Research integrity

How RSRC builds and maintains its library

RSRC maintains a curated discovery library and operates a source-linked evidence and synthesis system. This methodology distinguishes automated processing, editorial decisions, public review states, and the limitations that remain.

Methodology last reviewed: September 13, 2026

Methodology version: 1.16
Effective date: September 13, 2026
Status: Current public method

From a publication to a Research Insight

Library approval, mapping and synthesis inclusion are separate decisions. A newly approved publication does not automatically rewrite an Insight.

1. Discover and identifyFind a publication or dataset; check its identity, duplicates and relevance.

2. Approve for the LibraryAn editor approves the record. Source documents remain private; readers can follow the original source link.

Dataset: map for discoveryDescribe the data and assign relevant topics. No findings or future-work questions are added as synthesis evidence.

Stops here for syntheses. Dataset changes do not trigger Insight updates.

Study: check the evidenceExtract from the supplied PDF or Markdown. Check source excerpts, topic assignments and research eligibility.

3. Compare with evidence already consideredUnchanged inputs need no paid assessment. For new eligible evidence, assess whether it changes the accepted findings.

Study evidence: three possible outcomes

No material changeRecord the decision. Keep the accepted Insight. A new-evidence assessment may still incur a charge.

A revision is neededRevise affected passages. Check sources and all six editorial criteria before updating an eligible published Insight.

Checks or limits stop the updatePreserve the draft and spending record. Hold the update without automatic paid retries. Source problems can require withholding an article.

The three outcomes apply to study evidence. Dataset mapping ends at discovery. Mapping, AI completion and public publication are different milestones.

Scope and product boundaries

RSRC is a specialist research-curation project focused on romance scams, romance fraud, relationship-based fraud, and materially relevant adjacent evidence. It is not a victim-service intake system, a substitute for the original publications, or a completed systematic review of every database and language.

Public library

Bibliographic records approved by a human curator. A record may remain useful here even when RSRC cannot obtain a full PDF.

Full-document research corpus

Studies eligible for synthesis: approved records with a privately stored PDF or Markdown source, completed mapping, source checks and a current research assessment. Datasets can be mapped for discovery but are excluded from this evidence set.

Research syntheses

An operational, versioned claim-to-source workflow. Approved versions can be released through Research Insights as public Living Evidence Syntheses; they remain updateable evidence products, not final consensus statements.

Inclusion in the public library means that an editor judged the record useful and in scope. It does not mean that RSRC endorses every claim, rates the study as high quality, or has independently replicated its findings.

Search and discovery

Scheduled scholarly search

The application is configured to run a candidate search daily against Crossref, OpenAlex, and Semantic Scholar. The routine job currently searches works published from 2025 through the current year and requests no more than 500 candidate results in one run. Curator-directed backfills can use earlier date ranges; the automated candidate screen accepts years from 1990 through the next calendar year.

Current search queries

romance scam
romance fraud
online dating scam
sweetheart scam
catfishing
pig butchering

Automated discovery and enrichment sources
Source	Operational role	Information used
Crossref	Candidate discovery, DOI registration, and metadata enrichment	DOI, title, venue, type, dates, authors, subject terms, abstract when supplied
OpenAlex	Candidate discovery and metadata enrichment	Title, type, DOI, publication date, authorship, and venue information
Semantic Scholar	Candidate discovery and abstract enrichment	Title, authors, venue, year, DOI, abstract, publication type, and open-access indicators
Unpaywall	Open-access enrichment after DOI discovery	Open-access status and a reported open-access location when available
Publisher pages	Metadata and abstract enrichment	DOI landing-page metadata; this automated step does not download the publication PDF

Additional discovery routes

Researchers and readers can submit a publication for screening. A DOI or authoritative source record is required.
RSRC extracts DOI candidates from the reference lists and links of privately stored PDFs and Markdown documents. Crossref, DataCite, and the DOI resolver are used to validate identity and flag possible citation mismatches.
Duplicate checks use normalized DOI and title information before a candidate enters editorial screening.

Reference-list metadata triage

Pending DOI candidates extracted from reference lists receive a scheduled metadata review before a PDF is available. RSRC checks DOI registration and bibliographic metadata from Crossref and DataCite and supplements it with matched title, subject, abstract, and language metadata from OpenAlex when available. Requests run in small, variable batches with randomized pauses. This is a mild relevance screen, not full-document analysis and not a final publication decision.

