The latest features and improvements to help you plan, review and report your evidence.
The search builder warns when a query finds far more than it should
Search
When the saved PubMed query finds many times as many records as the builder’s own query did, the build now says so. It gives both numbers, names the step that added the most records and, when that was putting the plan’s missing terms back, the plan term that adds the most. Before, only a large drop was reported, so a query that had grown twenty-fold went unmentioned.
The builder no longer puts a broad MeSH heading from the plan into one part of a PubMed query when another part already requires a more specific heading beneath it. PubMed searches a heading together with every heading below it, so the broad heading matched everything the other part did and the two parts stopped narrowing each other: in one review it took a query from about 700 records to nearly 19,000. The term is left out and counted among the plan terms the queries do not carry.
A search links registry records to their publications again
Search
After a search saves its records, each ClinicalTrials.gov record whose registry entry lists its results paper or later analyses is linked to those publications, so screening and extraction treat a trial and its papers as one study. Linking now runs to the end after the search finishes, one record at a time, and tries a record again when ClinicalTrials.gov does not answer. Before, it could stop unseen and leave every registry record of a search unlinked, so screening saw a trial and its papers as unrelated records.
When linking cannot be started, the search says so in its warnings; the records it found are saved either way.
Under Run search, the Search page says how linking is going: how many registry records it has checked so far, and once it has finished, how many links it added and when, or that none was needed. When some records could not be linked, it names their ClinicalTrials.gov numbers, each opening its registry entry; when linking stopped or could not be started, it says so. Link again checks every registry record of the review once more and keeps the links already made.
A review whose registry records have no linking on record, such as one searched before this release, is offered Link registry records. Linking adds records to the review, so it is not offered once screening is locked, and a reviewer whose role cannot run searches is told who can.
Start the search over clears what the Search page says about linking from before it, with the records it hides; Undo brings both back.
A review holds at most 5,000 records
Search
A review now holds at most 5,000 records from its searches, next batches, database searches, citation chasing and reference imports together. Deleted records do not count.
A search that would take a review past 5,000 adds new records up to it, PubMed’s most relevant first, then ClinicalTrials.gov’s, then the other databases’, and says how many it left out; records already in the review take up no room. Before such a search runs, the Search page says how much room the review has left. Once a review holds 5,000 records, Run search, Load next batch, the database searches and reference imports say so and add nothing.
Living-review updates are not held to it, and the studies they add do not count toward it, so a living review keeps taking in new evidence. Records linked in automatically, such as a trial’s results publication, are still added.
A search adds 2,000 PubMed records by default, and at most 3,000
Search
On the Search page, the PubMed record limit now offers 1,500, 2,000 or 3,000 records a search, and is 2,000 unless you choose another. Before, it offered 1,500, 3,000 or 5,000, and 1,500 was the default.
One search adds at most 3,000 PubMed records. A limit you chose is kept, except 5,000, which now adds 3,000 a search. Load next batch adds the records past the limit with the same queries, and the run history says how many remain.
Start the search over
Search
The Search page’s Recent search activity card has a new control, Start the search over. It hides every record in the review, imported ones included, and sets its searches aside, so the next search starts from nothing. A record the next search finds again comes back as a new record, to be screened again.
The review keeps its settings, its search plan and queries, its strategy builds, its team, its sign-offs and its history. Screening decisions, full texts and extracted data on the hidden records are kept, but they no longer count anywhere: not in Found studies, screening, the counts on the sidebar and the Overview, PRISMA or the report.
Before anything changes, you are asked to confirm and told what will disappear: how many records, how many of them are already screened, how many have a full text and how many have extracted data. Once the search is started over, the message that says so offers Undo for 10 seconds, and there is no other way back. Undo is refused once a search has run since, or when some of the hidden records are in the review again, and for the same reasons as starting over.
Starting over needs the same role as changing the search, and the read-only demo does not offer it. It is refused while the review’s protocol is locked (unlock the protocol on Configure first; unlocking is recorded as an amendment), while the search strategy is locked (unlock it first), while a search is still running, while an automated run is in progress, and while a screening, extraction or risk of bias batch you started is running on the review’s records.
On a living review, starting over also switches living updates off, so that an update cannot bring the hidden records back as new ones. Switch them back on in the living review settings on Configure once you have searched again and locked the queries.
Restart from scratch, offered on Configure and on a finished run’s progress banner for a review that runs automatically, is unchanged: it permanently deletes the results of every stage, the searches included, and runs the automated workflow again from a fresh search.
Save and lock checks the protocol before it is locked
Configure
Save and lock now checks the settings before recording them for the report. Anything missing or weak is listed in one place, grouped by section: an outcome with no timepoint, a comparator with no other words, risk of bias on with no instrument, a criterion an abstract cannot fail. Each is marked Missing or Could be better.
Where the assistant can fill a field in safely, the row offers a one-click fix that says exactly what would be written; tick the ones you want and choose Apply and lock. Where only you can decide, Open takes you to the field; where the fix is wording you should steer, Ask the assistant puts the request in the chat.
You can lock the protocol as it is, or cancel and keep the settings saved but not locked. When nothing is missing, the lock goes ahead as before. Fixes you apply are recorded in the chat as the assistant’s changes.
The Configure assistant says when a reply takes too long
Configure
When a reply from the Configure assistant takes far longer than usual, the assistant now stops it and ends the reply with “I could not finish that reply in time. Send the message again.” Before, a slow reply could stay at Thinking… for several minutes and then end without a word, or with only part of an answer. Anything the assistant saved before it stopped stays saved.
Under that reply, Try again sends your message again, so you do not have to type it out. The sentence stays in the conversation, so it is still there when you come back to the page.
While a reply is on its way, the message box shows how long it has taken once it is slower than usual, for example “Thinking… 40 s”, then “Thinking… 1 min”.
When the assistant runs into a problem, the message you see says what to do next, for example to wait a minute before sending your message again, instead of a technical error.
Restart from scratch checks the review can run before it deletes anything
Automated workflow
Restart from scratch now refuses, deleting nothing, when the automated workflow could not run the review again, and says what to finish first, such as generating search queries and locking the search strategy. Before, it cleared every result and then the run stopped at its first check, leaving an empty review.
On Configure and on a finished run’s progress banner, Restart from scratch is shown only for a review that has run automatically. While that review is not ready to run again, the button is switched off and its hint gives the reason. A review worked one stage at a time no longer shows it.
When a restart is refused or fails, the message gives the reason in plain words, for example that automation is not available or that the previous results could not be fully cleared, instead of always asking you to try again.
A search runs only queries that were built or saved
Search
A search runs a query only when it was built from a search plan, edited and saved by you, or made from the search plan for a database with nothing saved. A review with no search plan still shows a query drawn from your review question, but as a starting point marked “does not run”: Run search stays off and says what to do, with Build the search strategy and Edit and save the query. Before, that drawn query could run with one click; on a review whose questions were still placeholders it searched every word of them, and could find hundreds of thousands of records, almost none of them on topic.
Once you edit the drawn query, Run search asks to save it first and then runs it. Saving it unchanged is refused, and so is a search of it from an automated run. Queries saved unchanged from the drawn query before today do not run either: the Search page says so, and building the search strategy replaces them. They cannot be locked as they are, since a locked search runs whatever was locked.
Before a search runs, the check of each database’s records now gives every database’s total against what one search adds. When one search would leave out more than nine in ten of a database’s records, the search is held back: a search that broad is rarely on topic, so the page suggests narrowing the query or building the search strategy, and runs it only once you tick that you understand most records will not be added.
The search builder no longer narrows a search that finds what it should
Search
The AI search builder judges a query by how much of what it finds is on your review’s topic, not by how much looks eligible. A thorough review search finds mostly records that screening will exclude, such as reviews and studies of other designs, so only a few in a hundred look eligible. Before, the builder read that as a poor query, added a randomised-trial filter or removed setting and intervention terms after it had finished, and could lose the trials your review is about.
The builder still makes a query’s population terms more specific when most of what it finds is about another population, and changes nothing else in the query when it does. Study design, comparators and outcomes are left to screening, as your protocol says.
On the Search page, Relevance shows how much of each source’s sample is on topic, for example “57% on topic”, and names the population or intervention most off-topic records miss when that share is low.
When the saved PubMed query finds far fewer records than the builder’s own query did, the warning now says which step cut it: one of the plan’s filters, with how many records it keeps on its own, or an automatic change to the query.
The out-of-date search notice takes you to Run search
Search
On the Search page, the notice that earlier search results are out of date now takes you to Run search, beside the queries a search will send, and Run search asks about unsaved edits and record limits as it always does. Before, the notice’s Re-run search only rebuilt the queries, which finds no records, so the results stayed out of date.
Searching one source on its own, such as OpenAlex from its panel, brings only that source’s records up to date. The notice stays until a search covers the other sources again; before, a search of one source cleared it although PubMed and ClinicalTrials.gov were not searched again.
The Configure assistant can set up everything a review needs
Configure
The Configure assistant can switch living review updates on or off and set how often they run: daily, weekly or monthly. As in the Living review dialog, switching them on needs locked search queries, and updates start once the protocol is locked.
The assistant keeps your screening criteria in your words. A criterion written as a statement, such as “Adults with resectable stage IB–IIIA lung cancer”, is saved as written, and screening asks it as a yes-or-no question, as it always has for criteria entered in the settings panel. Before, the assistant rewrote statements as questions of its own.
When you ask for an outcome’s timepoints, preferred timepoint, arm labels, units, how missing values may be derived, which arms to compare in a multi-arm trial, a fallback timepoint or other names for the outcome, the assistant saves each as its own setting, which extraction uses. Before, it could write them into the outcome’s reporting notes, where extraction does not read them as settings.
