Andreev corpus: backfill to 2000

This extension covers 2000-01-01 through 2016-09-08, separately from the frozen ten-year collection. It adds 3,468 accepted canonical records and restores 55 previously excluded records after reconsidering the earlier date boundary. The complete backfill view contains 3,618 accepted records; the expanded 2000–2026 view contains 6,647, plus 11 older foundations outside that window.

The construction report explains the overall research scope, original discovery, identity handling, and explorer derivation. This report records what changed during the backfill, including the screening problems found during validation. All relevance judgments remain agent assessments, not human confirmations.

Collections and date accounting

The two period views overlap by 47 canonical records: a preprint can qualify for the backfill while its journal manifestation qualifies for the ten-year snapshot. Adding the two period counts therefore double-counts these records. The expanded view counts each canonical ID once. Its counts are bibliographic records, not a fully deduplicated count of distinct intellectual works.

The backfill includes 95 records that were already accepted: 48 earlier foundations and the 47 overlapping records. Thus 3,468 new inclusions + 55 restored inclusions + 95 existing inclusions = 3,618. The original foundation view still has 59 records, of which 11 precede 2000.

The complete reviewed state has 9,147 records: 6,658 include, 2,442 exclude, and 47 defer. The 4,533 newly discovered canonical records received 3,468 include, 1,031 exclude, and 34 defer decisions. Within the backfill date boundary, including previously known records, the corresponding counts are 3,618, 1,072, and 34. The original register remains byte-for-byte unchanged; its SHA-256 is 9719c1184acacd287e3e86488194c239dc0f9180bfbb355f9a956ddd55d5b5ea.

Publication years use provider dates, which may contain estimated January 1 dates. For the backfill view the earliest qualifying backfill source date is used. The frozen ten-year view keeps its original qualifying date. These are collection dates, not claims about the exact date of scientific priority.

Discovery queries and continuation

The seven existing query families were rerun with the earlier date limits. The recorded native request files and ledger are authoritative. The core arXiv query was:

(ti:Andreev OR abs:Andreev)
AND submittedDate:[200001010000 TO 201609082359]

The related-literature query was:

((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:[200001010000 TO 201609082359]

Both arXiv families used ascending submission-date order, start: 0, and max_results: 200 initially. PatchScholar split a full recorded batch into provider requests of at most 100 records. The OpenAlex families used per-page: 200, initially page: 1, with the common filter from_publication_date:2000-01-01,to_publication_date:2016-09-08 and these terms:

These initial runs retrieved 7,837 provider records in 44 recorded batches. Those are retrieval counts before cross-query and cross-provider identity resolution. All seven reached a provider exhaustion condition. This establishes completion of the recorded queries, not completeness of the literature.

The corpus-specific backfill recipe saved each request before execution, the returned pagination state, original compressed payload path, start/end times, and new/touched canonical IDs. The next batch used the provider-native continuation specification. Completed batches can be resumed without replaying them. Failed requests are recorded separately; an error or safety-bound stop does not become a coverage claim. new_candidates in this ledger counts newly created canonical IDs, not newly accepted papers.

A separate raw-payload audit found that the numbered OpenAlex abstract pages contained 4,314 unique native IDs among 4,315 returned records: W103702087 appeared twice. The same query was therefore repeated using cursor: "*" instead of page: 1, with 22 separately recorded continuation batches. That run returned 4,316 rows representing 4,313 unique IDs; three IDs repeated. It reached next_cursor: null and provider exhaustion. Neither traversal alone matched the reported total of 4,315 unique records.

The union of the numbered and cursor passes contains 4,315 distinct native IDs. The cursor pass recovered W2159652144, Resonant proximity effect in normal metal/diffusive ferromagnet/superconductor junctions. It was already accepted as canonical ID 6066 through arXiv, so no new canonical record or review was needed. Two IDs appeared only in the numbered traversal: W1496829197 and W1971805134; their records remain preserved. This reconciliation supports coverage of this query at the recorded provider total, with a concrete warning that pagination order or provider results were not stable across traversals. It is not a guarantee of a transactionally frozen provider snapshot.

