The concrete failure is familiar: you open your reference library to prepare a rapid evidence brief, meet a wall of unlabeled PDFs, duplicate records, and imports with no usable metadata, and realize you cannot say which studies are eligible, which have been screened, or which reports belong to the same study. Teams often carry this burden for years because imports accumulate faster than screening decisions. The result is not merely clutter: it is a research-stage problem. A library that cannot separate an already-included randomized trial from an excluded conference abstract, or link two versions of the same study, cannot support a defensible review question.
This article works through a repair sequence for exactly that situation. We start from a focused clinical example: a small team updating an evidence brief on perioperative dexamethasone for prevention of postoperative nausea and vomiting in adult patients. The population and outcome are narrow, but the repair decisions generalize to unsorted imports in any topic area. We diagnose the starting state, then move through the decisions that turn an unsorted pile into a screening-ready, deduplicated, linked library, and we close with the corrected deliverable and the checks that show the result is usable rather than merely tidier.
What Does an Unsorted Import Pile Actually Prevent You From Doing?
An unsorted library blocks the earliest and most consequential step of a systematic or rapid review: moving from a defined eligibility question to a screened, deduplicated set of study records. The Cochrane Handbook describes this sequence as planned searches derived from eligibility criteria, searching multiple sources, documenting the process, deduplicating records, screening in stages, and linking reports of the same study. When imports accumulate without those decisions, the library contains records of unknown status, and no defensible screening log can be reconstructed afterward.
The practical symptom in our dexamethasone example is typical. One folder contains five near-identical PDFs from a conservation-oriented import routine, another folder contains registry entries for the same trials without any linked published report, and a third contains abstracts that were never marked included or excluded. A clinician preparing a brief cannot say whether the library contains one trial or three, whether a retracted version is still the closest available record, or whether the most recent search superseded an older one. The failure is not storage capacity. It is the absence of the decision trail that evidence work depends on.
This matters because initial screening is the gate through which confidence in later synthesis passes. A team that cannot reproduce which records were screened, against what criteria, and with what result cannot explain why a study appears in the final brief. The repair therefore has to restore that trail, not merely attach better tags.
Diagnose the Starting State: What Is Actually In the Library?
The first repair step is a structured inventory, and it is deliberately boring. Before any renaming, tagging, or merging, record what exists. For a typical unsorted import pile this means separating records that were imported but never screened from records that carry a prior decision. Mark three stacks: import-only records with no status field, records previously marked included or excluded, and near-duplicate or multiple-report clusters.
A useful artifact here is a simple triage sheet with four columns: source, coverage, decision status, and metadata quality. In the dexamethasone example, the team finds the library contains a search export from a database, a set of trial registry entries, a folder of full-text PDFs collected during a journal club, and one .bib file from an old manuscript. The coverage column shows overlap: several registry entries appear to describe trials that also exist in the database export, but no field links them. The metadata-quality column shows that many imported files have no DOI and no matched publication record.
This inventory changes the goal. The team is not organizing papers; it is reconstructing the question that the imports were once meant to answer. That is why the repair begins with a working eligibility frame rather than with file moves. Write the population, intervention, comparison, and outcome in one sentence. In this example: adult patients undergoing surgery, perioperative dexamethasone, placebo or standard antiemetic, postoperative nausea or vomiting by 24 hours. That single sentence becomes the filter for every later decision.
Where Should the Repair Start: Deduplication or Metadata?
The common intuition is to start by deduplicating, because duplicates are the most visible problem. But deduplication before metadata repair is fragile. If two identical PDFs are stored under different file names and neither has a DOI, deduplication fails silently. If metadata repair runs first, the same records acquire the stable fields that make duplicate detection meaningful. The correct order in the Cochrane sequence supports this: records are deduplicated, but deduplication works because the records have identifiers such as DOIs or trial registration numbers. In an unsorted legacy pile, those identifiers are often missing.
The DOI Foundation describes a DOI as a persistent identifier that resolves through the DOI system and can remain stable when the location or descriptive metadata of an object changes. That stability is exactly what a repair routine needs: once a record carries its DOI, the record can be matched to Crossref metadata, updates can be detected, and multiple copies of the same article can be consolidated. But a repair team should not assume every imported PDF has a DOI. In the dexamethasone inventory, the .bib file extracted from an old manuscript had the strongest identifier coverage; the journal-club PDFs had none. Repair therefore proceeds from the records with the most recoverable identity to those with the least.
The decision artifact is a ranked repair order: first, the .bib file and any export with DOIs or registration numbers; second, registry entries that can be matched to published records; third, full-text PDFs that need identity recovery from title and author. This order respects evidence strength at the organizational level: records whose identity can be verified from a registered source are repaired first, while records that depend on manual matching are held for later and flagged as lower confidence.
How Do You Recover Identity for a Legacy File That Has No Identifier?
For records that carry a DOI, identity recovery is direct. Crossref exposes registered scholarly metadata through its REST API, including DOI, title, author, publication, funder, license, and relation fields when deposited. That means a file with a DOI can be enriched without relying on memory: the registered metadata tells you the canonical title, the authors, the journal, and often relations such as a correction or a version. In the dexamethasone example, this step catches the case the team feared: one imported PDF is an earlier version of a trial published later with corrected outcome data. The relation information, not visual similarity, is what exposes it.
For files without any identifier, recovery depends on conservative matching rather than automatic assumption. The team uses title plus first author to search registered sources, and it treats a strong match as a candidate link, not a confirmed identity. Records that cannot be matched are not deleted; they are placed in a labeled unresolved set with the fields that were used for the attempt. This is an important evidence distinction: a matched DOI is a strong identifier, while a title-and-author match is an inference that should remain visible as such. The repair log should record which type of evidence supports each reassignment.