The assistant knows the exact name of every setting, so a change to living review updates or extraction defaults is saved the first time rather than failing several times.
A new review starts with a placeholder outcome. When you give the assistant your outcomes, it replaces the placeholder without asking you to confirm removing it.
Building queries no longer marks your search out of date
Search
Building or refining search queries no longer shows “These results are out of date” on the Search page. A build finds no records, so it has no results to be out of date; before, each build counted as an earlier run with different settings, even in a review that had not searched yet. A search that ran before you changed the plan or the queries is still marked out of date until you search again.
The search builder keeps specific MeSH headings
Search
The AI search builder no longer treats every MeSH heading as a broad one. Since April it read each heading it checked as if it sat at the top of the MeSH tree, so it was steered away from specific headings such as Programmed Cell Death 1 Receptor, and its notes on a build could call them broad. Now only a heading at the top of a tree, such as Neoplasms, counts as broad. A strategy built before this change may have left such headings out: refine or rebuild it to use them. The Configure assistant’s review of a search uses the same check.
A MeSH heading is now preferred over a supplementary concept, such as a single compound, whose name merely contains the term being checked. Before, “PD-1” could be accepted as an unrelated herbal preparation; now the builder is told it is not a heading and is pointed to Programmed Cell Death 1 Receptor.
Each search-plan filter says where it applies
Search
The search plan’s filters in the Configure settings panel each say where they apply: searched in a named source, or left to screening.
The ClinicalTrials.gov recruitment filter, now named “Only trials that finished recruiting”, applies to every registry query, whoever builds it: completed trials, terminated ones, and trials still active but no longer recruiting. Before, the AI builder chose its own statuses whatever the plan said, and the basic builder kept completed trials only, which left out large trials still in follow-up.
Study design, species, trial phase and language (except in OpenAlex) are left to screening by every builder, and the AI builder is told so. The basic builder used to add a randomised-trial filter that the AI builder did not; neither does now. A qualitative review keeps its qualitative study filter.
“With results only” is not applied to the ClinicalTrials.gov search, as before: a registry record without posted results can still lead to the trial’s published report. The panel now says it is left to screening.
With “Include outcomes in the search queries” on, the AI builder now sees the outcome terms and searches them, as the basic builder does. Before, it never saw them, and each such build warned that all the outcome terms were missing.
The AI builder’s explanation of a strategy says which filters it applied and which are left to screening, instead of saying the review does not restrict them.
Built queries keep the meaning of your search plan
Search
Extra terms are applied as one requirement: a record must mention at least one of them. Before, a term that also appeared elsewhere in the query was left out of the requirement, so the search demanded the remaining terms alone and missed relevant studies.
An exclusion that is part of one of your own population or intervention terms, such as “small cell” inside “non-small cell lung cancer”, is no longer added to the search, where it removed the very studies the search is for. The search plan shows a warning that names it; exclude those studies at screening instead.
Stage, grade and class terms in the plan, such as “stage II”, are no longer added to the search as alternatives to the condition, where they brought in studies of other conditions at that stage. The condition’s own terms already find every study they describe.
If the saved PubMed query finds far fewer records than the builder’s own query did before the plan’s date range, extra terms and exclusions were added, the build says so and gives both numbers.
A ClinicalTrials.gov query is no longer reported as leaving out a plan term that it carries in the singular or the plural.
A search runs the queries it shows
Search
A search now runs the saved query for each database, exactly as the Search page shows it, and records it with the results. Before, a search on a strategy that was not locked rebuilt the PubMed and ClinicalTrials.gov queries from the search plan, ran those instead and saved them over the saved queries, so a query you had edited, or one the AI builder had refined, could be replaced without a word.
Checking the queries, running a search and loading the next batch never change the saved queries. Load next batch continues the same search with the same queries, and Free full text only applies to the PubMed query the search runs.
Edits to a query in the search strategy are now saved when you choose Save, and Discard puts back the saved queries. Until then the strategy says Unsaved changes, Test queries checks what you typed without saving it, and leaving the page asks first. If the queries changed meanwhile, in another tab or by a rebuild, the save is refused and you are asked to reload. Before, each change was saved as you typed.
The query box now shows what a search sends to PubMed and ClinicalTrials.gov, and where each query comes from. Typing in it no longer changes a search: edit a query in the search strategy instead. Before, a change in the box ran for that one search and was not saved. A review with no search plan and no saved queries starts from a query drawn from the review question, which you edit in the search strategy; the first search asks you to save it as the review’s search queries. Where nothing is saved but there is a search plan, Save as the search queries keeps the plan’s terms as the review’s queries.
Changing the search plan no longer removes the saved queries. They stay, marked out of date with the reason, and a search still runs them as they are, with a warning. As before, they cannot be locked until they are brought up to date: refine them, or edit and save them.
Run search with unsaved changes in the search strategy asks first: save them and run, or run the saved queries as they are. If the saved queries changed after the page showed them, in another tab or by a rebuild, the search does not run, and the page shows the queries as they are now and says what changed.
Before a search runs, the Search page counts each database’s records and says when one has more than a search adds: run anyway, raise the PubMed record limit and run, or cancel. The PubMed record limit, 1,500, 3,000 or 5,000 records, is now set beside Free full text only and is a setting of its own: changing it does not change what the AI aims for when it builds or refines the queries.
A search that stops at a limit now says so after a reload too: the run history and Found studies read, for example, “Incomplete: 1,500 of 7,864 PubMed records added. 6,364 remain; the next batch adds 1,500 of them.”, and Load next batch still continues it. Before, a reload lost the next batch and the run history said Done.
Suggestions in the Suggest drawer are copied, not applied: paste one into the PubMed query in the search strategy, then save.
Going back to the Search page from another page shows it as you left it. Before, the page loaded again from the start.
The search strategy card writes its dates as the approvals history does, for example “Built 20 Oct 2026, 14:02”, in your own time zone.
An automated run never replaces search queries a person changed: it stops and asks you to review them on the Search page.
Approving a proposal to widen a search that found nothing saves the widened queries as your edit, and the next search runs them. While the search strategy is locked, approving one changes nothing and says why: to change locked queries, unlock the strategy on the Search page. Before, the approval failed.
Refine queries now starts from your saved queries, your note on what should change, the open findings of the newest search review, what the last search found and your known papers, instead of the search plan alone. A note sent with a Refine that could not finish is given to the next one. Rebuild plan and queries from scratch still starts from the review question only.
Search review findings now appear on the Search page, under the plan and above the queries. Dismiss one you disagree with, and Undo puts it back; after a refine they are listed as considered, which means the build saw them, not that it fixed them.
Configure
Saving the review’s settings on Configure no longer saves the page’s copy of the search strategy, which could put back queries, a plan or search settings that had changed since the page was opened.
Switching the review’s framework asks first when the switch would leave the search queries out of date.
The Configure assistant saves search plans but no longer builds search queries: build and refine them on the Search page. It still reads, counts and reviews them, and says when a plan change has left them out of date.
Regenerate plan on Configure asks first, then saves a new search plan drafted from your saved review question. It builds no search queries: the saved ones stay as they are, out of date until you refine them on the Search page. While the review methods have changes not saved yet, it is off: save them first.
Configure’s search section says when the saved search queries are out of date, and why, with a link to the Search page; Lock queries there says why it is off.
A change the Configure assistant offered before it stopped building search queries, such as Rebuild the search queries?, now says once you answer that the assistant no longer makes it, and points to the Search page. It no longer reads as a change that did not finish.
Restoring an earlier save of the review’s settings asks first when it would change the search queries, and says whether it brings back that save’s queries or clears them.
A search review by the Configure assistant is now kept with the review, not only in the chat, so its findings reach the Search page and the next Refine.
Configure reads in the order of your review
Configure
One save per thing. Most settings save as soon as you change them. The review methods (the title, framework, questions, screening policy, extraction approach, appraisal instruments and certainty) save together from a bar at the foot of the settings that shows only while one of them has changed, and the screening terms keep their own Save. Save and lock and Continue to search are in the page header; the settings’ own save button and header are gone.
Sections by stage. The settings read in the order of the review: Question, Screening terms, Screening, Search, Outcomes, Extraction, Appraisal, Synthesis, then the record (the version, why these settings, save history, export and templates). Each method sits in the stage it decides, and Full-text decisions opens the same settings as the full-text screening page.
Outcome cards fold their details. Each outcome shows its type and measure in words and its first few synonyms; everything else is under Details, in words.
Readiness is one notice. What is still needed comes first, as one list with Ask the assistant to fix these; when everything is set, one line says so. A locked protocol says so once, under the page header.
Clearer next steps for out-of-date CASP appraisals
Review settings
When CASP appraisals are out of date after a settings change, the notice on the page now asks you to check each affected study’s answers. Before, it offered Re-run, which started risk of bias rather than CASP, so the appraisals and the notice stayed as they were. Out-of-date risk-of-bias results still offer Re-run.
The Configure assistant asks before important changes
Configure
The Configure assistant now asks before it replaces the search plan, rebuilds the search queries, switches the framework or another of the review’s definitions, or removes an endpoint, a synthesis question or a whole setting. A card in the conversation says what would change, line by line for a search plan, and nothing changes until you answer. Small additions, such as a new endpoint or a title, are still made straight away.
Answer with the card’s two buttons, and add a note for the assistant if you like. The card then shows what you chose and when, and the reply continues below it. Until you answer, you cannot send a new message.
Answering in the tab where the card appeared continues that reply. Before, the reply showed twice until you reloaded the page, and the first copy still said it was making the change.
Tell Refine what should change
Search
Refine queries takes a note. Write what should change beside the button, such as a term to drop or a kind of study to bring in, and the builder starts from the saved queries and follows it. The last-build card says when a build followed your note.