A candidate can be classified as metadata not found only when its DOI cannot be verified and completed Crossref and OpenAlex title searches also find no sufficiently close title-and-year record. That result leaves relevance unknown and does not reject the candidate: missing index metadata does not establish that a work is irrelevant or nonexistent. Metadata supports a relevance decision only when provider identity evidence agrees with the candidate; a valid DOI for a different paper is not used. Timeouts, rate limits, identity mismatches, service errors, error payloads, and incomplete provider responses cannot produce rejection. A metadata-only zero-focus decision requires a matched, substantive, likely English-language abstract with no romance-scam, relationship-fraud, or adjacent-fraud signal. Relationship terms and potential impersonation technologies such as deepfakes and voice cloning prevent a clear-zero exclusion, without establishing direct romance-scam relevance. Non-English text, uncertain language, a missing abstract, or sparse metadata remains unresolved rather than being rejected automatically.

Under the default schedule, unavailable services, metadata not found, and insufficient metadata are eligible for another check after seven days, up to three automatic checks in total. After that limit, the record remains unresolved without repeated requests. Administrators can explicitly request another check. Identity conflicts and adjacent-topic assessments are retained without automatic repeated relevance checks; these states do not assign a final publication score.

Strong title, subject, or central abstract signals can mark a candidate as possibly above 3 on the romance-scam focus scale and prioritize it for document acquisition. That flag does not assign the publication's final focus score, approve the record, or replace the editor's manual PDF attempt and scope decision.

The scheduled metadata review can automatically exclude an unprotected discovery candidate only for a clear metadata score of 0 supported by the matched, substantive abstract described above. A candidate with an uploaded PDF or an existing publication link is protected from this action. Sources, reasons, score ranges, attempt counts, check times, and a bounded recent check history are retained in the administrative Reference Discovery workspace. Metadata triage is separate from validation of malformed or definitively invalid DOI identifiers.

Current automated candidate pre-screen

The pre-screen is deliberately broad. Passing it creates or updates a pending record; it does not approve a publication.

Accepted source metadata types

journal-article, proceedings-article, conference-paper, dataset, book, book-chapter, edited-book, monograph, reference-entry.

Required phrase match

At least one of the following must appear in the combined title, subtitle, venue, keywords, or abstract:

romance scam, love scam, romance swindle, sweetheart fraud, pig-butchering, romance fraud, dating scam, online dating scam, sweetheart scam, catfishing, pig butchering.

Known promotional or spam phrases, blocked DOI patterns, implausible years, disallowed publication types, and non-book records without a journal or venue are excluded automatically.

Controlled metadata for public Library filters

RSRC preserves the original publication type, method, group-studied, and tag wording stored for each publication. A separate, versioned normalization layer assigns reviewed controlled labels to known aliases and equivalent terms. These controlled labels consolidate the public Library facets and selected detail-page labels without changing publication approval or focus scores. Explicitly classifying a record as a Dataset controls its separate discovery-only role; recognized dataset aliases are stored as Dataset.

Each controlled assignment retains the exact source value and the matching rule used. Methods and groups that do not yet match a reviewed rule remain searchable through their original wording but are not added as named public controlled facets until the vocabulary is reviewed. A record with no recognized label appears under “Not classified.” Administrators can audit unresolved source terms, and future metadata changes are normalized automatically under the current vocabulary version.

The topic explorer uses the same approved records and normalized classifications as Library search. Its counts refer to publications, which may overlap across topics, methods, or groups; several publications can describe the same study. Missing labels and low counts describe catalog coverage, not an absence of research worldwide. Dataset records remain visible for discovery and are excluded from Insight evidence.

Optional reading lists store public record numbers in the reader’s browser. Viewing and citation exports recheck which records are currently public. Exports preserve recorded bibliographic details and translation provenance; readers should check them against the original source before submission or publication.

Author identity and public profiles

RSRC preserves the publication's original author line while maintaining a separate normalized authorship registry. DOI-matched Crossref and OpenAlex records can supply ordered authorships, ORCID iDs, OpenAlex author identifiers, public institutional metadata, and alternate name forms. For books and other records without a DOI, an OpenAlex result is used only when one exact normalized title match also agrees with the recorded publication year within one year.

ORCID and work-linked OpenAlex identifiers are the strongest automated identity evidence. RSRC does not assign an ORCID from a name search alone. When no persistent identifier is available, an exact name is linked across publications only when supporting coauthor or institution evidence agrees; otherwise, separate provisional identities are retained. Conflicting identifiers are marked for internal attention and are not made public.

