Ten-year corpus snapshot, 2026-09-09

For the full construction procedure, exact search expressions, screening and identity rules, date semantics, network derivation, classification evidence, and reproducibility limits, see How the corpus was constructed.

The collection covers 2016-09-09 through 2026-09-09, with earlier foundations kept separately. It contains 3,076 included records in the ten-year window, 59 earlier foundations, 1,466 exclusions, and 13 deferred scope decisions. All 4,614 local canonical records have a recorded screening decision.

These are agent assessments, not human confirmation. The deliverable is a screened bibliographic corpus with source provenance and selected full text. It is not a collection of 3,135 fully read or downloaded papers. Counts refer to records rather than a verified number of distinct intellectual works: 155 shared-title groups, containing 320 included records, remain flagged for relation review. Some are preprint/article pairs; others may be distinct contributions.

Completed discovery

The following provider requests were followed to their reported pagination ends. Counts overlap across searches and must not be added as unique papers.

The complementary arXiv query covers YSR/Shiba states, Cooper-pair splitting, gatemons, Josephson work mentioning subgap states, bound states, parity or current-phase relations, and superconducting proximity effects. Exact Boolean expressions, date constraints, page offsets and continuations are preserved in the request specs and search ledger.

An initial arXiv batch started at 2016-01-01, before the rolling-window clarification. It was superseded by the completed ten-year search. The final register applies the rolling window to source dates; it does not count the entire early-2016 batch as foundations. A record qualifies for the decade if at least one retained source manifestation has a date in the window. The year table uses the earliest qualifying source date, which can differ from the journal year in BibTeX. arXiv dates are initial submission dates; OpenAlex dates are publication dates.

Related literature and screening

The scope includes Andreev reflection and bound states, MAR, crossed reflection, spectroscopy and qubits, as well as substantive connections through YSR states, Cooper-pair splitting, Josephson transport, proximity devices, Majorana diagnostics, and physical analogues. Topical labels overlap and are incomplete; they are aids to navigation rather than independently validated subcorpora. For example, the decade register has 791 andreev-reflection, 492 andreev-bound-states, 163 yu-shiba-rusinov, 146 crossed-andreev-reflection, and 116 multiple-andreev-reflection labels.

Three Luna agents assisted with screening. Follow-up audits checked exclusions, publication types, title-only cases, scope, and identity evidence. Root review corrected false surname matches, unrelated geometry and humanities, data deposits mislabeled as articles, and confusion between software releases and research articles describing relevant methods. Andreev–Bashkin drag and Andreev–Lifshitz supersolidity alone do not qualify as reflection or bound-state literature.

Bulk theory/experiment labels were removed where their reliability had not been established. The final register retains reasons, evidence, date provenance and the originating review file. Earlier review files remain as history, including superseded judgments. Use the consolidated register or current database for the final decision, not an arbitrary individual batch file.

There are 254 included decade records explicitly marked title-screened. Thirteen broader physics records remain deferred because the available text did not establish their substantive connection. They are preserved with reasons and source links, and are excluded from the accepted bibliographies. Primary-source checks resolved several other cases; where only a syndicated abstract could be read, the evidence field says so.

Recall and citation checks

All 20 entries in the known-paper benchmark are present and included. They span qubits, spectroscopy, YSR states, pair splitting, Majorana diagnostics, quantum Hall/graphene interfaces, helium, unconventional superconductivity and hybrid devices. This finite benchmark is a recall check, not an estimate of total literature recall.

Bounded reference and citation expansions around spin-qubit, YSR, superfluid reflection and Andreev/Majorana-review seeds retrieved 172 provider records (99 outgoing-reference records and 73 incoming-citation records), creating 137 additional canonical candidates. The first expansion attempted a QAHI seed without an OpenAlex manifestation; its missing graph coverage is recorded. Subsequent expansions separately sampled the other seed neighborhoods to avoid spending the whole budget on the first seed. Every new candidate was screened.

Citation expansion stops at the recorded audit limits; many reference/citation neighbors remain unfetched. The protocol does not recursively include all background citations. Graph specs preserve exact reference IDs and reconstruct the native initial citation request from the supported client's parameters; continuations and acquisition states are retained in the ledger and local results.

Identity, source fidelity and software repairs

Twelve verified same-work consolidations used explicit identifier evidence. An earlier reviewed duplicate exclusion is also retained. An author correction remains distinct from its original article. Shared titles alone were not used to merge records; the outstanding groups are listed in the relation follow-up register.

Three companion-data source records had been attached to papers and corrupted their projected titles or DOIs. Supported explicit source splits restored the paper metadata and kept the data deposits as excluded records. Existing review history was checked against SQLite backups and remained intact. Provider-native identifiers can still express paper/data relations incorrectly; the explicit assignments and original evidence are preserved instead of deleting that history.

PatchScholar gaps were communicated through agentdispatch to the thread specified by the user. The live package supplied pagination continuations, identifier normalization/repair, conflict reporting and quarantine/retry, and creation of a new identity during a source split. The first repair restored 115 identifier projections. A later repair removed the original paper's arXiv ID from its correction. One quarantined identifier bridge was reviewed and retried successfully. Publication-type handling in BibTeX was also reported and fixed before export; books, chapters, conference papers and unclassified records no longer all become journal articles.

Raw provider downloads remain intact under ignored data/raw/. Five selected original full-text downloads from the first batch are cached: four arXiv source archives and one PDF returned by the source endpoint. Text derivatives are under ignored data/extracted/; downloaded source was not executed or compiled. See the full-text log.

Deliverables and limits

• Consolidated review register: every decision, identifier, source date, native publication type, reason and evidence pointer.
• Ten-year BibTeX, foundations BibTeX, and combined BibTeX: local, ignored exports.
• Search plan, exact request specs, ledger and batch/audit registers: tracked research provenance.
• state/patchscholar.db: current corpus and full review/identity history; data/batches/2026-09-09-completion/: raw results, reviewed action pack, verification results and before-change backups. These require a separate backup and are not included in Git.

The recorded discovery and screening protocol is complete for this snapshot. Provider coverage, indexing delays, inaccessible abstracts, incomplete citation neighborhoods and unresolved version relations remain limits. This is a broad ten-year corpus, not proof that every relevant work has been found.

Validation checked the register against all current corpus decisions, retained all earlier review events, and confirmed SQLite integrity and foreign keys. Pybtex parsed 3,076 decade entries, 59 foundation entries and 3,135 combined entries, with no duplicate citation keys. The 20 benchmark records are included. Checks and artifact hashes are recorded in the validation record.