The Configure assistant’s search review is kept with the review. When you next press Refine, the builder also starts from the problems the assistant found, and marks them as seen.
A query you edited by hand is kept when the AI builder cannot finish. Before, only a strategy the AI builder had saved was protected from the simpler fallback build.
The builder’s relevance check now reads a sample of 100 records, or one in ten of everything the query finds if that is more, drawn across the whole result set. Before, it read the first 15. The check still reads titles.
The Configure settings panel says each thing once
Configure
The settings panel beside the Configure assistant no longer repeats the review title, the screening questions and their synonyms below the Review settings card. Each is edited in the card: the questions under Screening questions and the synonyms under Screening terms.
A screening question is marked required when it is needed before you continue, and the list below the card, now headed “Needed before you continue”, agrees with that mark. In a PICO review the comparator question is needed to continue although screening treats the criterion as optional; before, the two disagreed on the same panel.
Screening terms sit last in the Review settings card, in their own box, and say that they save with their own button. Synthesis questions moved up to follow the screening questions.
The search section is headed “Search strategy”, and dates in the panel are written one way.
Resolve and Reopen work on comment threads
Collaboration
Resolve and Reopen in a comment thread’s menu now do what they say. Before, choosing either showed “Could not update the thread” and changed nothing. A resolved thread is marked Resolved and stops taking replies until it is reopened.
Long messages to the Configure assistant
Configure
Past 10,000 characters, the Configure assistant suggests attaching long text such as a protocol as a PDF, which it reads section by section instead of carrying the whole text through the conversation. Past 32,000 characters a message is too long to send; attach it, or send it in parts. The message’s length shows once it is long.
Replies show their steps in the order the assistant took them, and repeated steps of one kind fold into one line with a count. Messages the assistant could not accept say why in plain words.
A search strategy the AI builder could not finish is saved
Search
If the AI search builder cannot finish a review’s first search strategy, the simpler strategy built in its place is now saved, and the Search page says so. Before, the build ended with “Failed to generate queries.” and nothing was saved. Rebuild once the AI builder recovers: a strategy the AI builder saved is never replaced by a simpler one.
The strategy’s status shows such a strategy as a warning, and the message after the build no longer reads as a success.
A new search plan no longer keeps the queries built for the old one
Search strategy
Saving or regenerating a search plan now removes the queries built from the previous plan, so Search no longer shows them and they cannot be locked as your strategy. Build the queries again for the new plan.
The plan you save is the plan that is kept: a field you cleared, such as the date range, no longer comes back from the previous plan.
Rebuilding the queries for fewer databases removes the queries for the databases you dropped, and notes from an earlier build no longer stay with the new queries.
If the plan changes while the queries are being built, the build stops without saving and asks you to build again, instead of putting back the plan it started from.
A locked protocol keeps its search strategy and screening settings as signed off
Sign-off
Once a review’s protocol is locked, its search plan and queries can no longer change, even when the search strategy itself was never locked. Building or rebuilding the strategy, populating the plan and editing it on the Search page say the protocol is locked and change nothing, and so does the configuration assistant. Unlock the protocol on Versions to amend them.
A change is checked against the lock when it is saved, not only when you start it: queries still being built, or an assistant still working, when a sign-off is approved are not saved.
An outcome’s pooling model and network settings, the treatment grouping and the living-review schedule can still be changed on a locked review, and the automated workflow can still be started, as before.
Building the search strategy and populating its plan now need a lead reviewer or an admin, as editing the plan already did. Anyone else is told why, before anything runs.
Screening
The confidence threshold you choose in Auto-screen now applies to the run you start and no longer changes the review’s settings, so it works the same way on a locked review. The next time you open Auto-screen it starts from the review’s own threshold.
When a review’s outcome definitions need a repair before screening can start, Auto-screen now says what is in the way, instead of changing them itself: a locked protocol, which is unlocked on Versions, or a role that cannot change them, in which case a lead reviewer or an admin can start the run.
Drafting a search plan keeps changes saved meanwhile
Search
Changes saved to a review while its search plan is being drafted, such as a new title, an edited question or the search settings, are now kept. Before, populating the plan could undo them when it finished, and saving the plan on the Search page could undo a change saved a moment earlier.
If the search strategy is locked while a plan is being drafted, the draft is not saved and you are told why. A plan saved in the meantime is kept, unless you chose to rebuild the plan from scratch.
Clearer study-arm source review
Extraction
Cite shared study-arm facts once across outcomes. Unchanged fields keep their source confirmations when you save.
See every open validation requirement with a link to the field. Quote review names what you are checking and how many fields your answer covers.
Add quotes despite common spacing and line-break differences, while keeping the paper’s original wording. Review answers offer Undo.
The Verify lines and the open card share the pane by a handle
Extraction
On the Verify and Form pane, the open value’s card now sits under the lines behind a handle you drag, and the pane remembers where you left it. Before, the card took a fixed share of the pane, and one with several checks left the lines a strip. With no value open, the card’s one-line prompt takes only its own height.
The study’s notes above the lines, such as how many quoted sentences could not be found in the documents they cite, are one line; What this means opens what each means for the values.
The Overview answers “where is this review and what next” once: a stage strip that is also the study flow, and every number shown once
Overview
The review card now carries a strip of the six stages, Search to Report, each Done, In progress or Not started, with a Next marker on the stage the next step belongs to. It replaces the percentage, the phase chip and the Progress tile, which could disagree with each other; the summary line reads “4 of 6 stages done · Next: Analysis”.
The strip is also the study flow: Search, Screening and Full text show their count, their share of identified and what dropped out beneath them (duplicates removed, excluded at title and abstract, excluded after full text with its reasons a click away, PDFs still missing). Full text ends in the number of studies included, and Export PRISMA sits on the summary line. The separate funnel and PRISMA cards are gone.
Missing PDFs are counted the same way everywhere: the strip, Review health and the next step now agree. Before, the next step could say PDFs were missing when every report had been retrieved.
Review health is amber only where a number asks for something: conflicts to rule on, PDFs missing, values not yet validated, proposals waiting. Zero reads “None open” in the neutral tone, and Proposals takes the place of the Progress tile.
Next steps lists its items in the order that unblocks the most work and says so; when nothing waits it says the review is up to date and, for a living review, when the next update runs. Export in the header is now Open report, and it waits until a report exists.
Recent activity shows the last five rows, groups identical runs (“3 screening runs completed”), badges only a failed, cancelled, queued or running run, and opens the stage each row belongs to; Show all activity expands the rest by day.
The living-review card names the next and last update and how many new studies are waiting, labels each week’s bar with its count, and opens the living review and its recipes. Review methods is one line at the bottom with Show details, and the assistant is one short prompt with three suggestions, a past conversation scrolling inside it, with the full assistant on the Analysis page.
Network meta-analysis keeps distinct interventions apart, never combines a study’s arms without you, and shows what a grouping change would do before you save it
Network meta-analysis
Arms are now grouped only when they are the same intervention: the same treatment, dose, dosing schedule, regimen and route, given with the same co-interventions. Two dose-escalation schedules of one drug, or a drug given with and without laser, are separate treatment groups. A new Grouping control on Network setup names the grouping in use and lets you choose another: by treatment, by treatment and dose, biosimilars with their reference product, ignoring co-interventions, or one of your own. Every result records the grouping it was fitted under.
Two arms of one study are never combined into one unless you say they are the same intervention. When a grouping would put them together, the study is left out of that outcome and named until you choose Combine these arms, with your reason, or move one arm to another group. The new Combine with… action on an arm does the same from the Treatment groups list.
Approving a group, merging groups or moving an arm now shows, before it is saved, what it would do to the network: the number of groups and separate parts, the studies that would be left out, and the arms of one study that would be combined and how.
Preparing the treatment-group review again never changes a grouping you made or confirmed. Where an arm’s wording now reads as another group, the arm says so and keeps its group; Reset automatic groupings is the one way to send groupings nobody has checked back to review. A confirmation lapses when the arm’s treatment, dose, schedule or co-interventions are corrected afterwards.
The study-arm editor records each arm’s dose schedule and highest dose, what it was given alongside its treatment, and what kind of comparator a control arm is.
A network in separate parts now says which treatments are in each part and which part holds the reference treatment, and that results can be computed within a part but not between parts.
Results run again may differ from results saved before this change: arms that used to be combined automatically are now kept apart or wait for your decision, and a combined arm’s standard deviation no longer absorbs the difference between the arms it combines. Saved results are unchanged; each records the data it was fitted on.
Extraction’s row card leads with what blocks a value, says which quote a check is on, fills a quoted value in one click, and answers a quote at once
Extraction
A value’s card in Verify now shows what blocks it first: the checks on the value with their answers, then the quoted sentence. When the automatic check has already found the quote in the paper, the question about it steps back behind the checks.
A value quoted from the paper with nothing stored for it offers Use the quoted value: the empty fields take the numbers the sentence gives, saved with the study like any edit. Before, the numbers had to be typed from the quote.
Right and Wrong answer for every field the sentence is quoted for in one step, and the quote marks move as soon as you answer. Before, each field was answered in turn and a three-field quote took several seconds.
A value with several quotes behind it, such as an effect, its interval and its p-value quoted from different sentences, is one line in the Verify table, so every one of its quotes opened the same card and the same Needs you. The card now says, under Needs you, which quote the check is on and how the quote in hand stands: “The check is on Effect · 95% CI (pin 1), not on this quote; this quote was found in the paper.” The pins in the source keep their colours: red where a check stands, green where the quote is confirmed.