A public author page requires at least 2 distinct approved publications, each with an RS rating of 3 or higher, and identity confidence of at least 70%. Pending and rejected records never qualify or appear on an author page. These metadata-derived pages do not imply that an author is affiliated with or endorses RSRC, and they do not contain an automatically inferred biography.

These searches are reproducible as a configuration, not as a permanently fixed result set. External indexes change, records are corrected, APIs may return results in different orders, and rate limits or outages can affect a particular run. Internal ingest logs record run times, counts, failures, and incomplete processing.

Human screening and inclusion

Every publication that becomes public receives a human curation decision. Automated discovery narrows the queue, but it does not make the final public-library decision.

Identity and duplicate check. The editor compares the DOI, title, authors, year, venue, and authoritative source record and checks for an existing publication.
Scope review. The editor determines whether the work directly studies romance scams or provides useful adjacent evidence about mechanisms, populations, harms, prevention, enforcement, detection, or recovery.
Manual document attempt. The editor manually attempts to obtain each proposed PDF. RSRC does not automatically download candidate papers.
Access decision. If the paper is paywalled and the editor lacks access, or if no PDF can be found, the editor decides case by case whether the bibliographic record and available information justify keeping it in the public library.
Source-safety judgment. The editor may stop and reject a candidate when the download site appears unsafe or contains suspected malicious behavior or code. Access is never required at the expense of system safety.
Final curation decision. The editor approves or rejects the record and may record a reason or curation notes. Rejected records are not visible in the public library.

How the romance-scam focus score is used

When full-document analysis is available, AI assigns a 0-10 focus score: 0 means no romance-scam focus, 5 means a partial or secondary focus, and 10 means romance scams or closely equivalent relationship-based fraud are central. The editor uses this score as a screening aid, not as the sole public-library decision. Adjacent work may be retained when its relevance can be explained, and a seemingly on-topic work may still be rejected after source review.

Factors considered during publication screening
Generally supports inclusion	May support rejection or exclusion
A research work whose identity, source record, and relevance can be evaluated	Duplicate, promotional material, non-research commentary, or a record whose identity cannot be verified
Direct romance-scam evidence or a clear, material connection to an RSRC research question	No meaningful application to romance scams after editorial review
A DOI, publisher record, repository record, or other authoritative bibliographic source	An unsafe source, suspected malicious site behavior, invalid DOI, or materially mismatched metadata

A missing PDF is not automatically a Library exclusion. A supplied Markdown document can support text extraction and excerpt checks without page numbers. Without a usable private source document, a record can remain discoverable but cannot enter the synthesis evidence corpus.

Documents, extraction, and AI use

Private document handling

Only an authenticated administrator can upload a research document in PDF or Markdown (.md) format.
Uploads are limited to 50 MB. PDFs must pass file-type validation; Markdown must contain UTF-8 text with at least 500 characters of publication content.
The file is securely kept in private application storage outside the public web directory. Document access is restricted to authenticated administrators, and the file is not offered through the public library.
RSRC records the original filename, file size, upload time and user, and a SHA-256 hash so replacement or unexpected change can be detected.
The current upload check validates type and size and records a provenance hash; RSRC does not describe it as a malware scan. Manual source-safety judgment remains part of acquisition.

Text extraction

Approved PDFs are converted to UTF-8 text with Poppler's pdftotext layout-preserving mode. Page boundaries are retained so extracted evidence can be checked against cited pages. Image-only or scanned PDFs with too little extractable text fail this process unless a usable text layer is available. For long documents, the analysis input is limited to 180,000 characters and preserves material from both the beginning and the concluding pages.

Markdown text is read directly. Evidence excerpts are checked against the uploaded text, with section headings used for context and page numbers left blank. This check confirms that an excerpt appears in the supplied document; it does not establish that a transcription is identical to the publisher's original. Administrators view Markdown as plain text, without executing embedded HTML.

Structured AI extraction

Extracted document text is sent to the configured OpenAI model through the Responses API. The system instructs the model to treat publication text and metadata as source material, never as instructions, and requires a structured JSON response. Depending on what is missing from the record, the analysis can propose:

Methods and study design
Populations and geography
Identified gaps and future work
Limitations
Romance-scam focus score
Research-contribution score and rationale
Controlled topic assignments
Atomic findings with source excerpts and pages

The application records operational provenance such as task, model, prompt version, subject record, status, token counts, timing, and response identifier. AI activity logs intentionally omit the source text and generated prose, although the resulting research fields and evidence records are stored in the RSRC database.

Multilingual library queries

Public library searches that appear to be written in a language other than English can be translated into English before retrieval. The original query remains visible, and an accepted translation is shown with the results. A short query phrase is sent to the configured OpenAI model for language identification and translation; RSRC does not include the visitor's IP address or browser details in that request.