Turning a Repaired Pile Into a Screening-Ready Queue
A library becomes usable when every record has, at minimum, four properties: a stable identity where one exists, a single canonical entry standing for all duplicates, an explicit link to other reports of the same study where applicable, and an unreversed screening decision. The last property is the most overlooked. A record that has been deduplicated and metadata-repaired but never screened is still not ready for synthesis; it is merely cleaner.
For the dexamethasone example, the team now has the artifact it needed for the search strategy at the center of the project. WisPaper search provides the scholarly search interface for finding academic papers, and an updated search is part of closing the loop on legacy imports. The repaired queue is then screened against the single-sentence eligibility frame, title and abstract before full text, with each exclusion carrying a reason. The Cochrane Handbook's staged-screening guidance is reflected here directly: screening in stages makes the decision reproducible, and linking reports prevents double-counting a study that appears once as a registry record and once as a published article.
Decisions at this stage should be logged, not inferred from tags. A record marked excluded during title-and-abstract screening should stay excluded until a documented full-text reason changes its status. The queue is finished when a reader can open any record and see whether it is included, excluded, or awaiting full text, and why.
Where Does Citation Quality Fit Into the Repair Checklist?
The legacy import pile usually contains at least one reference list that was assembled by hand or generated under time pressure, often the bibliography of an earlier draft. That list is evidence in the same way the PDFs are evidence, and it deserves the same audit. The failure mode is specific and severe: a bibliography with invented or mismatched citations can propagate into the next brief before anyone looks. A verification pass checks each BibTeX entry against real academic databases and flags references that cannot be matched.
For the dexamethasone example, the team links the repaired library to manuscript-ready references only after the BibTeX check. TrueCite checks BibTeX against real academic databases to flag hallucinated references. This is a discrete checkpoint in the repair sequence: after records are deduplicated and identity is confirmed, run the bibliography through the check before it enters the next draft.
The outcome to verify is narrow. Confirm that every cited source in the brief resolves to a registered record, that the bibliographic fields match the registered metadata, and that no citation points to a record the team could not independently locate. In the example, one entry from the old manuscript references a trial that the audit cannot match; it is flagged and excluded from the final brief until the team obtains the registered details.
Can You Ask the Library Questions Instead of Re-reading Everything?
Once the repaired queue exists, the temptation is to re-read every PDF to recover the trial details. For a focused brief that is inefficient and often leads back into the same unsorted state, because the reading notes are created in isolation from the structured decisions already made. A more useful step is to query the collected set as a unit: which included trials report 24-hour vomiting outcomes, which used dexamethasone dosing above 4 mg, which recorded serious adverse events. Those questions map to the synthesis fields the brief will need, so asking them against the library produces answers attached to the records themselves.
Questions based on papers uploaded to the library or added from search results can be directed through Library QA. This is the right place for questions that depend on the user's own screened record set: the team asks about the included studies it has already selected, not about an open literature search. The work product is a set of field-specific answers tied to the included records, which the team can then check against the PDFs for the small number of trials that drive the estimate.
What Does the Corrected Deliverable Look Like?
At the end of the repair, the team has a compact set of artifacts. The first is an updated eligibility frame with the population, intervention, comparison, and outcome written on one line. The second is a deduplicated record set in which each included record has a DOI or trial registration identifier where available, and each unresolved record is labeled as unresolved. The third is a screening log showing staged decisions and reasons. The fourth is a study-level link table joining registry records to published reports so one study is not counted twice. The fifth is a bibliography in which every entry passed the BibTeX check. The sixth is a brief note applying the evidence to the narrow dexamethasone question.
The corrected deliverable is not a larger library; it is a smaller, asserted one. In the example, the unsorted pile of dozens of files collapses to a handful of unique included trials and a clearly separated excluded set. The brief can then say which trials were included, which reports represent the same trial, and which limitations the source records carry. Applicability is stated against the chosen population: adult surgery, perioperative use, a specific outcome window. Indirectness is stated where a trial enrolled a mixed surgical population or reported vomiting only as part of a composite. Each of those statements is traceable to the evidence table rather than to a global claim about the literature.
What Did the Reset Teach That Transfers to the Next Review?
The transferable lessons are unevenly distributed, which is worth saying plainly. The teams that benefit most are those that store a decision, not merely a file: an eligibility frame, a screening status, a duplicate-equivalence link, and a verified bibliography. Teams that only store PDFs will repeat this repair on the next topic. The sequence also transfers: diagnose what is in the library before choosing a tool, repair identity before deduplicating, link study records before counting studies, and verify citations before drafting. These are workflow properties that do not depend on the clinical domain.
One part of the reset is institutional. The team makes a small standing rule: any new import receives a source and status field on arrival, and any updated search receives a date before it is merged into the library. This is consistent with the Cochrane Handbook's recommendation to update searches before publication, but it is applied at the operational level of the standing collection. In the dexamethasone example, the team notes the date of the most recent search and the register query beside the queue so the next update does not start from memory.
New imports are then dropped into the saved space where the team keeps the working set. For teams that want the import-and-organize path to be explicit, My Library is the WisPaper area for saving or uploading papers and organizing them into the library. The governing rule is that an item does not enter the active evidence set until it has the four properties described earlier. New workflow projects, such as the next review or the next brief update, can begin from that same organized collection through Research Projects. The intent is to make the reset the last full repair, not the first.




