Scientific literature search, explained.

A scientific literature search, explained: from a question to a sentence you can stand behind

The route of one literature question, from wording to a kept answer

1. Write the question in partsThe target, the mechanism, the model system, and what an answer would have to contain: a rate, a count, a sentence.
2. Name the sources and write the strategyWhich databases, registries, patents and internal stores, and the exact search for each.
3. Run it and screen what comes backRemove duplicates, read titles and abstracts, read the full text of what survives, and record each reason for setting a paper aside.
4. Read the passage and tie each statement to itOne line per source: year, model system, how many, what it found, whether it supports the claim. This is the step this guide cares most about.
5. Keep the record, and run it againSave the strategy and the table. Check for retractions. Rerun before anyone relies on it months later.

The step numbers are a teaching aid; formal reviews add a protocol and a second reader at most stages.

Why a repeat search rarely matches, and three more surprises in literature work

Strategy, recall, precision and screening: four words for this work

Fourteen terms are defined in the one-page summary.

Quick question, rapid review or systematic review: how much proof each literature search owes its reader

In shortNo law sets a method for a discovery search. What a search owes depends on who will rely on it. A note for your own bench needs little. A systematic review for a journal or a payer owes a published strategy, two readers and a count of everything set aside. Most company searches sit between the two, with no written rule at all, while the pile of papers keeps growing.

The survey behind that four-hour figure also found that between 2011 and 2019, researchers read 10% fewer articles while spending 11% more time finding them.7 It was carried out by a publisher with a charity partner, and it is seven years old, so treat it as a rough guide. We found no newer public figure for drug companies, and no market size we could verify for literature search software. To get your own figure, ask ten scientists to log a week.

What each grade of literature search owes the person relying on it

Three grades cover most of the work. The names vary from one company to the next; the difference that matters is what you could show if someone asked.

Systematic reviews have two reporting standards. PRISMA 2020 lists 27 items for the whole review.9 PRISMA-S adds 16 items for the search alone, including the strategy for each source "copied and pasted exactly as run" and the date of the last search.8 Those standards bind journals and many payers. Inside a company, nobody enforces them. The useful question to ask of any search is the plain one: if this answer were challenged in a year, what could you show?

Where the free text is, and why it has grown

More of the literature can be read without a subscription each year. In PubMed, the share of each year's records marked as having free full text rose from 44% for 2015 to 65% for 2025.6 A 2018 study estimated that at least 28% of all scholarly articles were open access, and 45% of those from 2015.10 NIH-funded manuscripts accepted on or after July 1, 2025 must now reach PubMed Central on publication, with no embargo.11 That is cited from a university library guide; we did not read NIH's own notice.

What you may copy, share or mine from a free article depends on its license; sections 10 and 11 explain.

How to write a literature search strategy: split the question, list the words, trade recall against reading

In shortA strategy turns a sentence into blocks of terms joined by AND, with synonyms inside each block joined by OR. Widening it finds more of what matters and gives you more to read. The way to know whether it works is to test it on papers you already know should appear. Before relying on it, have a second person check it against a short published checklist.

From one sentence to three blocks of search terms

Take the question from this guide's example: what is known about resistance to covalent KRAS G12C inhibitors through MET amplification? It has three parts that a paper must touch to be relevant. Each becomes a block, using both the database's own subject headings (in PubMed, Medical Subject Headings, a controlled list that indexers assign) and the words authors really use.

A MEDLINE strategy does not transfer to another database as written: field tags, proximity operators and headings all differ. Each database needs its own translation, and each translation is a separate strategy to save. Cochrane's advice on keeping them is blunt: "The search strategies should not be re-typed, because this can introduce errors."3

Recall and precision worked through on a known set of 34 papers

You cannot measure recall on a question nobody has answered, because you do not know what you missed. The practical trick is a test set: before you search, list papers you already know belong, from a review or an expert. Then see how many of them your strategy finds. Suppose your list holds 34 papers (these numbers are illustrative).

Published reviews usually include about 3% of what their searches find: 2.94% on average across 195 reviews.13 A low precision is normal. A low recall on your test set is the warning.