Translation is accepted only when the model returns a structured result above the configured confidence threshold. Decisions are cached and model requests are rate-limited. If translation is unavailable, uncertain, or fails, the library searches the original query. Translation changes only the retrieval terms; it does not alter publication records or guarantee that a relevant item exists in the collection.

Publication synopses are a separate process

Synopsis generation requires extractable text from the privately stored PDF or Markdown document, or a sufficiently detailed abstract. Document text is used when available and extractable; the abstract is the fallback. Input is limited before generation, output is capped at 500 words, and the requested output language is English. Title-and-venue metadata alone is not sufficient for automatic generation, and the software does not substitute generic catalog text when model output is missing, short, or unusable.

Before public display, a versioned automated gate checks candidate length, complete sentence endings, repeated text, model-refusal language, retired boilerplate, unexpected code or active-markup indicators, numerical claims, and lexical grounding in the available English source. Sparse and cross-language sources require editorial review because those checks cannot establish semantic equivalence reliably. A candidate is withheld unless the current source-bound fingerprint passes the automated checks or an authenticated editor approves that exact candidate. Editing the synopsis, source metadata, abstract, or private-document identity invalidates the prior decision. The gate runs when candidates or sources change and during a daily reconciliation.

Passing this gate means only that the candidate met configured publication checks; it is not expert verification of every claim. The Evidence and review status panel identifies whether AI-generated content is displayed and whether human or subject-matter-expert review is recorded. A synopsis remains a discovery aid, not a substitute for the publication. Readers should follow the DOI or source record and can report a suspected synopsis error.

Review-state definitions

RSRC separates where information came from, whether a source excerpt was found, and who made a review decision. These states must not be treated as interchangeable.

Definitions of RSRC curation and review states
State	What it means	What it does not mean
Curation approved	A human editor approved the bibliographic record for public display.	Not an endorsement of every claim or a formal study-quality rating.
AI-extracted	A model generated or classified the field from available metadata or document text.	Not human-reviewed merely because the field is present.
Machine source-checked	The recorded source excerpt matched the supplied document: cited-page text for PDFs, or unpaginated text for Markdown. Matching does not establish the fidelity of a transcription to its publisher original.	Not proof that the interpretation is correct, the study is valid, or the claim generalizes.
Automated policy decision	A record met configured confidence, topic, focus, and source-match rules and was approved or rejected by automation.	No person reviewed the item unless a reviewer account is also recorded.
Editor reviewed	An authenticated human editor made or confirmed the recorded decision and is linked to it internally.	Not necessarily review by a subject-matter or methods expert.
Human source-verified	A person manually confirmed the supporting source connection.	Not automatically an expert appraisal of methods, bias, or certainty.
Subject-matter-expert reviewed	Would require an actual recorded review by a person with the relevant expertise. No outside-review program is planned.	RSRC does not currently apply this label to the publication corpus.

Current automated research thresholds

These defaults govern the narrower research corpus, not the human decision to display a bibliographic record. An unreviewed publication with a focus score below 1 is automatically excluded from research-corpus processing. Topic assignments and evidence normally require confidence of at least 0.70. Study evidence must also be tied to an approved topic and have an excerpt matched to the supplied source. Dataset topics can be reviewed for discovery, but dataset evidence remains excluded from syntheses. Human assessment decisions are preserved rather than silently overwritten by automation.

Public publication and Research Insight detail pages include a compact Evidence and review status panel. The side panel reports the recorded source basis, whether the page displays AI-generated content, generation or update dates, passed automated checks, administrative approval state, human content-review state, and subject-matter-expert review state. Explanatory details expand on request. The panel does not introduce a human approval requirement: routine processing remains automation-first, and missing or historical provenance is not inferred from publication or automated approval.

Synthesis workflow

Status: operational living-evidence workflow. RSRC publishes approved Living Evidence Syntheses through Research Insights. They are versioned, updateable evidence products and are not presented as systematic reviews, final consensus statements, or substitutes for appraisal of the original studies.

Each topic has structured research findings and a reader explanation. Both use the narrower full-document research corpus and preserve a path from a passage to its declared evidence, page-linked excerpts, publications, and document hashes. The Research Map records approved topic assignments and source checks before that evidence can contribute to an Insight. Approval of a new publication does not by itself require an article rewrite.

