How the Andreev corpus was constructed

The 27 September cleanup records the current publication-format policy, reviewed duplicate exclusions, and resolved deferrals. The 18 September quality audit records the earlier evidence gaps and follow-up inventories.

This report describes the source material, selection rules, collection procedure, and limits of the Andreev literature corpus. It distinguishes the frozen 2016-09-09–2026-09-09 snapshot from the separately recorded backfill to 2000. The ten-year coverage report gives its counts; the search ledger records individual requests. The completed 2000 backfill report documents the extension: 6,647 accepted records in the expanded date window, its separate 3,618-record backfill view, additional queries, screening corrections, unresolved metadata cases, and validation. Counts below describe the original snapshot unless explicitly identified as the extension.

Latest extension: 1985–1999

The 1985 extension report records a separate 1985-01-01–1999-12-31 discovery and screening pass. The main view now starts in 1985. The frozen ten-year and 2000 registers remain unchanged. This extension adds 1,971 accepted records and preserves 813 new exclusions and 287 new deferrals. Its OpenAlex searches ended at their recorded continuations; arXiv requests failed and are not described as complete. The linked report contains query-level native-ID coverage, audit corrections, date overlap, exports, and exact validation results.

Research scope and unit of collection

The corpus concerns Andreev reflection and bound states, including their use as mechanisms, probes, and device resources. It includes sufficiently related work on multiple and crossed Andreev reflection, Josephson subgap spectra, proximity effects, Yu–Shiba–Rusinov states, Cooper-pair splitting, quasiparticle parity, microwave spectroscopy, and superconducting hybrid devices. Majorana literature qualifies when its transport, detection, or control problem has a substantive connection to ordinary Andreev states or the accepted device platforms. Physical analogues in superfluids and bosonic systems can also qualify.

A passing mention, a reference to an author named Andreev, or general work on topology or superconductivity is insufficient. Andreev–Bashkin entrainment and Andreev–Lifshitz supersolidity are separate eponyms and need an additional reflection or bound-state connection. Data deposits, software releases, figures, and supplementary files are excluded as independent literature records. Research articles describing relevant computational methods remain eligible. Corrections and comments are retained as distinct works when relevant. Since 27 September, conference abstracts, presentations, proceedings papers and proceedings volumes are excluded, including preprints of those same works. Independently identified journal publications and standalone manuscripts remain eligible. Conflicting version evidence is deferred. This supersedes earlier conference inclusions without erasing their review history.

The counted unit is a local canonical bibliographic record. A canonical record can contain several provider manifestations, such as an arXiv preprint and an OpenAlex journal record. The word “canonical” describes a database identity, not a guarantee that all versions have been resolved correctly. Unresolved preprint/article pairs can therefore contribute two records. Citation counts, topical counts, and publication-type counts must be interpreted on that basis.

Sources and search design

arXiv supplies preprint metadata, abstracts, submission dates, and identifiers. OpenAlex supplies a complementary publication index, journal metadata, abstracts where available, authorship information, identifiers, and reference lists. Neither source is a complete census of the literature. No explicit language filter was imposed; screening depends on the metadata available for reading.

The initial collection began with small, relevance-oriented seed searches. Those were followed by seven larger query families with recorded pagination. The core arXiv expression was:

(ti:Andreev OR abs:Andreev)
AND submittedDate:[201609090000 TO 202609090000]

The complementary arXiv expression deliberately omitted records matching the core term, to discover related literature under different terminology:

((ti:"Yu-Shiba-Rusinov" OR abs:"Yu-Shiba-Rusinov" OR ti:Shiba
  OR ti:"Cooper pair splitting" OR ti:"Cooper-pair splitting" OR ti:gatemon
  OR (ti:Josephson AND
      (abs:subgap OR abs:"bound states" OR abs:parity OR abs:"current-phase"))
  OR (ti:"proximity effect" AND
      (abs:superconducting OR abs:superconductor)))
 ANDNOT (ti:Andreev OR abs:Andreev))
AND submittedDate:[201609090000 TO 202609090000]

OpenAlex used these native filters, each combined with from_publication_date:2016-09-09,to_publication_date:2026-09-09:

The saved JSON files are authoritative for exact spelling, punctuation, offsets, and provider parameters. See the original request directory. Each local batch requested at most 200 records. An arXiv batch could contain multiple underlying 100-record requests, which are retained in the returned pagination metadata. OpenAlex batches used 200-record pages.

The completed families returned 1,425 and 641 arXiv records and 1,552, 4,178, 184, 42, and 34 OpenAlex records, respectively. These counts overlap: adding them does not give a count of distinct candidates. Exhaustion means the provider reported no continuation for the saved query. It does not establish recall for different search terms, unavailable abstracts, or unindexed literature.

