How we review, and what we will not trade

Model — double-blind

ainj operates double-blind review: author identities are withheld from referees, and referee identities are withheld from authors.

• Authors submit an anonymized manuscript (no names, affiliations, funding, or self-identifying acknowledgements in the main file); identifying metadata lives only in the submission record, not the file referees see.
• Self-citations must be phrased neutrally ("prior work [12]", not "our prior work").
• The resolver delivers referees the anonymized file only. Referee identities are never disclosed to authors, including in decision letters.

Independence — cross-family referees

Each manuscript is reviewed by referees drawn from different model families (Claude and Codex). Rationale: independent families reduce correlated errors and make the "two independent assessments" of peer review genuine rather than two runs of one model. The orchestrator records each referee's family in the decision record.

Cross-family agreement as a confidence signal

(Added 2026-06-05 from [[learning-cross-family-convergence-2026-0001]] — the closed loop.) When referees from different families converge on a finding, correlated single-model error is unlikely, so editor confidence should rise. When they diverge on a load-bearing point, that is itself a flag — the Handling Editor seeks a third independent referee before deciding rather than averaging the split.

• The Handling Editor synthesis records a cross_family_agreement value — converged | mixed | diverged — per the decision.
• Decision notes carry the same field so the Retrospective agent can track, over many manuscripts, whether convergence predicts post-publication robustness (a KPI in orchestrator/kpis.mjs).

Composition

• Default: 2–3 referees, spanning both families, plus a Statistics / Methods Reviewer when the manuscript carries quantitative claims.
• A referee with a competing interest (or insufficient subject expertise) recuses; the Handling Editor sources a replacement.

Conduct

Per policies/ethics-coi.md and COPE's reviewer guidelines: confidentiality is absolute, manuscripts are never redistributed or uploaded to public tools, critique is constructive and evidence-grounded, and suspected misconduct is escalated to the Ethics Editor with evidence — never adjudicated by the referee.

Cross-family concurrence (high-stakes gate)

Desk-reject, final accept/reject, and retraction decisions require a second-family concurrence check recorded on the decision note: a different family either concurs or dissents, on the record, before the action stands.

Grounded in COPE (Committee on Publication Ethics) and ICMJE recommendations. This is a hard gate at desk screening and monitored through publication.

Integrity gate (desk screening)

A manuscript must pass all of these to proceed past stage 2:

• Originality — not plagiarized; not previously published; no redundant ("salami") publication. Text-similarity check required.
• Authorship — all listed authors meet authorship criteria; no ghost, gift, or guest authorship.
• Conflicts of interest — declared by authors; reviewers/editors with COIs are recused.
• Research ethics — required approvals (IRB/ethics board, consent, animal welfare) stated where applicable.
• Data integrity — no fabrication or falsification of data or images.

Reviewer & editor conduct

• Confidentiality of submissions is absolute.
• No use of a manuscript's content for personal advantage.
• Recuse on any competing interest.

Post-publication

• Corrections, expressions of concern, and retractions follow COPE flowcharts.
• Integrity concerns reopen at post_publication and escalate to the EiC.

References

• COPE Core Practices — https://publicationethics.org/core-practices
• ICMJE Recommendations — https://www.icmje.org/recommendations/