Eligible sources only. A source must be publicly approved, research-assessed, backed by a private PDF or Markdown source, and current with document-mapping integrity checks. Datasets are excluded even if mapped or historically assessed.
An accepted starting article. Incremental processing requires an accepted reader explanation tied to the current findings and sources. A missing starting explanation stops processing before a paid generation request. The September 10, 2026 starting package contains nine explanations edited and reviewed in Codex and one previously accepted explanation retained unchanged. These are recorded automated editorial judgments, not human or outside expert reviews.
Evidence comparison. The system compares the eligible evidence with its saved record of what has already been considered. Unchanged accepted content with unchanged evidence and validation inputs needs no AI calls. New evidence is grouped before assessment; the default batching delay is 30 minutes.
Contribution before revision. For changed evidence, an AI assessment distinguishes additional support, irrelevant material, qualifications, new findings, contradictions, and invalidated support. A separate model check tests that decision. Additional support alone does not require rewriting. The assessment can incur a charge even when the article stays unchanged.
Revise the affected material. Necessary revisions start from the accepted findings and explanation. Passages retain declared evidence links and citation snapshots. Valid source checks for unchanged passages can be reused; changed passages require new checks. Changed evidence, source text, or validation inputs invalidate affected prior checks.
Check before release. Revised findings must pass coverage and source-support checks. The reader explanation must pass source validation and all six editorial criteria described below. The system checks again that the evidence has not changed before accepting the result. Drafts, checks, model and prompt versions, and spending records are retained.
Preserve publication controls. A passing revision can update an already published, synchronized Insight automatically. Manually edited posts are protected. New posts remain drafts by default and require a separate publication decision; the incremental workflow does not create their missing starting explanations. Only published records are available in Research Insights.
Public access and change history. Published syntheses provide source-linked content, a contents list when the document is long enough to need one, a bounded on-page revision summary with access to the full approved public history, and a downloadable PDF. Drafts and unsuccessful generation attempts are not included in the public revision history.

Datasets support discovery, not synthesis findings

Datasets can be approved for the Library and mapped to research topics. Analysis describes documented populations, collection methods, coverage, access conditions and limitations. The application prevents dataset mapping responses from adding findings or future-work questions to synthesis evidence, including when reusing older saved analysis.

Existing dataset mappings and historical evidence are preserved, but the evidence cannot enter a synthesis. Dataset uploads, approvals, metadata edits and topic changes do not trigger Insight updates. Mapping may incur its normal AI charge. A separate study that analyzes a dataset can qualify for synthesis through the usual source checks.

What the editorial grades mean

The six criteria are structure, a clear answer, natural language, concrete examples, explanation of the evidence, and practical usefulness. Each must receive an A, and source checks must pass separately. Examples must be supported or clearly identified as illustrations; useful guidance must distinguish observed findings from interpretations and untested proposals.

An A is a recorded judgment against this communication rubric. It is not a reader-test result, human review, or a rating of the underlying studies. Source checks assess support in the declared excerpts; they do not establish the validity of the whole study. Findings also retain a default minimum evidence coverage of 90% and a claim-support confidence threshold of 0.75. Material changes to the method or acceptance rules require a new methodology version.

The reader explanation uses inline author-date citations and a reference list following APA 7 conventions where the stored bibliographic information supports them. Missing bibliographic details are not invented. Use Print / save explanation to print this accepted text or save it as a PDF through the browser. Download the full synthesis PDF exports the underlying technical research record, with its own source list; it is a different document from the reader explanation. Both remain available from the Insight page.

Bounded processing and held updates

Incremental Insight updates default to a $2 limit per update and a shared $5 limit per UTC day, with at most 12 AI calls and two editorial writing attempts per update. Charges are estimated conservatively from tokens, and unresolved request costs remain reserved. These limits cover incremental Insights, not unrelated publication ingestion, mapping, or search translation.

A failed request, exhausted allowance, uncertain usage, or a revision that cannot meet the checks leaves the update held. Its draft and spending history are preserved; repeating the same job or command does not reset its allowance or authorize another paid attempt. A wider revision, changed topic definition, or batch exceeding 20 evidence changes also remains held rather than starting a full rewrite. A held update does not mean its new evidence has been incorporated.

The previously accepted article stays published while a replacement is being prepared or is held for further work. Routine changes to mapping, metadata, or the working evidence set do not withdraw it: readers continue to see the accepted version and its preserved citations until a replacement passes the required checks. A recorded source notice, explicitly rejected source or evidence, or damaged citation or claim records can still require withholding. Queue pause, incremental processing, and legacy full-generation controls are separate. The reviewed-baseline deployment enables incremental maintenance while keeping legacy synthesis and editorial generation disabled. Pausing the queue preserves waiting jobs; an active incremental update checks the pause before its next AI request.