History keeps dual review blinded until both reviewers have decided, and a single reviewer can finalise risk of bias
Risk of bias
A single reviewer can now finalise risk-of-bias assessments from the Review queue. Before, the queue opened only when two reviewers screened the review, so with one reviewer an assessment stayed a draft and never counted towards certainty ratings. A review that agrees with the AI’s proposal finalises it; one that disagrees sends it to Conflicts, where, in a review with one reviewer, you rule on it yourself and give your reason.
An assessment’s panel now links the Review queue in every review, and GRADE’s note on an outcome with no confirmed assessment says to review them in the Review queue.
Sign-off status
In dual review, History, its counts and the exported approval record now show a reviewer’s vote on a study only to that reviewer until both reviewers have decided it, as the screening page does. Before, anyone who could read History saw a vote as soon as it was cast. The AI reviewer’s vote is withheld the same way unless the review shows the AI’s verdict first.
More ways to settle full-text criteria
Screening
The full-text screening settings, under Advanced, can let plain facts (the population, and whether a study is original research that reports results) count as met on one passage, while every other criterion keeps the evidence you chose. Ruling a study out still takes that evidence, and the check before an inclusion is confirmed for you also asks whether the population fits.
Two new options, both off unless you turn them on: propose excluding a study whose full text, read in full, never mentions the intervention or exposure (always for you to confirm), and take one more look at the passages that mention a criterion still open once the reading has ended.
The configuration assistant can record words that studies use in another sense, drafted from your protocol, and full-text screening takes them into account when it reads and checks a full text.
When the intervention or exposure counts as met on quotations that use none of its words, the study now waits for you, in reviews that use any of these options or words with another meaning. Reviews that use none of them decide studies by the same rules as before; the one change is that a study whose full text took too long to read is tried once more in the same run instead of waiting for the next one.
A file a study already has is not added twice
Full texts
Uploading a file that a study already has, on the Full text page or in a bulk upload, no longer adds a second copy: you are told the study already has it, and bulk uploads count those files apart from the ones added. A paper saved with the browser companion that the study already has is not stored twice either. Your own upload is always kept, even when the same file was fetched for the study automatically, and the same file uploaded for two different studies is kept for each.
Choose the criteria full-text screening requires
Screening
The full-text screening settings, under Advanced, can now choose which of a review’s criteria must be met at full text, including questions kept outside its framework such as a comparison or a study design. A note lists any criteria full-text screening does not use, with a button to use them, and Configure shows the same note. Reviews that change nothing screen exactly as before.
A review that requires the study itself to compare the groups can say what counts as the comparison, and the comparison then counts only where the study states a result for the groups compared.
A more thorough check before a decision is confirmed for you can also look at the population and for signs a study should not be included. When a decision waits for you because of these settings, the paper’s criteria panel says why, beginning “Held for you”.
Studies whose comparison looks incidental at the abstract stage can be sent to full text instead of being excluded, and studies already excluded for that reason can be reopened: the action counts them first, asks why, and keeps the reason in each study’s history.
Finalise approves only the answers you saw, and Confirm groupings names both kinds of arm
Critical appraisal (CASP)
If someone else changes an appraisal while you have it open, Finalise no longer approves their newer answers without you seeing them, even when you made no edits yourself. It says "Someone changed this since you opened it." with Reload, which shows their answers so you can check them before you finalise.
Treatment groups
The button in the warning above the groups now reads Confirm groupings, because it confirms arms grouped automatically (Automatic only) and arms approved but not confirmed (Approved). Its dialog names each kind it confirms.
Screening asks why whenever a decision needs it, and a cancelled vote is kept
Screening
A decision against the AI now always asks why first, including on a record the AI filed as Unsure because it was not confident enough, and on one whose latest screening run failed. Before, some of these decisions could not be saved at all.
A bulk decision is now saved as one: if it cannot be saved, none of it is, and your reason is still kept only with the records that go against the AI.
Full-text screening
With two reviewers, cancelling the question about why your vote goes against the AI’s earlier proposal now keeps your vote. The study says a reason is required until one is given, and choosing the same decision again asks for it. The vote itself can no longer be changed.
Living reviews
The inbox no longer asks why for a selection it cannot change because full-text screening has already decided some of it: it says so straight away. If saving is refused for want of a reason, the list now reloads, so trying again asks for it.
Critical appraisal never saves over a newer change, and treatment-group filters match their badges
Critical appraisal (CASP)
If someone else saves an appraisal while you have it open, even the first answers for a study nobody had appraised, your Save or Finalise no longer writes over their work. It says "Someone changed this since you opened it." with Reload, which shows their newer answers. The panel does not update by itself while it is open.
When Finalise saves your answers but finalising is then refused, for example because the review is locked, the list now shows the answers you saved.
A finalised appraisal whose finaliser was not recorded now reads "person not recorded", as GRADE does, instead of naming a reviewer.
Treatment groups
The Needs you, Automatic only and Confirmed filters now list exactly the groups whose arms show that badge. An arm awaiting approval puts its group under Needs you, and a grouping a person approved, which reads Approved, is no longer listed under Automatic only.
Moving an arm into a suggested group asks you to confirm in the same dialog as the other approvals.
Arms that were approved but not confirmed, which read Approved, must be confirmed before the network runs, just like arms grouped automatically. The warning above the groups now counts both kinds, and a new Approved filter lists the groups that have such arms.
Confirm full-text inclusions and exclusions automatically, separately
Screening
Full-text screening can now confirm inclusions and exclusions automatically separately: for example, let inclusions the evidence settles go through while a person confirms every exclusion, or the reverse. The full-text screening settings have one switch for each kind of decision; the switch on the Configure page is on while either is, and turning it on or off sets both. Both stay off in dual review.
Generated Methods sections now say which kind of decision was confirmed automatically, when only one was, and state separately any decision a rule settled from the record itself (for example a trial registry record that registers a single group), which is confirmed without a check of quoted passages. Reviews that set nothing new read as before.
The browser companion saves the page you are on, and teams can share the work
Browser companion
The next five changes arrive with the companion’s next update, which Chrome installs on its own.
Save this page, at the top of the companion’s Reviews tab, saves the article in your current tab into whichever of your reviews is waiting for it, on any journal site, including articles reached through your library’s sign-in. If it cannot, it says why.
When the companion works through a review on its own, every paper already in your reference library is saved from there before any journal page is opened.
Articles reached through your library’s sign-in are recognised as the journal behind them, and the PDF is fetched through the same sign-in.
On well-known publisher sites, the capture button appears on any paper one of your reviews still needs, without opening that review in the companion first, and disappears as soon as the paper is saved.
When the companion works through a review on its own, it now fetches papers several times faster: up to five journals at once, up to three papers a minute from any one journal, and shorter breaks between sittings.
On a shared review, anyone who can add its full texts can allow the companion to work through it with their own access. The list of what the companion fetched shows whose companion fetched each paper.
Papers that have a PubMed ID but no DOI and no free copy can now be fetched from the publisher page PubMed links to.
Screening asks why when you go against the AI or take a decision back
Screening
Including or excluding a record against the AI’s suggestion now asks why before it is saved, and so does Unsure on a record that was already included or excluded. The dialog says what the AI suggested, or whose decision you are taking back. Your reason is kept in the record’s history. Confirming the AI, deciding a record the AI has not screened, and Unsure on an undecided record save at once, as before.
Y, N and U open the same dialog with the cursor in the field: Ctrl Enter (⌘ Enter on a Mac) saves, Esc cancels and leaves the record undecided. In Focus mode, a note you have written starts the reason.
A bulk decision asks once for the whole selection, and only when some of it goes against the AI or takes a decision back. It counts which ones, and your reason is kept with each of them. A selection that agrees with the AI saves at once.
Send to full text no longer counts as going against the AI, so it asks nothing.
When a decision cannot be saved, the reason now shows under the decision bar, or under the bulk bar, with the next step where there is one, instead of a message that disappears.
On a record you have already decided, the decision bar shows that decision pressed, and pressing it again (or its key) changes nothing. The other two decisions stay a press away. A record the AI decided and no one has confirmed shows nothing pressed, since the decision that agrees with the AI confirms it. The same at full text.
Full-text screening
Excluding or including a study against the AI’s proposal, and Needs discussion on a study the AI proposed or that was already decided, ask why in the same way. An exclusion’s dialog starts with the exclusion reason you chose, which you can save as it stands or add to.
A bulk Confirm include or Exclude asks once for the selection, starting with the reason you chose for the batch.
With two reviewers, the vote that agrees with the other reviewer against a proposal the AI made before the review moved to two reviewers asks why. The other votes ask nothing.
A decided study now shows under its title who decided it, when, and the reason they gave.
Withdrawing a value, undoing a rejection and rejecting in the review queue ask why
Data extraction
Each outcome in the form now shows where it stands: Needs you, Validated or Rejected. Validate still saves at once, with Validate and next beside it. A validated outcome shows who validated it and when, with Withdraw beside it.
Withdrawing an outcome’s validation now asks why. Your reason is kept in the outcome’s history.
A rejected outcome now has an Undo rejection button beside it, instead of a hint you only saw on hover. Undoing a rejection asks why, and the values go back to Needs you, unvalidated, ready to check again.
Once a study is validated, its footer shows Validated with Withdraw. Withdrawing the whole study asks for one reason, which is kept with each value it withdraws.
In the review queue, Discard is now called Reject. Rejecting an outcome asks why before anything is recorded, one reason per outcome. The reason shows on the row, with Change to edit it.
In the study-arm editor for network reviews, Validate asks you to confirm, and Withdraw asks why.
When one of these changes is refused, for example because the review is locked, the reason now shows under the control, with what to do next, instead of a passing message.
Network certainty and treatment groups say where they stand
Network certainty
Each certainty review reads Draft or Finalised, in the same words as the rest of the review, and once finalised it says who finalised it and when.