The combined initial and verification runs comprise 66 recorded batches, 75 provider HTTP pages, and 12,153 returned rows, including repeats and cross-query overlaps. Discovery ran on 2026-09-11 from 15:16:10 to 16:08:14 UTC. The missing distinct-ID diagnostic was reported to the PatchScholar thread via agentdispatch (sequences 649325 and 649393, 2026-09-11). The local validation manifest records native-ID counts, the reconciliation, HTTP page counts, and stopping conditions. The original raw payloads remain intact.

Screening, corrections, and preserved evidence

The new candidates were split into disjoint batches with frozen local metadata inputs. Three gpt-5.6-luna subagents produced CSV assessments with canonical ID, include/exclude/defer, labels, reason, source evidence, assessment=agent, and human_confirmed=false. A separate 56-record batch reconsidered old exclusions affected by the date extension; 55 changed to include and ID 73 remained excluded.

The first-pass classifications contained substantial errors. In particular, some records were assigned generic physics reasons despite unrelated titles; other records were incorrectly excluded even though their abstracts explicitly studied Andreev transport. The root agent audited exclusions containing relevant physics terms, inspected the titles of the 1,201 provisionally included records in the later OpenAlex portion, and read abstracts for ambiguous cases. Follow-up source checks supplemented these metadata inspections. These were targeted checks, not a blinded independent review or a measured classification-accuracy study.

The retained override registers document those corrections:

These are assessment-event counts and are not necessarily disjoint record counts. Original first-pass files are retained beside the overrides. An additional proposed artifact-classification file was rejected because it included records outside its assigned batch and misclassified valid physics papers; it was not applied to the corpus.

Examples of corrected false positives include geographical/humanities material, algal and oceanographic data deposits, whole-journal issue records, and other Andreev eponyms without a reflection or bound-state connection. Corrected false negatives include point-contact Andreev spectroscopy of magnetic materials, quantum-dot transport, FFLO bound states, multiple Andreev reflection, and superfluid analogues. Relevant theses and chapters are eligible when their content is substantive; a provider type alone does not establish relevance.

The consolidation recipe checks exact input/output IDs, disjoint first-pass ownership, required evidence fields, and complete coverage of every new canonical record. It overlays accepted follow-up assessments, preserves the frozen register, writes the expanded and backfill registers, and prepares a PatchScholar review pack. Applying that pack records the final assessments in the database; it does not rewrite prior review events. No human-confirmed judgments were added.

The 34 new deferrals remain in the follow-up queue. Some are title-only candidates for which available evidence does not settle scope. Others have apparent title/abstract mismatches. They remain visible and filterable in the explorer rather than being counted as accepted.

Identity, source quality, and reading depth

No additional identity merges were made for the backfill. Identifier-based canonicalization supplied by PatchScholar was retained. Unicode-preserving normalized-title comparison flags 329 shared-title groups involving 673 accepted records for follow-up. This includes possible preprint/article pairs and title collisions; it is not evidence that every group should be merged. The identity queue preserves the individual canonical IDs. Normalization keeps Unicode letters and numbers and removes punctuation/spacing; an ASCII-only normalization was rejected because it overgrouped non-Latin titles.

Metadata checks found apparent title/abstract inconsistencies, including local IDs 8425 (10.1002/andp.200510155), 8580 (W1654666886), and 8593 (W636971457), and a garbled conference-overview abstract at 8867 (10.1088/0953-8984/20/47/470301). These are evidence-quality flags, not confirmed provider bugs. Provider payloads and reconstruction must be compared before assigning the cause. The reusable quality-diagnostic gap was sent through agentdispatch to PatchScholar thread 054f1c76-e075-4e6d-94c3-9a6dae3b22a8 (sequence 649028, 2026-09-11).