No automated gate establishes real-world truth, study validity, or scientific consensus. A public synthesis should identify the balance of evidence, study limitations, uncertainty, contradictory findings, and the level of human or expert review it has actually received.

Update frequency and change control

The public updates page and topic feeds list dated approved Library additions and the latest accepted reader explanation for each published Insight. Library dates use the earliest preserved approval timestamp, falling back to the current approval date where no earlier timestamp is recorded. They are not the source’s publication date. An accepted explanation can reflect an editorial correction without new evidence. Routine metadata checks, unfinished drafts, and spending activity do not create public announcements. These views use existing records and make no AI requests.

Configured update frequencies for RSRC research workflows
Process	Configured frequency	Important qualification
Candidate discovery and enrichment	Daily scheduled run	External API availability and queue conditions can delay completion.
Reference-list metadata review	Variable batch every 15 minutes	No PDFs are downloaded. Missing, unavailable, conflicting, sparse, or non-English metadata cannot produce an automated exclusion. Eligible unresolved checks retry after seven days, up to three automatic checks by default.
Document mapping, source checks, and automated research review	Small batches configured every 5-10 minutes	Approved records need a private, extractable PDF or Markdown document. Datasets are mapped for discovery only.
Publication-date metadata	Hourly checks with a daily failed-item retry	Authoritative sources can disagree or omit a precise publication date.
Private document integrity	Daily verification	A replaced or changed document invalidates stale mapping until it is processed again.
Synthesis integrity and incremental updates	Corpus guard and automation checks every five minutes; new evidence normally waits 30 minutes for batching	Only affected topics are marked for refresh. Evidence is assessed before any necessary revision. Processing must be enabled and queue activity running. Held updates do not retry automatically. A published insight is withheld when its cited evidence no longer passes public-display checks.
Human curation and corrections	As reviewed; no fixed service-level deadline	Publication remains an editorial decision and is not made automatically.

Changes to curation, research assessment, focus and contribution scores, document identity, and selected chronology fields create internal corpus revisions. Impact analysis limits relationship remapping and synthesis consideration to affected research questions and topics. A completed no-change assessment records the evidence considered without creating an article revision. A held assessment remains unresolved. Approved public synthesis versions appear in the linked Research Insight revision history; publication metadata and synopsis edits do not yet have a comparable public item-level history.

Scheduled work does not restart failed paid mapping or related-work checks merely because time has passed. An exhausted or interrupted synopsis batch stays held for unchanged inputs; an explicit retry can authorize another paid batch. Related-work comparisons reuse a saved result when the question and evidence inputs match. Concurrent requests for the same uncached Library translation do not each start an AI call. Limited output-length retries require reported usage and room below the output ceiling; a timeout does not authorize a larger retry. The dollar caps described above apply to incremental Insights, not to all site activity. Other AI stages have separate call, output, queue, and scheduling controls.

Ongoing source-status checks

RSRC checks accepted DOI records for incoming publisher and Retraction Watch update notices through the Crossref production API. The routine batch checks up to 20 due records per hour and ordinarily revisits completed checks after 30 days. Failed or unresolved checks retry after increasing delays from one to 30 days. A provider outage or rate limit stops the batch.

For DataCite records, RSRC also checks deposited withdrawal dates and correctly directed replacement relationships, then checks supported Zenodo, arXiv, Harvard Dataverse and verified Georgia State repository records. Successful DataCite checks ordinarily repeat after seven days; incomplete checks use delays of seven to 30 days. ResearchGate and unsupported repositories retain visible coverage limitations. The public DataCite API can omit withdrawn records, so a missing response is not treated as clearance. Repository notices, exact observed versions and available reasons are preserved; an identified removed version can require reassessment even when another version is available. An inaccessible or unattributable tombstone remains unresolved.

These checks use bounded public metadata requests and make no paid AI calls. New versions and metadata edits alone do not trigger synthesis writing or replace the stored source document. Where the preserved file has been matched to an exact repository version, that binding is retained. Datasets remain excluded from new synthesis evidence selection. Actual source notices enter the existing hold and reassessment process; they do not authorize an automatic full rebuild.

Titles and translations. The Library uses English display titles for accessibility. For reconciled non-English records, it also preserves the original published title and document language. Publisher-supplied English titles take precedence; otherwise the translation is labelled as supplied by RSRC. This does not imply that the full publication is available in English. Search includes both titles. Copyable citations retain the original title followed by the English translation in brackets; reference-manager downloads retain the original title and an English-title note. Historical evidence snapshots are preserved. A verified publication-date correction takes precedence over later automated metadata enrichment.