Finalise appears once every domain is rated, none is left on the earlier scale and your ratings are saved, beside a line saying that finalising counts this outcome towards the Analysis freeze milestone. It still asks you to confirm, and says that the ratings cannot be edited afterwards and that the outcome is locked if your role may lock analyses.
If finalising is refused, for example because the review is locked, the reason now shows under the Finalise button, with what to do next.
Treatment groups
Each study arm now says where its grouping stands: Needs you, Automatic only for an arm grouped by its name that no one has checked yet, Confirmed once a reviewer has checked it, Approved when it was approved with nobody recorded, or Rejected when it is left out. A group that needs a look reads Needs you, and an approved comparison reads Approved.
Approve and Confirm now ask you to confirm first, and say what follows.
Leave out now asks why in a dialog, and a reason is required. It is kept with the arm’s grouping.
The Word report now ends with the review’s sign-off record
Report
Every Word report now ends with a Validation and sign-off record, as it stood when the report was generated: the milestones, with who approved each and when, then each approval point under its stage, with how many items a person approved, how many were decided automatically and how many still need a person, and its status, in the words of the Sign-off status page. Numbered notes explain labels such as AI only.
If the sign-off status cannot be read when the report is made, the appendix says so in one line instead.
The record names people as the Sign-off status page does.
Conflict rulings ask why when they go against a reviewer
Conflicts
A lead reviewer’s ruling is now a set of choices you can see at once, instead of a drop-down. For risk of bias: Keep the AI assessment, Keep the reviewer’s assessment, Set aside, or Decide later. For screening and extracted values: Include, Exclude or Unsure.
Resolve now asks before it records anything. Keeping the AI’s assessment, including or excluding a study, and including a set of values ask you to confirm. Keeping the reviewer’s assessment instead, setting an assessment aside, sending a study back as Unsure, and excluding extracted values ask why, and your reason is kept in the record. Setting an assessment aside cannot be undone. Decide later, and Unsure on extracted values, decide nothing: the conflict stays open for another ruling.
The optional reasoning box has gone: the reason is asked in the dialog when the ruling needs one. Root cause and its note stay on the card.
If a ruling is refused, for example because the review is locked, the reason now shows under the Resolve button, with what to do next.
The inbox asks why when you go against the AI’s screening
Living review
In the living-review inbox, Accept, Reject and Import data now ask why when any of the studies you picked goes against the AI’s screening, or was not screened by it yet, for example rejecting a study the AI included or accepting one it was unsure about. One reason covers the whole batch, and for a batch the dialog says how many of the studies it applies to. The reason is kept in each of those studies’ history. A batch that agrees with the AI saves at once, as before.
When the inbox cannot make a change, the reason now shows under its buttons instead of in a passing message, with Reload where reading the list again helps.
Proposals
Each proposal on the Proposals page is now decided on its own card: Approve asks you to confirm, and Turn down asks why. The bar with Submit at the foot of the page is gone.
Each proposal reads Waiting, with when the AI proposed it. The fixes the automated review applied by itself read Approved · Applied by policy.
Risk of bias reads Finalised, and says what finalises an assessment
Risk of bias
An assessment’s status now uses the same words as every other approval: Draft for the AI’s proposal, In conflict, Finalised (what used to read Validated, or Adjudicated when a ruling kept one side) and Set aside (an assessment a ruling left out of certainty ratings). An assessment the AI assesses again after its data changed goes back to Draft.
The foot of an assessment’s panel says what finalises it: a review that agrees, with a link to the Review queue when two reviewers appraise the review, or a lead reviewer’s ruling on Conflicts for one in conflict. It names the Analysis freeze milestone the assessment counts towards, and a finalised one shows who finalised it and when.
The Review queue now says an assessment you agreed with is finalised.
Validate several values at once on the Verify pane
Data extraction
On the Verify pane, tick the values you want, or tick “not yet validated” beside the filters to take every one shown, and a bar at the foot of the pane says how many can be validated and which are held back and why, such as a value flagged by a check or one with a missing field. Validate them in one step after a confirmation that names the number; the ones held back stay for you to check one by one.
The open value’s card now shows its state, Needs you, Ready, Validated or Rejected, with who validated it and when, and Validate or Withdraw at its foot.
Withdrawing a validation, reopening an answered check, marking a quote as not in the paper and withdrawing an answer about a quote now ask why. Your reason is kept in the value’s history. If the change is refused, for example because the review was locked meanwhile, the card says why, with what to do next.
Reopening a GRADE, CERQual or CASP rating asks why
GRADE
Each rating’s status now reads Draft, In review or Finalised, the same words the Sign-off status page uses.
A draft rating is finalised after review: select Start review, or change a factor, and Finalise appears. Finalise asks you to confirm, and the foot of the panel says the outcome then counts towards the Analysis freeze milestone. The foot stays in view while you scroll the factors.
A finalised rating shows who finalised it and when. Reopen now asks why you are reopening it; your reason is kept in the outcome’s history, and the rating goes back to Draft.
When a rating cannot be changed, for example because the review is locked or someone changed the rating meanwhile, the panel says so under its buttons, with the next step.
CERQual
Each finding’s status now reads Draft, In review or Finalised. Edit any rating to start the review; Finalise then asks you to confirm.
A finalised finding shows who finalised it and when. Reopen now asks why you are reopening it; your reason is kept in the finding’s history, and the finding goes back to Draft.
Critical appraisal (CASP)
Reopening a finalised appraisal now asks why. Your reason is kept in the appraisal’s history, and the appraisal goes back to In review until it is finalised again.
Each appraisal’s state reads Draft (the AI’s appraisal, not yet checked by a person), In review or Finalised, in the table and in the study’s panel. What was shown as In progress is now In review.
Until an appraisal is finalised, the foot of the panel says that finalising counts it towards the Analysis freeze milestone; on the AI’s draft it also says when the AI proposed it, and once finalised it says who finalised it and when. Finalising with answers you have not saved saves them first.
See the review’s approval history, and export it
Sign-off status
The Sign-off status page now ends with the review’s History, latest recorded first, each step with who did it, what it was about, when, and the reason they gave. It records screening decisions, on titles and abstracts and at full text, with their overrides and withdrawals; extracted values and their quotes, with validations, withdrawals, rejections and undoing them, and answers about quotes; finalising and reopening GRADE and CERQual ratings and CASP appraisals; rulings on screening and extraction conflicts; and automatic locks. Other stages will follow. History starts when this feature arrived; earlier steps are not listed. Choose Last 30 days, Last 90 days or All time; a line above the list counts the approvals, withdrawals and unlocks in that period, beside a small chart of the last 30 days. Load older shows more.
Export, in the page’s header, downloads the review’s approval record, the steps History lists, as a spreadsheet (CSV) or as one record per line (NDJSON). Like History, it starts when this feature arrived. On a phone it is in the More actions menu. Anyone who can review may export; the demo cannot.
In a history, a milestone locked without a sign-off reads Locked automatically, and a step taken by Axelium support names the person with “Axelium support”.
Sign-off status now says when a fixed rule decided something no person has checked, and who locked the search strategy
Sign-off status
Quotes the quote check found in the paper and treatment groups matched by their names now read Automatic only, with a short explanation, until a person checks them. A point whose only unchecked decisions were made this way says Automatic only as its status. Proposals the automated review applied by itself still say by policy under their count; when nothing else waits, their status reads Automatic only.
Under the search strategy’s Locked you now see who locked it and when, for example “Locked by J. Doe · 2 Apr 2026”. A strategy locked by the AI or automatically says so; when nobody was recorded, the count says person not recorded instead.
A review with nothing to approve yet now offers Open search even after its search strategy is locked.
Overview
The Approvals card now lists treatment groups and proposals that only a fixed rule has decided, as Automatic only, and notes it under a waiting count when the same stage has such decisions too.
Requesting a sign-off shows what the milestone checks first
Sign-offs
Before you request a milestone’s sign-off, the request now lists what that milestone checks, each with a tick or a warning and its counts: Protocol lock checks the search strategy, and holds the request until the strategy is locked; Screening lock checks screening and full-text screening, Extraction lock checks extracted values and quotes, and Analysis freeze checks risk of bias, critical appraisal and the certainty ratings. Manuscript lock has nothing to check. Anything still waiting for a person links to its page, and holds the request until it is done. Where nothing is approved yet and everything is still in review, the row shows a neutral mark instead of a tick, and does not hold the request. Values marked validated with nobody recorded are named, and do not hold it.
The request says who approves it, an admin of this workspace, and its note is now labelled Note for the approver.
Approving a sign-off now asks you to confirm first. Turning one down, withdrawing your own request and unlocking the protocol ask for the reason in a dialog, shown on the review’s Versions page; if the action is refused, the reason shows under the button.
Rejected values you can bring back, and decisions that say when they were not recorded
Data extraction
A value excluded from the analysis now stays excluded when it is edited, extracted again or changed in the review queue. To bring it back, select Rejected on the outcome in Form: it returns unvalidated, ready to check again, and its History records who brought it back. This works the same way in a network meta-analysis, unless a saved network analysis holds the outcome: then the undo is refused, with a message saying so.
A quote you add by selecting words in a source now reads Confirmed by a reviewer, and the value’s History records that you added it.
When a decision on the outcomes the checks flagged cannot be recorded, for example because extraction was locked meanwhile, the review queue now names each outcome and says why, and that the others were recorded, rather than reporting every decision submitted.
Validating a study’s values and acting on its inbox items at the same moment no longer fails: the second waits for the first.
Screening
When two reviewers agree to exclude a study at full text, the first reviewer’s exclusion reason is now kept, and the Conflicts page shows each person’s reason.
