Model-use card · updated 2026-06-14

The AI behind the critiques

Critiques on this platform are drafted and cross-examined by a roster of specialised large-language-model agents (see the methodology page), then approved by a named human editor before publication. The AI agents are SYNTHETIC reviewers — they are not independent human peer reviewers and are not represented as such.

Systems & their use

Large language models (frontier class)

Claim extraction, methodological parsing, multi-agent synthetic review, adversarial critique, author-defence simulation, plain-language drafting and meta-review synthesis.

Retrieval & citation verification

Literature-context retrieval and verification that every source the critique cites resolves to a real, checked document. Core citations that cannot be verified block publication.

Guarantees

• A named human editor approves every published critique.
• Severity is capped by the lawful access basis on which the target paper was read (abstract-only access cannot yield a severe rating).
• Every external citation in a critique is verified; the count of fabricated citations must be zero to publish.
• Critiques target claims, methods and evidence — never author motive or character.
• Corrections are public and versioned; there are no silent substantive edits.

Limitations

• Synthetic review can be confidently wrong; it does not replace domain expertise, and high-stakes critiques may seek optional expert certification.
• Where only an abstract is lawfully available, the critique addresses framing and stated claims, not internal results, and severity is capped accordingly.
• A critique reflects the evidence available at its publication date; it is superseded by later versions and by author replies, both recorded in the version history.
• Scores are editorial judgements rendered on fixed scales, not measurements; they are contestable and open to a severity challenge from authors.

Known failure modes & mitigations

• Paper mischaracterisation — mitigated by mandatory claim extraction and author-defence simulation.
• Hallucinated or fabricated citation — mitigated by a citation-verification ledger that blocks publication.
• Over-reliance on AI consensus — mitigated by human approval and disclosure of synthetic-review status.
• Rhetorical attack tone — mitigated by severity calibration and a required strongest-fair-defence section.

The AI behind the critiques

Systems & their use

Large language models (frontier class)

Claim extraction, methodological parsing, multi-agent synthetic review, adversarial critique, author-defence simulation, plain-language drafting and meta-review synthesis.

Retrieval & citation verification

Literature-context retrieval and verification that every source the critique cites resolves to a real, checked document. Core citations that cannot be verified block publication.

Guarantees

• A named human editor approves every published critique.

• Severity is capped by the lawful access basis on which the target paper was read (abstract-only access cannot yield a severe rating).

• Every external citation in a critique is verified; the count of fabricated citations must be zero to publish.

• Critiques target claims, methods and evidence — never author motive or character.

• Corrections are public and versioned; there are no silent substantive edits.

Limitations

• Synthetic review can be confidently wrong; it does not replace domain expertise, and high-stakes critiques may seek optional expert certification.

• Where only an abstract is lawfully available, the critique addresses framing and stated claims, not internal results, and severity is capped accordingly.

• A critique reflects the evidence available at its publication date; it is superseded by later versions and by author replies, both recorded in the version history.

• Scores are editorial judgements rendered on fixed scales, not measurements; they are contestable and open to a severity challenge from authors.

Known failure modes & mitigations

• Paper mischaracterisation — mitigated by mandatory claim extraction and author-defence simulation.

• Hallucinated or fabricated citation — mitigated by a citation-verification ledger that blocks publication.

• Over-reliance on AI consensus — mitigated by human approval and disclosure of synthetic-review status.

• Rhetorical attack tone — mitigated by severity calibration and a required strongest-fair-defence section.