & sovereignty (RODO/GDPR). No PII enters the repository; anonymisation and local-first processing are defaults.
• Transparency. AI involvement is disclosed per artifact, in line with open-science norms and EU AI Act transparency expectations.

Stage 1 — Design & orchestration (investigators)

The investigators (authors) own this stage: research questions, protocol, item content, registration artifacts (ClinicalTrials.gov / OSF), and consent / RODO text. Claude / Cowork may help draft and scaffold these artifacts, but the investigators decide, edit, and approve everything.

Stage 2 — Independent item review (GPT-5.5, LLM-as-judge)

An independent LLM (GPT-5.5) evaluates each item against explicit criteria: clarity, ambiguity, double-barrelled phrasing, leading wording, cultural bias, response-option balance, and content-validity relevance. It returns structured flags and suggested fixes that loop back to the investigators, who adjudicate. Where formal content validity is required, LLM relevance ratings may be reported alongside — never instead of — a human expert panel (I-CVI / CVI).

Stage 3 — Instrument build & data capture (Claude / Cowork + REDCap)

Once the investigators approve the items, Claude / Cowork builds and operates the instrument in REDCap — data dictionary, branching logic, consent, RODO collapsible, validation rules — via MCP / browser automation, under investigator direction. REDCap provides access control, an audit trail, and API access. De-identified exports feed Stage 4.

Stage 4 — Analytics dashboards (Python / React)

AI assists in generating reproducible analysis code and interactive dashboards for descriptive statistics, hypothesis tests, and figures, published via GitHub Pages and showing aggregated, de-identified data only.

Quality control — the "Judge" pass

Each stage carries a self-audit: source quality, double-counting / leakage, outdated assumptions, and unjustified extrapolation. AI-derived artifacts are explicitly labelled and separated from expert-verified ones.