Dates and snapshot boundaries

The first probe used 2016-01-01 as its lower bound. The user subsequently specified a rolling ten-year corpus, so the completed snapshot used 2016-09-09–2026-09-09. Earlier results were preserved and screened against the clarified rule, rather than silently discarded or all relabeled as foundations.

For the ten-year register, a record qualifies if at least one retained source manifestation has a date within the window. Its statistical year is the earliest qualifying source date. An earlier arXiv preprint can therefore qualify through a later journal publication. This date can differ from the canonical publication date and the year in a BibTeX export. Both source dates and the qualifying date are recorded, so this choice is inspectable.

The native arXiv upper timestamp was 202609090000; it is not an end-of-day timestamp. The requests were executed during the recorded September 8 UTC session. Consequently, the September 9 snapshot label and intended date window do not imply that submissions indexed later on September 9 were retrieved. Provider indexing lag is an additional limit. The run timestamps in the ledger describe what was actually available at collection time.

The requested backfill uses 2000-01-01–2016-09-08: arXiv timestamps 200001010000 through 201609082359, and the corresponding inclusive OpenAlex publication-date filters. This fills the earlier interval without redefining the original snapshot. Work before 2000 remains eligible only as explicitly selected foundational literature.

From a request to a review decision

1. Save a provider-native JSON request under searches/ before executing it. Use this project's Pixi environment and PatchScholar's fetch raw operation.
2. Retain the original provider payload, fetch summary, and pagination metadata under ignored data/ paths. Import provider manifestations into the local database using PatchScholar's identifier and identity rules.
3. Record the retrieved count, new canonical candidates, timestamps, continuation, and status in the ledger. Candidate counts are distinct from provider record counts and from corpus membership. A failed request remains recorded and is not counted as a completed page.
4. Attach candidates through supported review actions. Merely ingesting a provider result does not make it an accepted corpus member. Preserve existing decisions when the same work appears in another search.
5. Read the title and available abstract. Record include, exclude, or defer, topical labels, a paper-specific reason, and an evidence pointer. Inspect original source or full text when needed and accessible; otherwise retain the uncertainty explicitly.
6. Consolidate the reviewed batches, audit questionable decisions and identities, and apply the final decisions through a recorded PatchScholar pack. Export and validate the resulting register and bibliographies.

Representative commands are:

pixi run scholar fetch raw arxiv search --spec-file searches/REQUEST.json
pixi run scholar fetch raw openalex works --spec-file searches/REQUEST.json
pixi run scholar run-pack packs/REVIEW.yaml
pixi run scholar corpus export andreev-literature

These placeholders explain the operations, not an instruction to replay old review packs over later decisions. Archived packs contain local canonical IDs and document historical actions. The backfill's bounded run recipe is scripts/backfill_2000.py; its provider operations remain in the sibling PatchScholar package.

Screening and quality checks

Three Luna agents assisted with the original screening. Their assessments were followed by root review of exclusions, publication types, title-only judgments, ambiguous scope, and identity evidence. The audit corrected false surname matches, unrelated geometry and humanities, data deposits projected as articles, and confusion between software artifacts and research methods articles. Earlier batch files are retained even where their judgments were superseded.

The consolidated register—not an arbitrary batch file—records the snapshot's final decisions. Every assessment is labeled as agent work and is not human-confirmed. This was not a blinded multi-reviewer systematic review; neither inter-reviewer agreement nor a statistical estimate of screening error was measured. Agreement among automated passes would not establish correctness.

The ten-year snapshot contains 254 included records explicitly labeled title-screened. Thirteen scope cases remain deferred because the available text did not establish substantive relevance. The accepted set is therefore not a set of fully read papers. Only five original full-text downloads were retained from the initial batch: four arXiv source archives and one PDF returned by the source endpoint. Their log is reviews/2026-09-09-fulltext.csv. Unpacked derivatives remain under ignored data/extracted/; downloaded source was not compiled or executed.

Identity resolution and source fidelity

Identifiers, rather than title equality, support automatic source association. Twelve same-work consolidations were explicitly reviewed using identifier evidence. Corrections were distinguished from their originals. Three companion data sources had contaminated paper projections; supported source splits restored the paper metadata and retained the data artifacts as excluded records.

Provider identifiers are not infallible: a dataset can expose the identifier of its associated article. Conflicting identity claims were quarantined for review, and a reviewed bridge was retried. Original source records and review history were retained through these repairs. A title-based follow-up list contains 155 unresolved groups covering 320 included records; these groups were not merged merely because their normalized titles agree.