Once full text has decided a study, that decision stands. A ruling on its title-and-abstract conflict is recorded and closes the conflict, and the conflict card says the study stays where full text put it. Reopening the study, or accepting, rejecting or importing it in the living-review inbox, is refused; change the decision on the full-text screening page.
See what still needs a person: Sign-off status
Sign-off status
A new Sign-off status page, after Report in every review’s sidebar, shows who approved what in your review and what still needs a person. It lists, stage by stage, everything a person approves, from screening decisions and extracted values to risk-of-bias and certainty ratings, with how many a person approved, how many were decided automatically and how many still need a person. Each row, and every count in it, opens its stage.
Counts that look finished but that no person checked say so: AI only, when the AI’s decisions stand unchecked, and person not recorded, when something is marked approved but nothing records who approved it. Anything your review does not need is marked Not required, with the reason, rather than shown as zero.
A team review also shows its milestones, from Protocol lock to Manuscript lock: which are locked and who approved them, which wait for approval and which are still open. A lead reviewer or admin can request a sign-off from the page.
Overview
The Overview has a new Approvals card, below the study funnel: what is waiting for a person, largest first, each linking to its stage, and where the review’s milestones stand. See sign-off status opens the full page.
In Review health, Extraction is now Validated: the share of extracted values that a named person has validated, out of every value that has not been rejected. Values marked validated with no record of who validated them are left out of the share and counted beside it as person not recorded; validate one again to put your name on it.
Screening decisions that count, and a clearer value history
Screening
In single-reviewer screening, each person’s decision on a record is now kept as theirs. When two people decide the same record differently, it goes to the conflicts queue, and analysis waits until the conflict is resolved.
An automated re-screen never overturns a decision a person made, including one made while the AI was still screening. Where the automated review disagrees with a person, it leaves the person’s decision in place.
Once screening is locked, the AI no longer changes any screening decision. A registered trial found in a paper’s full text shows as pending for a few seconds while the AI screens it.
Leaving a full-text exclusion restores the decision a person made at the title and abstract stage.
Data extraction
A value’s History now leads each step of approval with what was done (Validated, Withdrawn, Proposed) and who did it, with their role in the review and the reason they gave. The AI’s extraction reads as an AI proposal, an automatic step says why it happened, and a value marked validated whose validator was not recorded says so. The newest four steps show first, with the earlier ones a click away.
When you ask the Analysis assistant to change a validated value, the value keeps its validation and its History shows the change as the AI’s.
Withdrawing a validation no longer clears a rejection, and a box drawn by hand around a quote can no longer be added once extraction is locked.
Clearer messages
When a screening decision, a conflict ruling, a validation, an answer about a quote or a restart cannot be saved, the message now says why in plain words: your session has expired, you are in the read-only demo, the stage is locked and where to unlock it, someone else changed it since you opened it, or your role cannot make that change.
Clearer messages when a rating cannot be saved
Certainty and appraisal
When a GRADE or CERQual rating cannot be saved, the message now says why: your session has expired, you are in the read-only demo, or someone else changed the rating since you opened it. A refused change to a CASP appraisal now says why in plain words too.
Approving a sign-off now also keeps a copy of the finalised CERQual ratings, as it does for GRADE and risk of bias.
Milestone names match the rest of the app
Sign-offs
Milestones now use the same capitalisation as other labels: Protocol lock, Screening lock, Extraction lock, Analysis freeze and Manuscript lock. You will see the new names when you request a sign-off, in pending requests, on version badges and in notifications. Only the names change; your sign-offs and versions are unaffected.
Each criterion's evidence marked in the abstract and the paper
Screening
In a record's abstract, the passage the AI relied on for each criterion is marked on its own, numbered like the criterion's chip and shaded by its verdict: met, not met, or not established. Where two criteria share words, both numbers show.
Select a criterion's chip to outline its passages and bring them into view, and select it again to clear them; clicking a passage selects its criterion. In Focus, the chips beside the record outline the passages in the abstract. At full text, the Abstract tab marks the same passages, and choosing a quote in the paper outlines its criterion there too.
At full text, the paper marks every quote found in it at once, numbered by its criterion and shaded by what it shows; a quote found but not counted is muted. A criterion's chip now steps through its quotes one by one, showing which of them you are on, and brings the quote into view; clicking a marked quote chooses it. On a trial registry record, every part that holds a quote opens, the raw record included.
On a PDF, every quote but the one you are reading is outlined on its page. A dashed outline is approximate: it marks the paragraph the quote is in, with the numbers of the criteria quoted there. Once a reviewer has opened a quote and it was found on its page, it is outlined where it is, for everyone on the review. Click an outline to choose its quote.
Only passages found word for word in the abstract are marked. Records screened before 10 June 2026 kept no quote to check their passages against, so their abstracts are no longer highlighted; their verdicts are unchanged.
Data extraction
After an extraction, an outcome the paper does not report reads "Not reported in this paper." in the Verify table; it read "Not extracted yet", like an outcome no extraction had reached. An outcome for which the extraction found no passage to read from is listed with those it could not finish. A value already stored is never replaced by either.
Every extracted value says whether to use it, check it, or read the paper
Data extraction
The Verify table files each value as Needs you, Unconfirmed or Ready, with a chip and a one-line reason on the line and the full reasons in the row card. Unconfirmed is new: nothing speaks against the value, but its quote has not been checked against the paper, was found only as a close match or not at all, or the automatic checks left no record. The study list, the study header and the triage matrix count the same way.
A suspected swap of the arms, a trial registry that disagrees, and a paper reporting a different effect measure now file a value under Needs you, with the reason in words. Within Needs you the most pressing line of an outcome comes first, and each check names its severity.
Validate study and Validate all take Ready values only; the confirmation offers the Unconfirmed values as a choice you tick. The row card says who confirmed a quote and the page it was found on, and the Key explains the quote marks.
An outcome the extraction could not finish is listed as such, with its reason, instead of reading as not extracted. The check that asks whether the arms are the wrong way round now runs on every extraction, a quote the extraction cannot find in its evidence buys one repair attempt, and evidence not found in the stored document is set aside before a value is read from it.
Review saved multi-arm studies before preparing their network
Network analysis
When arm-level editing is enabled, adopt saved positional arms by explicitly matching every outcome against the paper. Original values, sources and prior comparison statistics remain in review history; missing Ns are not inferred.
Quote confirmations resolve the matching source check without clearing unrelated checks. Complete and validate arm values and context, prepare treatment-group review, then confirm groups, reference and follow-up windows before refreshing evidence. No model is fitted by these preparation steps.
Add source quote offers a keyboard-accessible way to attach exact words from the selected source, including parsed tables. Repeated text requires a passage choice; missing text cannot be attached.
Clearer entry and limits for multi-arm network studies
Network analysis
When arm-level editing is enabled for a review, enter the randomized arms together for each outcome and follow-up, with source evidence and outcome-specific denominators kept separate from randomized group sizes.
Single-trial networks can show direct comparisons after an explicit common-effect choice, but do not report estimable between-study heterogeneity or independent inconsistency; prediction intervals and treatment rankings are withheld.
Subgroup analysis in the Analysis assistant
Analysis
Ask for a subgroup analysis by a study tag, an extracted variable or a study detail such as the journal. Each group gets its own pooled estimate, with the test for subgroup differences; studies without a value are left out and counted. Meta-regression on a numeric variable works the same way.
The message box is ready as soon as the page opens, and a failed step now says what went wrong and what to try, such as too few studies or a model that did not converge.
Living-review searches in your search history
Living reviews
Automatic update searches now appear in Recent search activity, with the dates searched, studies found and new studies added. Earlier updates are not listed.
Updates search up to 2,000 studies per database. Incomplete searches identify the affected database and revisit those dates at the next update.
Ask questions in the live demo
Live demo
Ask about the demo review with a limited number of questions per visit. The remaining count is shown beside the message box; conversations are not saved.
Request access from any demo page or when your questions run out. Visits last up to eight hours; the demo remains read-only, with living-review updates paused.
Clearer GRADE ratings for ratio outcomes
GRADE
Imprecision ratings now assess hazard, odds and risk ratios on the same scale as your results, using 1 as the no-effect value. Explanations and minimal important differences use that scale too.
Re-run an analysis to update its earlier ratings.
The right appraisal checklist for older reviews
Appraisal
Older trial reviews now open on risk of bias with their existing assessments. Mixed-methods reviews show both risk of bias and CASP; qualitative reviews keep CASP.
Your chosen appraisal settings are preserved. Opening the page does not start an assessment.
More reliable participant counts for network analysis
Network extraction
Participant counts are read more reliably from table headings and checked against reported event percentages. Percentages are not used to invent missing group sizes.
When no outcome provides a group’s size, network setup may use its explicitly reported randomised size, with a note for each study. Counts that contradict the outcome’s percentages are not used.
Re-extract a study to apply these improvements to its existing values.
Open the Triage matrix from Extraction
Extraction
The Triage matrix is available from the Extraction header in both Verify and Form. Switch back to Study view to return to the study you had open.
Choose how much evidence full-text screening needs
Full-text screening settings
Open the settings beside Screen full texts to require 1, 2 or 3 supporting passages per criterion. Two is the default; Advanced lets you require more evidence for exclusion and change automatic confirmation.
Preview how undecided AI-screened studies might change before saving. New settings apply to future screening; existing results stay until screened again, and your own decisions are preserved.
Decisions and generated Methods sections record the settings used, including changes made during screening.
More reliable quotes from ClinicalTrials.gov
Full-text screening
Quotes containing symbols or quotation marks are now recognised more reliably, helping settle eligibility criteria. Existing decisions stay until the studies are screened again.
More accurate living-review alerts for ratios
Evidence-change alerts
Alerts now correctly assess changes in hazard, odds, risk and rate ratios, including protective effects and confidence intervals crossing 1.
