Reviewed author identity cleanup, 27 September 2026

The collaboration graph contained multiple source identities for the same researcher. This batch uses automatic candidate detection followed by Luna validation. Detection nominates candidates; only explicit, high-confidence gpt-6-luna merge groups become identity corrections in PatchScholar.

Scope and outcomes

The frozen scan examined 22,721 author identities attached to non-duplicate corpus records. It nominated 17,389 pairs in 2,070 candidate groups. The first bounded review selected 36 groups (at most eight identities per group), prioritizing the reported Akhmerov example and well-connected authors in accepted records.

Three gpt-6-luna reviewers assessed 115 nominated pairs: 73 merge, 41 defer, and one distinct. Their explicit group decisions produced 33 merge groups and 54 source-author reassignments. Pair counts and group counts describe different objects: several pair decisions can support one multi-identity group.

The five Akhmerov IDs (69, 1856, 2388, 6276, 6303) resolve to canonical author 69. Luna checked the primary arXiv paper 2604.22650 and its link to the supporting dataset, alongside source-reported affiliation, authorship and collaboration evidence. The initial, more conservative judgments and the revised evidence-backed decisions are both preserved.

The remaining 2,034 candidate groups have not been reviewed. Deferred pairs remain separate, including conflicting source-reported ORCIDs and insufficiently supported fragments. This is a bounded cleanup, not complete author disambiguation.

Evidence and safeguards

• Name normalization, initial compatibility, existing canonical coauthors, shared records and ORCID evidence nominate pairs. Priority scores are queue ordering, not identity probabilities. Names alone do not authorize a merge.
• Dossiers show source identities, affiliations, publication histories and coauthor evidence. Long publication histories are explicitly sampled.
• Luna records merge, distinct or defer judgments with confidence, reasons and work IDs. A review may use the whole group's evidence, but candidate chains do not themselves establish identity. Proposed groups must agree with their nominated internal pair decisions.
• Application rejects conflicting nonempty ORCIDs, distinct native author IDs co-occurring on one source work, inconsistent judgments, missing evidence and source mappings that changed since review. These checks enforce the review; they do not replace Luna's identity judgments.
• Corrections use PatchScholar's atomic assign_identities operation. Each source assignment has an audit event, previous ID and reason, and can be reversed by a further explicit assignment. Source records and historical canonical IDs remain intact. No work identities or include/exclude/defer decisions change.

Replay and package fix

The first isolated replay exposed a projection bug: name-only assertions could still attach to departed identities after their source authors were reassigned. At the user's request, the fix was implemented and validated in a dedicated PatchScholar thread (2e5aecff-e04b-430d-8af3-88d0277c5293).

The completed fix keeps exact source-name variants attached to explicitly reviewed canonical groups. Empty departed identities remain in history but do not compete for name-only assertions. Live homonyms remain ambiguous and unreviewed identities retain their prior matching policy. There is no fuzzy name resolution in the graph.

A fresh replay and live application check all expected mapped authorships, additional exact-alias evidence, source assignments, audit-event counts, SQLite integrity and foreign keys. Hashes of all 17 non-author tables must match before and after. The register remains 5a42bb33b350da13b0bbf9324c2e1d2a8d098055de3193abb78cd95ead4ce5ad.

Reproduction record

• Detection limits and input hashes
• Complete candidate queue
• Frozen review batch
• Luna decisions: batch 1, batch 2, batch 3
• Validated application plan
• Fresh replay validation
• Live validation

The application reports record the PatchScholar base commit and hashes of its uncommitted files, along with the frozen review and application-script hashes. The corpus runner is pixi run author-cleanup-pilot prepare|preview|apply; it refuses to overwrite existing replay evidence or repeat an application. Historical inputs are preserved under data/author-cleanup-2026-09-27/, including the failed replay and before-application backups. Do not rerun the detector over the frozen review files; create a new batch for subsequent cleanup.

This pilot is superseded for new work by the systematic identity workflow. Its source and frozen evidence remain unchanged; the current author-cleanup task uses the new workflow.