The six things a second person checks in your search

Peer review of a search is cheap and it works. Cochrane "strongly" recommends it before the search runs.3 The PRESS 2015 guideline keeps the check to six points:14

It catches real faults. In 137 reviews from January 2018 with explicit PubMed strategies, 92.7% of the search strategies contained some error, and errors that hurt recall were the most frequent, at 78.1%: mostly missing terms and mistakes with subject headings.12

PubMed, Embase, Google Scholar, registries, patents: which sources a scientific search has to reach

In shortNo single database holds the evidence. Four general sources together reached 98.3% recall in one study, and an estimated 60% of published reviews fell short of 95%. Beyond the journals sit conference abstracts, trial registries, preprints, patents and your own reports, none of which share a search box. Writing down what you left out is part of the method.

That 2017 study came from one information specialist's recorded searches, so its setting is narrow. Its best combination was Embase, MEDLINE, Web of Science Core Collection and Google Scholar, with 98.3% overall recall and 100% recall in 72% of reviews. Its authors estimated that 60% of published systematic reviews do not retrieve 95% of the available relevant references, because they skip important databases.2

The sample question in this guide comes from cancer biology. A chemistry-heavy or device question would call for other specialized sources; the same study notes that specialist databases add unique references when the topic fits.2

Search engine or database: what a repeat search needs

A database with fixed fields lets you write a statement that can be run again. A web-scale search engine is good at finding something relevant quickly. A 2020 test of 28 academic search systems found that only half could be recommended for evidence synthesis without substantial caveats, and its authors wrote that their findings "demonstrate why Google Scholar is inappropriate as principal search system".15

That is a judgment about the main search of a formal review. As a second source for a working question, or for reports and theses, it remains useful, and the 2017 study above includes it for that reason.2

The evidence that never reached a journal

Conference abstracts show how much stays out. Across 425 reports covering 307,028 abstracts, only 37.3% were published in full, and abstracts reporting a significant result were more likely to make it (risk ratio 1.31).16 So a search of journals alone is a sample tilted toward results that worked. Librarians call the rest gray literature: abstracts, registry results, theses, regulatory reviews and company reports.

Your own reports are the part of the gray literature only you can search. A translational interim report or a signed notebook entry may hold the most relevant result in the world on a narrow question. It is also the source that a published-literature tool cannot read. Section 9 explains why that matters, and section 14 how teams handle it today.

Why literature searches cannot be repeated, and the six facts that make one reproducible

In shortStudies have tested this directly. One review in 100 could be fully repeated from its published details, and about half of searches re-run by independent specialists differed by over 10% in records found. The good news from 2026: those differences seldom changed the final result in a meaningful way. The fix is clerical. Save six facts for each search on the day you run it.

Those figures come from two sets of published systematic reviews: 272 articles from high-impact pediatrics, cardiology and surgery journals in 2012,18 and a random sample of 100 reviews indexed in MEDLINE in November 2021.1 In the later study, six searches differed by more than 1,000% from their reported result.1 These are the formal reviews whose methods get written up. Working searches inside companies usually leave less behind, though we found no public measurement of them.

A 2026 replication adds a useful check. Two information specialists repeated the searches of 10 reviews, 43 searches in all. Record counts differed by more than 10% in 58% of searches for one replicator and 49% for the other. Eight of the 10 meta-analyses first looked not fully replicable; after screening and data discrepancies were resolved, five did. Differences were judged meaningful in one blinded replication and none of the unblinded ones.19

Its authors' reading is fair: results were not always consistent, and the inconsistencies seldom changed summary estimates in a meaningful way. An unrepeatable search is mainly a problem of trust and checking. You cannot tell which search is the rare one that mattered.

Six facts to save for each search on the day you run it

Item numbers follow PRISMA-S.8 The 2024 study of 100 reviews checked six key items that differ slightly: database name, multi-database searching, full strategies, limits and restrictions, dates, and total records.1

When a saved search goes stale

A search answers the question as of its date. Cochrane requires a rerun if the search date is more than 12 months, and preferably six, from intended publication.3 A company can set its own clock by what is at stake. A study plan about to be signed deserves a rerun that week, in a field where a key paper can appear any month. What matters is that the reader can see the date and decide.

Screening thousands of abstracts: how often careful readers err, and what a second reader catches

In shortScreening is reading what came back and sorting it against written rules. Trained reviewers make mistakes in about one decision in ten. One reader alone misses a median of 5% of the eligible studies, and a less experienced reader misses more. A formal systematic review takes over a year on average, which is why teams look for shortcuts. Knowing the size of the human error helps you judge any tool that claims to replace it.