This extension is predominantly title/abstract screening. It acquired no new full texts. The five selected full texts from the earlier construction remain available locally: four arXiv source archives and one PDF. Their originals stay intact; derivatives are ignored data files. Downloaded source was not compiled or executed. This corpus does not imply that every included paper has been read in full.

Explorer, network, and classification

The explorer offers separate backfill, expanded, original ten-year, original foundation, and all-screened collections. Each supports the existing statistical charts, record details, search, provenance, CSV download, network, and classification filters. The backfill collection uses its own qualifying dates; a canonical record can belong to both period views without duplicating its row in the expanded dataset.

The expanded snapshot has 42,406 directed cached citation edges among the 9,147 reviewed records (40,525 edges between accepted records), and abstracts for 7,698 records. Graph views retain only edges whose endpoints pass the active filters. Edge direction means “source cites target.” Missing references, unresolved identifiers, and provider coverage limit this graph; no similarity edges are invented. The default node limit remains 250, with truncation and isolates disclosed and a full connected selection option.

Physics, platform, and approach classifications retain the distinction between agent review tags and unreviewed title/abstract suggestions. The rules and their evidence are inspectable, and categories are nonexclusive. Counts describe metadata annotations, not validated scientific prevalence. The backfill screening errors reinforce the need to treat these filters as navigation aids and inspect the underlying record when drawing conclusions.

Backfill by qualifying year

The expanded series retains the original ten-year date assignments for its records. Publication-volume trends are sensitive to provider coverage, missing abstracts, version duplication, date conventions, and the broad related-work scope. The partial 2016 and 2026 boundaries should be considered when comparing period views.

Validation and reproduction

The decision manifest, local-state audit, and bibliography validation record the checks. The final register matches the database's review decisions, reasons, and labels. SQLite integrity and foreign-key checks pass; all 9,225 prior review events are retained and 4,589 new review events record this extension. The explorer snapshot is tied to the expanded CSV by its SHA-256. Its graph contains unique directed pairs and no self-links.

PatchScholar exported the full accepted bibliography. Pybtex parsed it and the two date-restricted bibliographies, and their exact key sets were checked against the corresponding accepted canonical IDs. The corpus-specific partition/export check uses Pybtex; this run used the installed interpreter at ../patchscholar/.pixi/envs/test/bin/python. It does not replace PatchScholar’s bibliography exporter:

• Expanded 2000–2026 bibliography: 6,647 entries.
• Backfill bibliography: 3,618 entries.
• All accepted, including older foundations: 6,658 entries.
• Expanded review register: all 9,147 judgments.
• Backfill review register: all 4,724 date-eligible judgments.

Seven Python checks and the JavaScript filter/count/graph checks passed. Chromium browser checks passed on desktop and mobile, covering collection counts, year charts, filters, details, source links, CSV downloads, history, network neighborhoods, classifications, report access, and empty/error states. A mobile overflow introduced by the two extra collection buttons was corrected.

The original ten-year and foundation exports remain separate. Local exports, raw downloads, and the database are ignored by Git and require backup alongside the tracked provenance. Git alone can reproduce the register-only explorer; it cannot reconstruct absent provider payloads or rich metadata exactly.

# No network or review mutations: validate the preserved local run.
pixi run python -m scripts.validate_backfill
pixi run explorer-build
pixi run explorer-test
npm test
npm run test:browser  # Requires the local preview server and Chromium.

# Discovery recipe example. Requires the normal OpenAlex authentication workflow.
# Existing successful pages are resumed, not overwritten.
pixi run python -m scripts.backfill_2000 openalex-core --pages 1

Re-running consolidation prepares files and a pack; applying a review pack is a separate mutation and should not be replayed over later human review. Exact research reconstruction needs the preserved local state and the PatchScholar checkout/environment, not just a fresh query against changing provider indexes.

Completion here means all recorded query families have been traversed, every new candidate has a recorded judgment, unresolved cases are retained, and the reviewed collection is explorable and exportable. Further citation expansion, full-text screening, identity adjudication, and human validation can improve recall and precision beyond this completed collection pass.