The detailed repair summary and source-split evidence are linked from the snapshot report. The local PatchScholar dependency is editable and was not a clean, frozen release; its revision and dirty-worktree status are recorded in DEPENDENCIES.md and the validation file.

Citation expansion and the network

A bounded citation audit expanded selected spin-qubit, YSR, superfluid-reflection, and Andreev/Majorana-review seeds. It fetched 172 provider records: 99 references and 73 citing works, creating 137 additional canonical candidates. All were screened. A seed without an OpenAlex manifestation exposed a graph-coverage gap. The audit did not recursively retrieve all references and citations.

The browser network is broader than those 172 fetched records. OpenAlex results also carry reference lists; PatchScholar resolves references whose target identifiers already occur in local state. A resolved edge means that the source record cites the target record. It does not mean that an agent read both papers or judged their scientific relationship. The explorer omits identity self-links and displays edges only between records in the active selection.

The initial explorer snapshot had 15,184 non-self directed edges across all 4,614 screened records. That number includes links involving exclusions and foundations; an included-only or ten-year view has fewer. The bounded display defaults to 250 connected records and reports its omissions; the full graph is available. Missing links and isolated nodes do not establish absence of a citation or a scientific connection. Layout distance is a visualization aid, not a measured similarity score.

Coauthorship view

The Collaboration tab projects PatchScholar's canonical authorships onto the records selected by the corpus filters. Each node is a canonical author ID; each undirected edge counts distinct selected records naming both authors. A record contributes once per author pair regardless of repeated source assertions. Reviewed duplicate records are omitted. Single-author records remain as isolated nodes. No additional name matching or identity merging is performed by the viewer.

The overview displays the 250 or 750 authors with the most selected records. Author search covers the full selected author set; focusing an author ranks their collaborators by shared records. The displayed subset and missing author coverage are reported explicitly. Shared-paper lists and CSV exports identify the underlying record IDs. Layout coordinates are for display only; node size reflects selected records and line thickness reflects shared records.

Author identities may be split or conflated upstream, and unresolved work duplicates can inflate collaboration counts. Records with no canonical authorships cannot contribute nodes or links. This view does not claim a complete collaboration census or resolve the identity issues described above. scripts/build_collaboration.py uses PatchScholar's public authorship inspection API and writes the separate, register-hash-bound collaboration.json artifact. The regular explorer build refreshes it without changing reviews or identities. The 27 September author cleanup records the automatic candidate queue, Luna judgments, 54 explicit source-author assignments and verified replay. Only the reviewed identities were corrected; deferred cases and the unreviewed queue remain separate.

Classification used by the explorer

The physics tags originate in the screening register, with explicit display aliases for synonymous labels. The later classification view adds platform and research-approach facets. It keeps existing review tags separate from unreviewed title/abstract suggestions generated by the versioned rules in scripts/classify_explorer.py.

For a text suggestion, the public record retains the matched field. Supporting title excerpts are visible; abstract excerpts remain local. A match can describe background rather than the paper's main result; users can restrict the view to review tags. Categories can overlap, and an unclassified record is not a negative scientific judgment. Theory/experiment labels from unreliable bulk screening were removed in the original audit; the display rules do not retroactively make those judgments reliable. Classification annotations never change inclusion decisions or identity links.

Validation, preservation, and reproducibility

The original validation matched all 4,614 consolidated decisions to database state, preserved all 4,611 earlier review events, checked SQLite integrity and foreign keys, and parsed the exported bibliographies with Pybtex. The exports contained 3,076 ten-year entries, 59 foundations, and 3,135 combined entries, without duplicate citation keys. All 20 known-paper benchmarks were included. Those benchmarks are a finite recall check, not a measure of total recall.

Git tracks scope, request specifications, the ledger, review registers, reports, and validation summaries. Raw downloads, the SQLite database and its histories, extracted full text, and generated exports are ignored and need separate backup. A fresh checkout can inspect the tracked judgments but cannot reconstruct every provider manifestation or citation edge without that local state or a new fetch.

Hashes and original checks are in reviews/completion/validation.json. The backfill records its own baseline backup, requests, decisions, and validation separately. It must preserve the original register and review history even when the expanded date scope justifies a new decision.

Public projection and ongoing identity review

The publication policy keeps stored abstracts and quoted abstract evidence local; the browser retrieves abstracts from sources on demand. Published review copies can therefore differ from the originals whose hashes appear in historical validation reports.

The systematic author workflow replaces the initial bounded pilot with complete candidate accounting, evidence-bound Luna review and explicit pending, deferred and stale states. Reported examples are regression checks, not special-case cleanup rules.