What one reader misses that two would catch

Reporting standards prefer two independent screeners because people miss things. A review of the evidence found a median of 5% of eligible studies missed when one person screened, with a range of 0 to 58%.21 The experience gap is the part to remember.

Those numbers rest on only four evaluations and nine reviewers, so read them as a size and a direction. When the missed studies were traced into meta-analyses, findings changed substantially in three of seven single screenings, all by the same less experienced reviewer. The authors' view is that one experienced screener can be an acceptable shortcut in a rapid review, and that single screening is not equivalent to double.21

Write the rules before you read the first abstract

Screening goes wrong when the rules form as you read. Write them first: which species and model systems count, which dates, which study designs, whether reviews and conference abstracts stay in, and what to do with a paper you cannot get in full. Record a reason for each paper set aside at full text. PRISMA 2020's flow diagram expects those counts, and a working search benefits from the same habit in a short form. Six months later, the first thing anyone asks is why a paper they know about is missing.

Evidence tables and citations back to the source: one row, one passage, and the numbers behind it

In shortCitations go wrong more than people think: one in four to one in seven in medical journals, depending on how you count. Extraction into tables also goes wrong. The defense is a table in which every row carries its model system, its count, its effect and the exact passage. Reading the count beside the percentage stops a small sample from looking like a finding.

The two quotation figures use different denominators. The 2015 meta-analysis found 25.4% in total,4 and the 2017 recalculation, counting per quotation, found 14.5%, of which 64.8% were major: the source failed to support or contradicted the claim.23 On extraction, a methods review reports "up to 50%" from only four studies of error frequency, and warns that the evidence base is weak.22

Six columns that let a reader judge one source line

Reading 12 of 23 against 1 of 4: why the percentages mislead

Take the two numbers from the example at the top of this page. They are fictional, and the arithmetic is what matters. The published series found MET gain alone in 12 of 23 samples, 52%. The internal report found it alone in 1 of 4, 25%. They look like a clear disagreement, so we worked out how much these counts can tell you.

On a Fisher's exact test, a standard test for small counts, the two-sided p value is 0.60. Those four samples cannot show the published series wrong. They also cannot be dismissed, because the internal report describes a different population and a second alteration, confirmed in two matched biopsies. The right response is to go and find out: more samples, a look at the assays, a word with the author. A good search hands the person that question, with the counts in plain sight.

How to find retracted papers before you cite them, and why preprints need their own label

In shortPubMed marks retracted papers, and you can filter on the mark in minutes. The hard part is the citations made before the notice, and the answers you already kept. Few later citations mention the retraction at all. Preprints are a separate issue: they are growing fast, and they are useful as long as the row says what they are.

Each bar counts papers by the year they were published. A retraction often comes years after publication, so the low numbers for recent years will rise, and they show no fall.

What PubMed marks, and what the mark leaves out

The National Library of Medicine gives a retracted article the Retracted Publication type and the phrase "Retraction in:" on its record, and links a separate notice carrying "Retraction of:".25 It covers partial retractions and corrections under errata, which can touch the very table you cited.25 It also does not separate articles retracted for honest error from those retracted for misconduct, plagiarism or the like.25 So a mark tells you to go and read the notice. It does not tell you why the paper was withdrawn or who was at fault.

Three checks cover most of the risk. Run the filter on your reference list before anything leaves your team. Read each notice and erratum it turns up: Cochrane expects reviewers to read relevant retraction notices and errata, including for studies already included when a review is updated.3 And recheck what you already kept: a saved table has no way of knowing that a source was withdrawn after you made it.

The citations made before the notice

One study followed the publications of a single research group with 113 published integrity concerns. Of 3,925 references to papers with a published concern, 3,688 appeared in articles published before the concern, and very few later publications assessed the effect on their own findings.26 Across PubMed, retracted papers kept being cited, and "only 722 (5.4%)" of 13,252 post-retraction citation passages acknowledged the retraction.24

Preprints: useful, early and labeled

A preprint is a paper shared before peer review. In a fast field it may be the only source for several months. Treat it as you would any row: name what it is, say it has not been reviewed, and check again for the published version. PubMed indexes only some preprints, so the count above understates what preprint servers hold.

When published findings and your own lab disagree: what the reproducibility studies do and do not show

In shortCompany scientists have long reported that published results often fail to match their own work. One much-quoted count from Bayer found published and in-house data in full agreement in about a fifth of 67 target projects. A large open project repeating cancer biology experiments found effects a median 85% smaller. A disagreement usually has ordinary causes, and finding one early is cheap, if someone is assigned to look.