The DOI and publication title must agree before a notice is attributed. A verified original-language title can establish identity when the Library displays an English translation. That verification is tied to the recorded titles, authors, year, journal, DOI and stored document; entering a translation alone does not establish a match. Retractions, corrections, expressions of concern, reinstatements, and other incoming source updates are preserved with provider attribution, notice identity, date when available, and first observation. An outgoing notice relationship is not interpreted as a retraction of the notice itself. Provider-not-found, incomplete, and conflicting results remain unresolved.

A recorded incoming notice pauses the source's use in AI research content. The bibliographic record, citation tools, source links, and dated notice remain available. Affected public syntheses are withheld and marked for reassessment using eligible evidence. The incremental workflow retains its starting-article, source, and spending requirements: a source hold does not authorize an automatic full rebuild. A replacement that passes the current checks can restore a previously public synthesis withheld by this monitor, provided no later administrative change supersedes that restoration. No original curation decision, human-review record, source excerpt, or historical citation snapshot is overwritten. A later missing notice or a reinstatement alone does not remove the source hold: stored documents and affected evidence have not thereby been reassessed.

A documented reassessment can release a correction hold after checking the corrected source and the evidence used by RSRC. The decision records the exact document, reviewed notices and evidence, its scope, and whether Codex performed the review. Repeated detection of that same reviewed notice does not reopen the hold. A new notice or a changed source document requires another assessment. The public record retains the notice and explains the reassessment; historical articles are not republished by releasing the source hold. Codex review is not human or outside expert review.

This is automated metadata monitoring, not a complete integrity investigation. Notice coverage is incomplete, especially for corrections and expressions of concern; some works lack usable DOIs or have notices not supplied to the checked providers. A repository copy may not carry notices affecting a separately published article. “No notice found” describes the dated provider response and does not certify source validity. Checks and notices are retained as internal audit records; the public record displays known notices, the last attempt date and DataCite repository coverage limitations. No outside review or automatic email is part of this process.

This methodology is updated when a material process, source, review definition, threshold, or publication rule changes. Each public revision receives a new version number, effective date, and entry in the revision history below. Minor copy or accessibility corrections that do not change the method may update the page without changing the major method description.

Correction procedure

Anyone can report a possible citation, metadata, synopsis, source-alignment, broken-link, or accessibility problem through the correction form. A correction report does not change public content automatically.

Receipt. RSRC stores the page, publication title and DOI snapshot when applicable, concern type, description, supporting-source link, and a stable reference number.
Triage. An editor moves the request through the recorded states received, triaged, or verifying while the concern is reviewed.
Verification. The editor compares the report with the original publication, DOI registry, publisher or repository metadata, and other authoritative sources appropriate to the field being challenged.
Decision and change. A confirmed correction is made through the editorial system and the request is marked resolved. A request can be declined when the current record is supported or the proposed change cannot be verified.
Decision record. The system records status, editor notes, reviewer, review time, and resolution time internally.

Reporter contact information is optional and is used only when a human editor needs clarification. RSRC does not send an automated acknowledgment to the reporter. Reports must not contain victim names, financial records, private messages, account details, or other sensitive case information.

Public item-level correction histories are not yet displayed. Until that feature is available, a reporter can retain the stable reference number, and RSRC maintains the correction decision internally.

Known limitations

The public library is a curated discovery collection, not a completed systematic review or evidence-of-effectiveness rating.
Routine automated discovery currently emphasizes recent work and uses English search phrases. This can miss older, non-English, poorly indexed, unpublished, or locally distributed research.
External metadata and access status can be incomplete, inconsistent, or changed after an RSRC search.
Paywalls, unavailable PDFs, unsafe download sites, and image-only documents create uneven processing depth across otherwise useful public records.
AI can omit, misclassify, mistranslate, overstate, or introduce unsupported details. Structured output and page matching reduce some risks but do not eliminate them.
Machine source matching shows that similar wording was found at a cited location; it does not assess truth, causal validity, bias, representativeness, or certainty.
RSRC does not yet apply a single validated risk-of-bias instrument across the varied qualitative, quantitative, legal, computational, review, and conceptual works in the library.
Public provenance panels report recorded source, automation, administrative approval, human review, and subject-matter-expert review states. Corpus-wide subject-matter-expert assurance is not currently available, and missing historical review data is not inferred. The synopsis gate detects configured warning patterns but cannot prove semantic accuracy or study validity.

