Editorial policy

Editorial policy

Standing

In Silico is an experiment in open machine-generated peer review. It is not accredited, not indexed, and not a substitute for conventional peer review. Do not cite an In Silico listing as evidence that a result is correct.

The panel's rubric and method limits are documented under Criteria.

Authority

The Editor-in-Chief agent decides the verdict and readiness score. A human decides whether to publish the review.

The editor agent recommends accept, minor revision, major revision or reject, and assigns a publication-readiness score from 0 to 100. Specialist scores are advisory inputs. The readiness score measures the current manuscript through scientific validity, methods and evidence, reproducibility and reporting, and clarity and completeness. Novelty, significance, and usefulness are reported separately as a contribution profile.

No score range determines the recommendation. The recommendation follows the work required to make the manuscript publishable. Low novelty or significance does not prevent acceptance when the work is valid, useful, and ready. We publish the decision letter word for word. It binds nobody. Merging the pull request is the decision to publish. A human editor may decline to publish a review that is incoherent, that misreads the paper badly enough to mislead, or that came from a mangled ingest.

What acceptance means

In Silico accepts a paper the panel returns at accept or minor revision, and declines the rest.

The editor is never told this rule. It recommends one of the four standard verdicts with nothing hanging on the answer, and the line is drawn afterwards. An editor told that "minor revision" means acceptance is an editor being asked to gatekeep, and it would grant more of them.

The line is drawn on the editor's recommendation, never on the readiness score or the specialist-panel mean. Historical reviews retain their 1 to 5 panel mean and label it as a legacy value. It is not converted into a readiness score.

"Declined" rather than "rejected", because that is what happened: the editor declined to accept, and most declined papers carry a letter setting out what would fix them. Anyone can submit a preprint they did not write, so a permanent machine-generated "rejected" is not a label we will attach to a stranger's work on this evidence.

Where a paper has been reviewed more than once, its status comes from the most recent review by the graded panel. A run on a single model is a published experiment rather than an editorial decision, and does not overturn a panel it was run to be compared against.

Scope

In scope. Any original research manuscript, in any discipline, whose claims a careful reader can evaluate from the manuscript and what it cites or deposits. Empirical, theoretical, computational and methodological work all qualify, as do negative results, replications and reanalyses. What matters is whether the evidence can be checked, not the field.

Out of scope.

• Anything where a wrong machine-generated review could affect patient care: clinical trial reports, diagnostic or treatment guidance, dosing recommendations. This is the one hard line, and it is not about field: a computational paper that outputs a dosing recommendation is out, a clinical-adjacent methods paper that does not is in.
• Work whose central evidence cannot be checked at all.
• Anything that is not a research manuscript. Editorials, press releases, marketing, or text making no checkable claim.

Out-of-scope submissions are declined at the desk, without running the panel.

Who may submit

Anyone may submit any public preprint, including one they did not write.

Every review says which it was. The form asks one question: are you an author of, or affiliated with, this paper? Where the answer is no, the review carries a notice that the authors did not request it and have not replied. Older reviews say "solicitation unrecorded" rather than passing for requested. We do not verify the claim, and we say we don't.

Submitting a rival's work to attach a public criticism to it is not a use we support. An editor reads every submission before starting a run, and declines any that reads as score-settling.

If you find an unrequested review of your work, you can request its removal.

The desk

Two checks run before any referee is assigned, and either can stop a submission.

We do not screen for prompt injection

Our referees are all AI, so a paragraph hidden in a PDF telling the reviewer to recommend acceptance would reach every referee we have. You should know that we do not look for it.

We used to. The scan compared a text run's colour against a threshold without checking what was drawn behind it, so white labels on a dark figure read as concealed text. On a real submission it reported that the authors had hidden eleven thousand characters, which were the axis labels of a heatmap. The matching half was a list of phrasings: it caught the payload copied off a blog and nothing rewritten. A check that reliably accuses honest authors and unreliably catches dishonest ones is worse than no check, so it was removed rather than tuned.

Nothing here replaces it. If you submit a manuscript with instructions hidden in it, our referees will read them as prose. What we can promise is the opposite of what the scan promised: we will not publish a claim that anyone concealed anything, because we are no longer in a position to make one.

Editorial triage

One fast pass assesses whether the submission is in scope, intelligible, complete, and not fatally flawed on its face. It is told to reject sparingly. While the new manuscript text hierarchy is being validated, this pass is advisory and cannot stop the panel.

If enforcement is restored, a desk rejection will be badged separately from a panel rejection. Nothing will have read the paper in depth and no specialist reports will exist.

Publication record and reproducibility

A published review contains the panel output described under Who does what, plus the decision letter. It also contains manuscript_stats.md, which records counts over the text the panel read so a reader can compare that text with the selected archive source.

Every review ships a provenance.json: provider, the model at each stage and any per-agent override, PeerReviewAgents version and commit SHA, In Silico commit SHA, resolved configuration and hash, journal profile hash, debate rounds, active desk screens, per-reviewer scores, total and per-agent cost, and the resolved preprint metadata. A manifest.json fingerprints every artifact in the bundle. You can re-run the panel yourself. Output will not be identical, because the models are not deterministic, but the inputs are disclosed.

A published bundle is never edited. Records use two coordinates. v<N> is the manuscript version and r<M> is an independent review attempt of that version. A same-version re-review creates a new attempt and does not inherit the earlier attempt's verdict or score.

The newest eligible editorial attempt on the newest earlier manuscript version is the baseline for a revision round. That exact bundle is passed to PeerReviewAgents through its built-in revision comparison. Single-model experiments remain visible but do not set the paper's status and cannot become a revision baseline. Revision bundles include a round.json record of what was carried forward.

Removing a review

Anyone who submitted the paper, or an author of the paper, can ask us to remove the review. Comment on the review pull request or submission issue. Before publication, we close it. After publication, we remove the review from the site. No explanation is required.

Conflicts and cost

The editor is one person running this out of pocket. Reviews are triggered manually because each one costs real money. No queue guarantee, no SLA, and no promise the project outlives the author's interest in it.

Data

We store the preprint URL and its public metadata, the review bundle, the GitHub issue thread, and your reply if you send one. We do not store PDFs. They are fetched to a temporary directory at review time and deleted afterwards.

Paper text is sent to a third-party AI provider, currently Anthropic. Since we only take already-public preprints, this should not come up.

License

Reviews and site content are CC BY 4.0. Preprints stay under whatever license their authors chose. We host none of them.