That analysis covered most of Bayer's work in oncology, women's health and cardiovascular medicine over four years, and it was reported in 2011.27 It counts in-house projects, and it is the company's own retrospective with no shared data. It is fair to say the two sets of results disagreed. It is less fair to say the papers were wrong. A year later, scientists at Amgen reported a success rate of 11% in their own check of preclinical cancer papers.28

The open attempt to measure this properly was the Reproducibility Project: Cancer Biology. Of 193 experiments it planned to repeat from high-impact papers, it completed 50, because the original papers often lacked the detail or the materials. Across those, the median effect in the replications was 85% smaller than the original, and 40% of 97 positive effects replicated by three or more of five binary tests.

The authors add a caution worth keeping: "A successful replication does not definitively confirm an original finding or its theoretical interpretation", and a failure does not disconfirm one, though it suggests more investigation is needed.28

The best-known dollar figure is an estimate: more than 50% of US preclinical research irreproducible, about $28 billion a year. Its authors built it from four categories of error, study design, biological reagents and reference materials, laboratory protocols, and data analysis and reporting, and their own bounds run from 18% to 88.5%.29 It shows the likely scale of the problem and does not measure one company.

The ordinary reasons two good labs get different answers

Most of these sit in the methods sections and the internal report's appendix, which is why the check needs both documents open at once. It also needs a named person to rule on it, usually whoever owns the internal data. A disagreement with no owner stays a surprise until the wrong moment: a reviewer's question, a failed study or a board meeting.

Copyright for corporate research: what a company may copy, keep and share from a journal article

In shortAccess to a paper does not settle what you may do with it. In the US, the leading case on company scientists copying journal articles went against the company. Your rights come from your subscription or license, any open license on the article, and the law in each country. Most companies run on a collective license plus publisher agreements, and the terms of both decide the answer.

The Texaco case: why copying for company research is judged differently

US law allows fair use after weighing four factors: the purpose, including whether it is commercial; the nature of the work; how much is used; and the effect on the market for it.30 In American Geophysical Union v. Texaco, 83 publishers sued over a company scientist who photocopied eight articles from a chemistry journal for his files. The appeals court held that this was not fair use.31

The court was careful about its scope. It treated the scientist as one example of 400 or 500, and the question as "institutional, systematic copying". It did not decide on copying "by an individual, for personal use in research". It also gave weight to the working market for licenses, including the Copyright Clearance Center (CCC).31

Scientists sometimes assume that research itself makes a copy fair. That court said the idea of a use with a new purpose would be stretched "beyond recognition" if it applied simply because the copier "acted in the course of doing research".31

What each kind of article lets a company do

Reprints and team folders: where everyday copying happens

The everyday risks are small and frequent. A shared folder of PDFs built by a project team. A set of articles sent to a partner or a contract research organization, which is an outside party. A reprint a medical team hands to a doctor, which normally needs its own permission. A PDF uploaded into a consumer AI tool to get a summary. Each of these copies a work, and each one may or may not be covered by the license.

A good rule is to share links where you can, keep citations and short passages in the record, and have your library or legal team write down what is allowed so scientists do not have to guess.

Text and data mining at a company: the EU, UK and US rules, and why AI search raises a new question

In shortThe EU lets anyone mine lawfully accessible works unless the rightholder reserves the right. Its stronger research exception is closed to most companies. The UK's exception is for non-commercial research only. US law rests on fair use, and in September 2026 an appeals court ruled against a company that copied a rival's content to build a tool that returned passages. Licenses do most of the real work. Ask yours three questions.

The three places a company's rights differ

Other countries have their own rules; check with counsel wherever your scientists sit. A German appeals court has already shown how much detail matters: in the LAION case, it found a rights reservation "not machine-readable" on the facts, and the case can still go further.40 We read a law firm's note on that judgment, not the judgment itself. The Directive, section 29A, the UK government's report and the Third Circuit's opinion below were read in their official texts.

Training, indexing, quoting and summarizing are four different uses

Writing about AI and copyright mostly discusses training models on books. A company's literature search does something narrower, and it matters which of these it does.

The US Copyright Office's May 2025 report, still labeled a pre-publication version, treats retrieval separately from training. In fair use terms, it says retrieval is less likely to serve a new purpose when the aim is to generate outputs that "summarize or provide abridged versions of retrieved copyrighted works", as opposed to hyperlinks, and that it can also substitute for the original in the market.38 The report is the Office's analysis, and courts are not bound by it.