Official Resources is a separate directory

The Resources directory links visitors to official reporting, support and information pathways by country. These entries are not research findings or synthesis sources. Only published entries appear publicly; uncertain or unavailable entries can remain private.

Entries retain their official source links, coverage details and verification history. Source-change monitoring flags entries for follow-up; it does not guarantee a service's availability, eligibility or response time. Visitors should check the linked official service for current requirements. RSRC does not receive case reports or provide case management through this directory.

Revision history and citation

Three most recent Methodology revisions
Version	Effective	Material changes
1.16	September 13, 2026	Keep the previously accepted Insight public during routine evidence updates and held replacement drafts. Preserve original publication history while retaining source-notice, explicit-rejection and immutable-record integrity safeguards.
1.15	September 12, 2026	Added public topic coverage with matching Library counts and explicit missing-classification limits, dated updates and topic feeds, browser reading lists with selected citation exports, and source-checked corrections removing legacy dataset references from two reader explanations.
1.14	September 11, 2026	Extended source monitoring to DataCite and supported repositories, with visible coverage limits, exact repository versions, withdrawal and supersession attribution, and no paid AI calls or rewrites for routine version changes.

Earlier revisions (14)

The complete history is retained here and in the downloadable PDF.

Earlier Methodology revisions
Version	Effective	Material changes
1.13	September 11, 2026	Added English display titles with original-language titles, translation provenance, bilingual search and citations. Documented verified title identity and document-bound correction reassessment while retaining notices and evidence decisions.
1.12	September 11, 2026	Stopped scheduled paid retries of failed mapping and related-work checks and unchanged exhausted synopsis batches; documented duplicate suppression, uncertain-usage safeguards, and the distinction between Insight dollar caps and other stages.
1.11	September 11, 2026	Added the workflow overview and dataset discovery-only boundary; reconciled PDF and Markdown source eligibility and excerpt checks; documented the public official Resources directory separately from research synthesis.
1.10	September 11, 2026	Documented accepted starting explanations, contribution assessment before selective revision, six editorial grades with separate source checks, APA-style citations, automatic updates to existing published Insights, protected manual edits, spending and writing limits, and held work without automatic paid retries. Clarified that source holds do not authorize a full rebuild and that Codex review is not human or outside expert review.
1.9	September 9, 2026	Separated missing metadata from relevance, added bounded discovery retries and adjacent-topic safeguards, and clarified recent-history retention. Added ongoing DOI source-status monitoring, preserved notice provenance, conservative research-use holds, and affected-synthesis refresh rules.
1.8	September 9, 2026	Updated reference-list metadata triage from a one-time curator tool to a scheduled, rate-conscious workflow; documented its narrow automated-exclusion boundary, uploaded-document protection, retained audit details, and continued prohibition on candidate PDF downloads.
1.7	September 8, 2026	Documented the conservative reference-list metadata triage process, including DOI and title existence checks, language-safe zero-focus rules, transient-failure safeguards, and the limited meaning of possible-relevance flags; added datasets to the accepted publication types.
1.6	September 7, 2026	Documented the normalized author registry, identifier provenance, conservative identity resolution, and the eligibility rules for public author profiles.
1.5	August 12, 2026	Updated the synthesis workflow from development status to its current operational state; documented automated approval and separate public-release controls, targeted corpus refresh, public synthesis contents and revision histories, PDF distribution, current provenance labels, and private document access protections.
1.4	August 12, 2026	Added item-level public provenance panels that distinguish source availability, AI-generated page content, automated checks, administrative approval, human content review, and subject-matter-expert review without inferring missing review states.
1.3	August 12, 2026	Documented the versioned controlled-metadata process used to consolidate public Library filters while preserving original source wording, matching-rule provenance, and conservative treatment of unresolved terms.
1.2	August 11, 2026	Documented confidence-gated translation of suspected non-English public library queries, including privacy boundaries, caching, rate limits, and fallback behavior.
1.1	August 10, 2026	Added the permanent, source-bound synopsis quality gate, fail-closed public display, daily reconciliation, and an authenticated editorial review queue; removed generic synopsis fallback generation.
1.0	August 10, 2026	First comprehensive public method covering discovery, manual screening, document handling, AI extraction, review definitions, synthesis development, updates, corrections, and limitations.

Suggested citation

Romance Scam Research Center. (2026). RSRC Research Methodology (Version 1.16). https://romancescamresearch.org/methodology

Request a correction Submit research Contact RSRC