How we review submissions
Before you submit, know exactly what happens to your preprint — the automated AI review, the AI quality score, and the human decision. No surprises.
The short version
- You submit — with your verified GitHub or ORCID identity, which signs your submission as its responsible author.
- Automated screening runs immediately: security checks, spam detection, and an AI scientific review that produces an AI quality score.
- A human moderator decides. The AI's score is advisory — the final accept/reject is always made by a person.
- You get an email with the outcome, and (on rejection) the score breakdown so you know what to improve.
The three screening layers
Every submission passes through three automated gates before a human sees it:
1. Security
Deterministic checks. LaTeX sources are scanned for shell escapes, file access, and non-whitelisted packages; Markdown for scripts and dangerous URLs. Insecure sources are never compiled or served.
2. Content quality
Heuristic filters for spam and low-effort text: lexical diversity, repetition, entropy, and known spam signals. High-confidence spam is blocked; borderline cases are flagged for the human.
3. Scientific review (AI)
An AI scientific reviewer — running the DiDAL review protocol — evaluates the text against a fixed rubric and produces the AI quality score. Details below.
What is the AI quality score?
The AI quality score is a number from 1 to 10, produced by an AI reviewer, that rates the scientific quality and likely impact of the review as written. It is an advisory assessment of the text, not a judgment about you and not a prediction of real citation counts.
It is built from five sub-scores, each 1–10:
| Criterion | What it measures |
|---|---|
| Originality | Does the review bring a fresh angle, synthesis, or framing beyond a summary of papers? |
| Methodological rigor | Is the review's approach clear — search strategy, inclusion criteria, or organizing framework? |
| Clarity | Is the writing precise, well-structured, and easy to follow? |
| Relevance | Does it address a question that matters to the field? |
| Bibliography | Are the references current, relevant, and adequate for the scope? |
Each score is accompanied by a confidence value (0–1) and a one-line summary, and the model used is stored with the result. Scores are advisory — they inform the moderator, they never replace them.
What the AI quality score is NOT
- Not a journal impact factor. It rates the quality of this single review; it does not predict real citations.
- Not a judgment of your worth as a scientist. It evaluates the text submitted, nothing else.
- Not a guarantee of citations. It is an AI prediction with a stated confidence, produced by a model that can make mistakes.
- Not the final decision. A human moderator reads the review and decides publication.
The DiDAL review protocol
The scientific review runs the DiDAL (Dialectical Dual-Agent Loop) protocol, adapted from the research pipeline developed for EcoSeek. It is a multi-stage AI review that does more than give a number:
- Classify — the reviewer sizes up the submission to choose the right depth of review.
- Retrieve — it searches scientific literature (OpenAlex, CrossRef) for the key works in the review's area, to check whether the bibliography engages with the field.
- Draft — it writes a draft evaluation against the five criteria, citing what it found.
- Critique — a second AI reviewer re-examines the draft, challenges its claims, and catches over-or under-rating. This dialectic reduces single-model bias.
- Revise & score — the two voices are reconciled into the final AI quality score, sub-scores, flags, and one-line summary.
The critique is stored with the screening record. Where no literature was retrievable (network limits, niche topic), the review proceeds with the text alone and notes the limitation. AI review is never the last word — the human moderator is.
So you don't get surprised
- No bibliography → low "Bibliography" score. It is nearly impossible to score high on a literature review without engaging the actual literature. If you don't cite, expect a low score.
- Generic text without method → low "Methodological rigor". Describe how you searched and synthesized; it is read as a method.
- Short, shallow, or filler-heavy → low overall. The AI reviewer is deliberately strict. A 5–6 is average; 8+ is exceptional.
- AI-assisted writing is fine. Using AI as a drafting and synthesis partner is the norm here — declare it openly, it never disqualifies.
- You can always improve and resubmit. A rejection email includes the score breakdown so you know what to fix.
Privacy and transparency
We store the screening record (scores, flags, one-line summary, the model used) alongside your submission. Your identity and contact details are visible only to the moderation team. Scores shown publicly on accepted reviews are labeled as AI-predicted and advisory. For full details see the About page.