The Third Circuit decided Thomson Reuters v. ROSS on September 29, 2026, under seal, and filed the public opinion on September 30. Contractors working for ROSS used Westlaw headnotes to write the memos that trained ROSS's legal research tool, a point ROSS did not contest on appeal. The court noted the tool "was not a generative AI" and "would only return text passages", found that ROSS's use shared "the same ultimate purpose" as Westlaw's and added little new, and held it was not fair use, affirming a partial summary judgment on 2,243 headnotes.

In a footnote it set this apart from the generative models in Bartz v. Anthropic and In re OpenAI, and it weighed the fact that ROSS meant to compete with Westlaw.37 Our reading: a drug company reading papers for its own research builds no rival to a journal, so the opinion does not decide that situation, and it shows that a tool which only returns passages gets no automatic pass.

Three license questions to settle before any tool reads your full text

Non-patent literature and patent search: keeping a claimed dose apart from a measured one

In shortPatents show what other groups think is worth protecting, often before they publish. They also lag: applications generally stay hidden for 18 months, so a clean search says nothing about the most recent filings. The largest trap is reading a claim as a result. A claimed dose range states what an applicant wants to protect, and it measures nothing. Whether a position is safe is a question for counsel.

Papers, posters and theses as prior art

Patent examiners search both patents and everything else, which they call non-patent literature. A document counts as a printed publication if it was "disseminated or otherwise made available to the extent that persons interested and ordinarily skilled in the subject matter or art exercising reasonable diligence" can locate it.42 Public availability is the test, and peer review is not required. That is why a poster or a thesis can count. It cuts both ways: your own conference abstract can count against your own later filing, depending on timing and country.

US rules also put a duty of candor on those substantively involved in filing a patent application: they must disclose information known to them that is material to patentability.43 So what your team's searches turn up can matter to your counsel, and who sees those results should be a deliberate decision.

Why a claim and a result need different labels in one table

In a merged list, a patent and a paper look alike: a title, a date, a dose. They mean different things. A paper reports what someone measured. A claim of, say, 200 to 400 mg draws a boundary around what the applicant wants to protect. Count it as support on a claim about doses in people and you have invented evidence. Keep patent rows visibly separate in any table, count them apart, and write "no result in people" when that is the truth.

What AI literature search gets right and wrong: hallucinated references, missed papers and who answers for them

In shortModels asked to produce references from memory invent some (often called hallucinated references), even in 2026. Models with web search still write statements their sources do not support. In one test, a purpose-built research tool found 39.5% of the included studies, against 94.5% for traditional searching, though it needed far less reading. AI screening does better, at about 92% sensitivity. FDA's draft guidance leaves drug discovery out, and evidence groups and journal editors keep responsibility with people.

The 2024 test counted a paper as invented if any two of title, first author or year were wrong, and models found few of the right papers: recall was 11.9% and 13.7% for the two GPT models.44 The 2026 test ran 350 queries in one specialty, and 8% of queries had only fully accurate citations.45

Giving the model the web helps with this failure and leaves another. Across seven models and 58,000 statement and source pairs, between 50% and 90% of responses were not fully supported by the sources they cited, and about 30% of individual statements from one web-searching model were unsupported.46

High precision, low recall: one AI research tool against traditional searching

Those four case studies concern one version of one product, and its authors note that it keeps changing. They also found that the tool turned up some studies the original searches missed, and they called it a useful adjunct and too insensitive to replace traditional searching.47 That points to a fitting use: a fast first look, with a formal search still required when missing a paper would matter.

Screening is where the evidence is strongest. A systematic review of 148 performance assessments reported median agreement with human reviewers of 0.92 on included records and 0.89 on excluded ones for title and abstract screening, with a wide spread between studies.48 A 2026 meta-analysis of 18 studies pooled sensitivity at 0.92 and specificity at 0.94.49 Set that against the 10.76% human error rate from section 6 and you can see why teams are interested, and why a tool's result still needs measuring on your own question.

Five jobs called "AI literature search", and how far each can be trusted

What FDA, EMA, editors and evidence groups say about AI in this work, as of November 2026

Once a literature answer enters a regulated record, such as a filing or a quality document, the usual rules for that record apply to it: who wrote it, who checked it, and whether it can be traced to its source. Controls in a tool can support that work. The responsibility stays with your people, and so does the checking.