Alert emails and summaries show results on their reported scale. Settings are unchanged, but a review may receive an alert at its next update when earlier changes are reassessed.
A more consistent app experience
Navigation and controls
Page names match the sidebar, sub-pages link back to their parent, and notices keep the next action easy to find. Review is now called Proposals.
Buttons, fields, tables and tabs follow a consistent style. Extra actions sit in a More menu, and important confirmations appear within the app.
Dark theme and smaller screens
Improved contrast makes text and status messages easier to read in both themes.
On narrow screens, multi-pane pages show one pane at a time without losing your place or unfinished edits.
Account pages
Settings brings Profile, Notifications and Connected apps together. Team managers can open Billing from their team page, and usage shows one row per kind of work.
Send unsure studies on to full-text screening
Send to full text
Sending unsure studies to full text no longer counts as your personal decision to include them, so full-text screening can exclude them when the evidence supports it.
Studies you included yourself still wait for your decision before exclusion. Studies sent before this change keep their earlier inclusion status.
History labels the handoff as “Sent to full text by”.
Better full-text screening of trial records and papers
ClinicalTrials.gov records
Registry records can now receive conclusive eligibility assessments. Missing posted results alone do not hold a record back, since results may appear in a separate paper.
Single-group studies do not meet a required comparator. Studies you included yourself, or whose abstract described a comparator, wait for your review before exclusion.
Papers and bulk decisions
Screening better distinguishes genuine comparison groups and primary research. Criteria listing alternatives can be met by any one of those alternatives.
Bulk exclusion can keep each study’s suggested reason or use one reason you choose for all. A study without a suggestion needs its own reason.
Existing screening decisions stay until the studies are screened again.
Keep every treatment group in network extraction
Network extraction
Trials with three or more treatment groups now retain every group, with source quotes for their values. Values can be validated when at least two groups have the numbers needed for network analysis.
Pooled dose groups are not counted twice alongside their individual doses. ClinicalTrials.gov results use recognisable treatment names.
Re-extract existing studies to update their values, and fetch older registry records again to update treatment labels.
Clearer results, abstracts and review activity
Analysis results
Ratio outcomes now display consistently as ratios across results, forest plots, review history and reports. Forest plots label the effect measure and use 1 as the no-effect value; proportions and rates use their own scales.
Evidence-evolution charts now show protective effects correctly. Re-run an analysis to update an older forest plot.
Reading and navigation
Abstracts display symbols and spacing correctly across screening, search, extraction and citation exports, while preserving highlighted passages.
Activity descriptions and labels use plain language. Living-review schedules distinguish overdue, pending and unscheduled updates.
Clearer screening lists and smoother handoffs
Screening
Empty lists distinguish completed screening from records hidden by filters and point you to the relevant tab. Changing the sort order keeps your open record and selections.
Records the AI could not screen show the reason, without implying an eligibility judgement. Exclusion reasons and decision history are clearer.
Send to full text
The handoff respects the studies, search and filters shown in its menu, including large lists.
Full-text screening and layout
Moving between studies is faster, and long sessions stay more responsive. ClinicalTrials.gov entries display as readable trial summaries with highlighted quotes and a link to the full record.
Screening adapts more smoothly to narrow windows. A dismissible introduction points out the layout and keyboard shortcuts.
Screening messages in plain language
Screening
Unsuccessful saves and screening actions explain what happened and what to try next. Unsaved decisions return to their previous state, with the same message announced to screen readers.
Exclusion reasons and conflict cards use readable labels such as “Intervention mismatch” and “No comparator”. Records without a readable abstract say so.
More accessible screening lists and abstracts
Accessibility
Screen readers announce empty lists, loading states and errors. Long abstracts and side panels in Focus can be scrolled from the keyboard.
Loading errors are easier to read in both light and dark themes.
More room to read full texts
Full-text screening
Papers have more reading space in Split and Focus, with working page controls, a page counter and keyboard scrolling.
Switch between Pages and searchable, screen-reader-friendly Text; your choice stays as you move between studies. Open in new tab, Download and Close are available above the paper.
Screen readers announce the page as you turn it.
Network analysis with settings for each outcome
Network setup
Follow three steps: check the data, confirm treatment groups and run the analysis. Readiness checks identify missing values and disconnected treatments, with links to the relevant studies.
Each outcome has its own effect measure, direction of benefit, reference treatment and follow-up window. Outcomes can run independently, and changed evidence or settings mark results as out of date.
Comparisons use treatment-group data where available, with reported estimates as a fallback. One timepoint per study is selected within the outcome’s window; omitted studies are listed with reasons.
Confirm suggested treatment groupings before analysis, or rename, merge or exclude them. Long analyses can finish in the background.
Results and certainty
Results bring together a model summary, league tables, rankings, consistency checks and a network graph, with accessible table views and downloads. Tables and figures can be pinned to reports.
Comparisons show estimates and prediction intervals on the effect measure’s natural scale. Certainty ratings include reasons and supporting evidence, with an overall suggestion you can change with a reason.
Locking an outcome fixes its evidence, treatment groups and settings while leaving other outcomes open to correction.
A redesigned screening workspace
Screening
Review the study list beside the open record, or use Focus for one record at a time. Both title-and-abstract and full-text screening support bulk decisions, saved notes and a remembered layout preference.
Use Y, N and U to decide, J and K to move, X to select and F to switch layouts. Press ? for all shortcuts.
Full texts open at the selected quote. In dual review, AI assessments remain hidden until you decide unless your settings allow them to appear first.
Both layouts support keyboards, screen readers, dark mode and phone-sized screens.
Clearer explanations in screening conflicts
Conflicts
Full-text conflicts show each criterion’s assessment and reason, alongside publication checks. Title-and-abstract exclusion reasons use plain language, and every conflict names its screening stage.
Full-text conflict decisions include an exclusion reason
Full-text screening
Lead reviewers choose a reason when resolving a conflict by exclusion. The reason is saved with the decision and counted in the PRISMA flow.
The AI reviewer’s suggested exclusion reason is shown beside its decision.
Reliable switching to two-person screening
Screening
Switching from AI + human to two human reviewers now requires both people’s decisions at each screening stage. Earlier AI decisions do not take a reviewer’s place.
Assessments stay hidden until both reviewers decide, and disagreements go to Conflicts. In single-reviewer mode, disagreeing with the AI does not create a conflict.
Extracted values beside their source evidence
Extraction
Verify shows one row per outcome and timepoint beside the source document, with status, quotes, checks and history. Switch to Form for detailed editing without losing the source context.
Flagged values offer clear ways to resolve them: confirm, correct, clear or re-extract with guidance. Confirmations record who answered and reopen if the value changes.
Confirm whether a quote is in the paper, or select source text to add a missing quote. Bulk validation and imports leave unresolved checks for review; Validate row also waits for edits to be saved.
Keyboard shortcuts support study and outcome navigation, quote checks and validation. Press ? for the full list. Assistant links open the relevant outcome in Verify.
Edit extraction guidance in Configure and upload PDFs from Full text.
Qualitative reviews
Qualitative reviews point to Coding, Synthesis and GRADE-CERQual where available, with a summary of studies, coded passages and themes. Findings are managed in Coding and Synthesis.
Easier navigation through long study lists
Full text
Browse 25, 50 or 100 papers per page and filter by title, PMID or DOI. Selections carry across pages, and bulk actions cover all selected papers.
Bulk upload uses the same paging options and matches each PDF against every study, not just the current page.
Review pages
Screening panes scroll independently, and key actions stay visible on Analysis and Proposals.
Quality-review summaries describe the evidence they used in familiar review terms.
Full-text decisions backed by relevant quotes
Full-text screening
Only passages that answer an eligibility question count towards its decision. Related but inconclusive text remains visible for you to review.
Decisions wait for you when the supporting quotes do not justify automatic confirmation. Quotes that cannot be found are clearly marked and excluded from evidence counts.
Save settings with older screening questions
Configure
Reviews with questions from an earlier framework can save their settings again. Existing questions are preserved; new or edited questions are checked against the current framework.
Re-extract one flagged outcome without changing the others
Extraction
Re-extracting a flagged outcome preserves the study’s treatment labels and all other outcomes.
Flagged-outcome decisions in value history
Extraction
Discarding or re-extracting a flagged outcome now appears in its history under your name, including withdrawn validation or changed values. Keeping a value leaves it unchanged.
More reliable verification of table quotes
Extraction
Quotes from results tables are recognised more reliably. A verified quote without a known page is labelled as found without a page reference.
Value history includes reviewer decisions and corrections
Extraction
History records second-reviewer agreement, resolved disagreements and changes to validation.
Automatic corrections show before-and-after values and are clearly distinguished from reviewers’ edits.
Safer imports from the living-review inbox
Living reviews
Import Data validates only complete values with no unresolved checks. Flagged or incomplete values wait for a reviewer in Extraction.
The import summary lists what was validated and what still needs attention. Validation records the reviewer and requires the appropriate role.
Full-text screening
Eligibility checks
Screen included and unsure studies against the full paper before extraction, with quoted evidence for each eligibility criterion and checks for primary research and reported results.
Conclusive decisions require two independent supporting passages. Conflicting or insufficient evidence goes to Needs review; missing evidence or an unavailable full text never excludes a study by itself.
Review individual studies or confirm and exclude several at once, with recorded exclusion reasons. Studies awaiting a decision or marked for exclusion do not proceed to extraction.
Settings and collaboration
Single-reviewer reviews can confirm conclusive decisions automatically, or require every decision to be checked. Automatic decisions can be reversed; automatic confirmation is disabled in dual review.
AI + human review hides the AI assessment until you decide unless you choose to see it first. Two-human review requires both reviewers, and full-text conflicts are resolved separately from abstract conflicts.
