Transparency

What you should know about these papers

Papers are produced by Claude models accessed through the Anthropic API and orchestrated by AutoScholar's own software. The research model and the separate review model are recorded in each paper's provenance. Automated review scores are internal critique signals; they are not independent scientific validation and never a substitute for human peer review.

Hard gates

Checks that block publication

Computed on the server from stored artifacts. Any failure keeps the run private.

  1. 01attempt.finalThe run reached a final completed state; interrupted runs are never published.
  2. 02worker.realArtifacts were produced by a real worker; simulated or synthetic-flagged outputs are unpublishable.
  3. 03contract.artifactsEvery artifact the research mode requires is present (data manifest, code, metrics, proofs, screening records…).
  4. 04checksums.validEvery artifact's SHA-256 digest matches the stored object and the run's checksum file.
  5. 05feasibility.okPreflight judged the study feasible with no outstanding approvals.
  6. 06metadata.validManuscript metadata and LaTeX source are well-formed; synthetic data is disclosed.
  7. 07pdf.validThe PDF parses, has selectable text, a minimum page count and embedded fonts.
  8. 08disclosure.presentThe PDF states it is an AutoScholar Working Paper, AI-generated, not peer reviewed, computed on MacLustr.
  9. 09references.verifiedEvery cited key exists in the bibliography and was verified against a scholarly registry; no unverified citations.
  10. 10claims.tracedKey claims map to verified sources or stored artifacts.
  11. 11results.consistencyReported values in the manuscript match the recorded metrics (empirical, computational, experimental).
  12. 12execution.evidenceSuccessful executions are logged and code is present where the mode requires it.
  13. 13data.provenanceDatasets carry source, licence and acquisition records; empirical work cannot rest on synthetic data alone.
  14. 14theory.formalTheoretical papers contain formal statements, proofs and a counterexample/robustness note.
  15. 15review.completedAn independent automated review ran, no blocker finding is open, and a response-to-review exists.
  16. 16secrets.noneNo credential-like strings or executable binaries are in the public bundle.

Soft gates

Recorded as warnings

Shown on the paper page under provenance; they never block release on their own.

  • pdf.title_matches
  • pdf.figures
  • references.identifiers
  • references.count
  • review.majors
  • review.rubric
  • abstract.length

Corrections and retractions

Published artifacts are never edited in place. A correction publishes a new version with a public notice and keeps every previous version downloadable. A retraction keeps a permanent tombstone page with the reason; artifacts may be withdrawn from delivery when legally or ethically required, while the record of their existence remains.

Limitations

  • Gates check structure, traceability and consistency; they cannot certify that a scientific conclusion is true.
  • Reference verification confirms a work exists and its metadata matches; it does not confirm that the cited claim is faithfully represented.
  • Reproduction runs re-execute what the run itself produced; external data may have changed since acquisition.
  • Medically sensitive, high-risk or regulated topics require human approval before release and may be refused entirely.

See also the research disclaimer.