How R&D and medical teams run literature searches now: the scientist, the librarian and four separate search boxes

In shortIn most companies, the scientist does the searching alone, a library or information team helps on bigger jobs, and patents and internal reports are searched separately, if at all. The answer goes into slides or email, and the next person starts again. In the illustrative week below, most of the time goes on finding, tabulating and checking, and reading gets what is left.

Scientist, information specialist, program lead: who touches a literature question

Six steps and four systems behind one answer about resistance

Where a scientist's literature hours go in a week

Estimated, and shown to give a shape to test. The only published figure here is the survey's "just over four hours" of searching for articles and more than five reading them, in 2019.7

From reference managers to question-answering search: four ways teams work

"What is known about resistance through MET amplification?": eleven seconds to numbered claims, then an amber row

In shortThis is the example from the top of the page, inside the product. Dr. Maya Chen, a translational scientist on Program CVT-416, asks in plain words. Agents, pieces of software that each do one defined job, search papers, patents, internal reports and signed lab notebooks together. They answer as numbered claims, each made of evidence rows, and each row opens its passage. One row turns amber because a company report disagrees, and only the owner of that report may settle it.

1

Yesterday's answer is still there, with its rows

It is Tuesday, October 6, 2026, a little after 8 a.m. Ask opens on Maya's last question from October 5: does MET copy gain show in plasma before tissue? That answer is kept in the program notebook: 3 claims, 7 rows, 6 sources, 1 of them the company's own, no disagreement found.

The scope sits under the question: Program CVT-416, all four source types, 2021 to 2026, checked against the company's own work. Clicking a row shows its passage on the right.

2

The run shows each step as it happens

Maya asks about resistance through MET amplification. The panel shows the question scoped to what she may open, then 128,412 sources searched and 36 passages read in full. Five claims follow, from 17 evidence rows across 12 sources, 4 of them the company's own.

Only after the claims are written does the citation checker read every row back against its passage, and then the agreement check starts.

3

Numbered claims, each a block of rows

The short answer comes first. Under it, claim 1: MET copy gain is the most common single new alteration at progression, in about one in ten people. Its four rows are a plasma series at 10.8% (23 of 212), a review of 11 series, the company's VLT-302 interim 2 at 10.5% (4 of 38), and a colorectal study that limits the claim to lung.

The meter reads 3 support, 0 contradict, 1 limits it. The citation checker read all 17 rows: 17 supported, 0 removed.

4

One row turns amber

Claim 2 says that when MET gain appears, it is usually the only escape route. The published series found it alone in 12 of 23. The company's VLT-302 interim 2 found it alone in 1 of 4, with three samples carrying a second alteration. The meter reads 2 support, 1 contradict, and the claim lands amber.

The agreement agent's note starts with the weakness in its own finding: small numbers, four samples, and a lead for a person to check. Section 7 showed why that caution is right.

5

The models, and what no one has shown in people

Claim 4 reports that adding a MET inhibitor restored pathway suppression in resistant models: 62% xenograft shrinkage at day 21, and 71% growth inhibition against 18% for CVT-416 alone in the company's RX-0942 study. The last row limits the claim, with regrowth within 11 days of stopping.

Claim 5 is a gap: no report of the combination in people. Its first row is a patent claim, with its own label.

6

A patent row reads as a patent

Opening that row shows the passage from sample patent family WO 2025/998432: a daily MET inhibitor dose of 200 to 400 mg, which overlaps the planned 300 mg arm. A note under the facts explains that patent text describes what an applicant wants to protect.

The row counts as a patent claim on the meter, never as support. Whether the overlap matters is for Omar Haddad, the patent counsel.

7

From the row to the report, at its page

"Open at page 18" takes Maya to the VLT-302 translational interim report with the cited sentence highlighted. Beside it, the panel says why this passage came back, what the source studied, who wrote it and who may read it: owned by Program VLT-302, program only. The note at the top says the citation checker read the row against this passage and found it supported.

8

Both passages on one screen

"Compare side by side" lays the published passage beside the company's own, each with its n and model system. Below them, a note on why they may differ. If the finding holds, a MET inhibitor alone would leave three in four of these patients with a second route open, and the cohort should enter on a panel. That is a suggestion for a person to weigh. Maya clicks to send it to Lena Ortiz, who owns the VLT-302 cohort.

9

On one list until its owner settles it

It joins the Disagreements list, sent at 8:17 a.m. Four sit there in the last 30 days: one new, one open with Dana Okafor on twice-daily dosing, one explained by Sam Patel on assay length, one settled by Lena on a panel threshold. The rule is on the screen: Lena Ortiz only, because an agent never closes a disagreement. Until she does, the row stays amber on every answer that uses this claim.

10

Kept with its rows, so the next person starts ahead

Maya keeps the answer in the CVT-416 notebook at 8:18 a.m. as "Resistance through MET copy gain: what we know today": 5 claims and 17 rows from 12 sources, each opening at its passage. The disagreement stays attached. The notebook also lists the questions the team asked this week, so Sam Patel's question about bypass alterations does not get worked a second time.

Morning matches, the source library, a refusal to guess, and the rules a library administrator sets

In shortAround the answer sit the screens a research group uses every week: what is new against each person's topics, every source with its model system and count, the evidence table to take away, an honest "no passage found", and the settings that decide how strict an answer is and who may see which program.

Five agents read and check; the scientist keeps the answer and the report's owner settles the disagreement

In shortFive agents index the library, write claims, check rows, count agreement and match new sources. People do the deciding: a scientist decides what to keep, and only the owner of the company's source may explain or settle a disagreement. A person checks every answer that matters. Numbers below are design targets to measure on your own library.

From library agent to watch agent: what each one does

Maya keeps, Lena settles: what stays with people

Roles and separation of duties, an audit trail of who asked what and which passages came back, exports, connections to your own systems and clear ownership of each decision are part of the design. The platform is SOC 2 Type II audited, with information security managed to ISO 27001. If an answer later goes into a regulated document, these controls support your own process; the checking and the responsibility remain with your people.

Papers, patents, reports and notebooks in; claims, rows and kept answers out

Systems a research literature rollout connects to

What slow searches and missed disagreements cost your research group

In shortAlmost nobody publishes what a literature question costs a company, so start from your own log. A few public figures show the scale: the hours a researcher reports, how rarely a search can be repeated, and how often one company's labs and the literature did not match. The costliest case is a study plan written without a fact you already owned.

How to count your own literature hours and missed papers in one week

What an unchecked literature answer costs that no spreadsheet shows

What faster sourced answers return: scientist hours, license costs and the point where the tool pays for itself

In shortPut in your own numbers. The calculator counts search and table-building hours released, against the tool and any added content license. Only the survey's four hours has a source, and it is a 2019 self-report. Hours count as value only if you decide what they are for. The disagreement caught early is the bigger prize, and you should measure it apart.

Search time, repeats, retractions and early disagreements: where literature automation saves

Your scientists, their search hours and your hourly cost

A group of 20 scientists at $90 an hour over 46 weeks, saving 20% of four search hours and 30% of one table hour a week, releases about $91,000 a year: about $89,000 short of a $150,000 tool and a $30,000 license. For a small team, the case rests on missed disagreements and traceable citations. If your log shows little repeated work, say so before buying anything.

Is your research group ready for AI literature search? Eight questions on records, licenses, internal reports and owners

In shortScore your team from 0 to 2 on eight questions. A low score tells you what to fix first, usually your license terms and the rule for who settles a disagreement. A ninth question has no score, and it may be the most useful one.

Reading your score: fix the record, pilot on known questions, or roll out by program

A 0 on rights mapped to sources, question 3, holds you back at any total. Settle that first: it decides what a tool may read.

Questions for any AI literature search vendor: recall, invented citations, rights, your own data and proof

In shortAsk us too. Polished demonstrations on a vendor's own sample show none of what matters. These questions go to the failures Part A described: papers quietly missed, statements a source does not make, rights nobody checked and internal data seen by the wrong team. Ask to see each one on your own questions.

Testing AI literature search on one target: a 10-week pilot with known papers and planted disagreements

In shortPick one target your team knows well. Log a baseline, build a test set from past reviews, settle the rights question, plant a few disagreements, then run the tool beside your scientists for a month. Decide on measured recall, passage accuracy and time saved after people have checked. The target and pass marks should be written before you start.

Which kinds of question to test first

The measures that decide a literature search pilot

Scientific literature search questions: PRISMA-S, Boolean, recall, AI tools, retractions and copyright

In shortShort answers to the questions people type, each of which can be read on its own.