Reporting
PRISMA and RevMan exports include full-text exclusion reasons. Generated Methods sections describe the screening approach, evidence requirements and automatic confirmations.
The analysis assistant can explain which studies were excluded and why.
Include studies that report an effect estimate only
Analysis
Studies reporting an odds ratio, risk ratio, mean difference, standardised mean difference or correlation can now be pooled without treatment-group data when they provide a confidence interval or standard error.
Estimates must identify the effect measure being pooled and pass consistency checks. Omitted studies have an explanation in the data-availability summary.
Confidence intervals use their reported level. RevMan exports identify estimate-only studies where treatment-group data are required.
Clearer sign-offs and role-appropriate actions
Collaboration
Other workspace admins can approve or turn down sign-off requests from Configure or Versions, with direct links from notifications. Milestones use readable names.
Pages show the actions available to your role. Lead reviewers can invite, resend and withdraw invitations for roles no higher than their own.
Removing members or ending guest access asks for confirmation. New reviews and review moves offer only eligible workspaces.
Reviews and account
Deleting a signed-off review warns you first. Versions explains how versions are created when sign-off is unavailable.
Activity times display correctly across time zones. Sign-in returns you to the intended page, Settings has mobile navigation, and Rename saves with Enter.
Clearer access messages and review navigation
Reviews
Unavailable review links show a helpful not-found page. Analysis and Provenance open more reliably, and Provenance labels its times in UTC.
On phones, screening uses tabs for criteria, records and study details.
Collaboration
Team admins can delete a team from its page. Its reviews become personal reviews of their creators.
Analysis, report generation and starting workspace reviews are available to lead reviewers and admins.
An expired invitation offers sign-in instead of a teams page you cannot open yet.
Earlier excerpts recorded since 24 August now show whether they were found in the cited paper. Some table quotes remain unchecked rather than being incorrectly marked missing.
Older citations now open at the correct page where the quoted sentence could be located.
Clearer extraction evidence and value history
Evidence and editing
Resolved warnings are marked as answered. New checks flag mismatched effect measures, possible subgroup results, inconsistent estimates and quotes needing confirmation.
Repeated quotes are combined, source labels are clearer, and reported estimates and confidence intervals can be corrected directly in the form.
The source opens beside your work, with quotes for the active field shown first.
Validation and history
Quote confirmations record who answered and when, can be changed or removed, and apply only to the sentence reviewed. Values with disputed quotes are left out of bulk validation.
Value history shows extraction checks, edits, before-and-after numbers and validation decisions.
Saving preserves source quotes and validation for unchanged values. Moving between studies is faster and opens each study’s own timepoint.
GRADE uses confirmed appraisals and study weights
Certainty
Risk-of-bias ratings now account for reviewer-confirmed assessments and each study’s weight in the pooled result. Imprecision also considers the events available for the outcome.
Current ratings have been recalculated and may be lower. Explanations state the basis for each rating and highlight AI assessments still awaiting confirmation.
Risk-of-bias assessments match the right outcome
Risk of bias
Assessments now stay linked to the outcome and timepoint they assess, so they appear in the appropriate views and certainty ratings.
Assessments for outcomes the study does not report are not retained.
Import published risk-of-bias assessments
Risk of bias
Import published assessments with their source review and the lead reviewer who checked them clearly recorded.
Existing assessments using the same instrument are kept and identified in the import summary. Imports remain unavailable after the analysis is frozen.
Resend team invitations
Collaboration
Invitations show who sent them, when they expire and whether the email was sent. If delivery fails, copy the invitation link to share it yourself.
Resend pending invitations from the team page. The new link lasts seven days and replaces the previous one.
Confidence intervals use the reported level
Extraction and analysis
Extraction records each confidence interval’s stated level, using 95% only when none is reported. Analysis, network analysis and RevMan exports use that level when calculating uncertainty and study weights.
Forms and evidence summaries label the interval level and flag disagreements with the source quote.
Earlier values remain at 95% until updated; check flagged intervals before relying on their study weights.
Resolve risk-of-bias disagreements
Risk of bias
Lead reviewers can keep either assessment, set it aside from certainty ratings or decide later. Conflict summaries show judgements in plain language.
Viewers can read assessments; reviewers can assess and run them; importing published assessments requires a lead reviewer. Unavailable actions explain why.
New extraction quotes are checked immediately
Extraction
New extractions show whether each quote was found in its cited paper. Unconfirmed quotes ask for your review.
More control over re-extraction
Extraction
Extract all and Extract filtered ask for confirmation and show how many studies already have values. Validated and manually edited outcomes are preserved.
Paused extractions offer Keep, Re-extract or Discard for flagged outcomes. Re-extract uses the guidance you provide.
After a run finishes, flagged outcomes show their reasons and a way to re-extract the study.
Clearer labels distinguish automated checks from reviewer validation. Coverage counts, notes, badges and legends are easier to read.
Two-reviewer extraction compares the relevant values
Extraction
Second reviewers use the same outcome fields as the extraction form, including hazard ratios and confidence intervals. Single-group outcomes are supported too.
Comparisons recognise equivalent numbers and highlight missing or different values. Lead reviewers can choose which value to keep for each disagreement without changing the rest.
Bulk validation leaves flagged values for review
Extraction
Validate study and Validate all studies preview complete, unflagged values before approval, with reasons for anything left out. Flagged values can be checked and validated individually.
Reported estimates with confidence intervals count as data; inconsistent estimates are flagged. Validate all studies covers only included studies.
Leaving with unsaved changes asks for confirmation.
Clearer source details for extracted values
Extraction
Quotes show page numbers only for sources that have pages. Documents use recognisable source names, and field labels match the form.
Abstracts preserve symbols and spacing, and extraction summaries distinguish studies with and without extracted data.
See only the relevant timepoints
Extraction
The editor lists scheduled timepoints and those reported by the current study, with a single timepoint selector. Trend charts appear only when at least two timepoints have values.
Arm comparisons open at the timepoint with the most studies and show event rates as percentages.
One entry for each extraction timepoint
Extraction
Equivalent timepoint labels now appear once across the editor, Triage matrix, comparisons and validation lists.
Timepoints use readable names and keep their matching source quotes and extraction history.
Faster loading on Extraction
Extraction
Study lists open faster in large reviews and show a loading message while you wait, rather than an empty-study message.
Transfer a team review to another member
Collaboration
Workspace admins can transfer a review from its overview to another workspace member. Scheduled updates and alerts follow the new owner; other members keep their role-based access.
The overview alerts admins when the owner has left the workspace.
Team roles in plain language
Collaboration
Invitations, notifications and read-only notices use familiar role names such as lead reviewer, admin, reviewer and viewer.
Keep your place when resizing the window
Review pages
Resizing or zooming no longer resets your open study, Triage matrix, expanded sections or worklist.
Clearer Triage matrix validation choices
Triage matrix
Validation previews explain which values cannot be approved there. If none are eligible, the matrix shows “Nothing to validate”.
Consistent validation status across Extraction
Extraction
The study list, Triage matrix and validated-study count now agree: a study is marked validated only while all its saved values are validated.
Qualitative frameworks for every review
Review settings
SPIDER, PICo and SPICE are now available to all reviews. The qualitative starter includes SPIDER questions, CASP appraisal and a qualitative search filter.
When results use earlier settings, stage notices let you re-run supported steps with current settings. Earlier results are kept; screening preserves your own decisions.
More reliable search-plan saves
Search strategy
Search plans drafted by the Configure assistant save reliably, including reviews without a title. Building a strategy preserves changes made elsewhere while it runs.
If a generated query omits terms from your plan, Search identifies the source, question element and number of missing terms.
Clearer review-settings history
Settings and next steps
Save history compares changes by outcome and setting. Stage notices explain what to re-run or revisit, and the overview summarises methods and results using earlier settings.
Configure offers clearer explanations and lets you restore removed suggestions. Settings exports identify anything that could not be included.
Team review methods and templates (preview)
Browse built-in and team-authored methods, create your own versions and start reviews from team templates. Published versions stay unchanged; retired methods remain with existing reviews.
Pages that follow your settings (preview)
The sidebar, screening criteria, appraisal and certainty pages reflect your chosen review methods. Qualitative frameworks use their own questions, and reports identify the methods used.
Versioned review settings (preview)
Review settings
Save method choices, framework questions and reusable starters, with settings-file export. Framework changes highlight questions needing attention.
History supports comparisons and restoring earlier settings as a new save. Stage notices identify results made with earlier settings.
Saves preserve unrelated edits and warn about conflicting changes. Search strategy amendments still require unlocking.
Timepoint-by-timepoint review in the Triage matrix
Coverage and validation
Expand an outcome to see each study’s timepoints, values and validation status. Find gaps with Missing timepoints and Timepoint gaps, then open the relevant value.
Select a coverage count to see the studies behind it and validate eligible values together. The preview identifies exclusions and any studies that could not be updated.
Charts and comparisons
Timepoint charts show estimates and confidence intervals without joining across missing values. Corrected zero-event estimates and possible duplicates are clearly marked.
Compare intervention and comparator arms across studies to spot possible reversals or implausible values, with links to the editor.
Ratio trajectories in Provenance use a log scale around no effect.
Automatic full-text collection across journal sites
Browser companion
With your permission, the companion can collect papers from journals your institution gives you access to, including library-proxy access. It reads pages only for requested papers, not your everyday browsing; permission can be withdrawn in browser settings.
Collection is faster across journals and resumes automatically after temporary publisher pauses.
Once a review is authorised, new missing papers can be collected while Chrome is open and the companion is signed in. A run stopped by hand stays stopped until you restart it.
Authorise a review directly from Full text. Earlier permissions may need renewing to cover automatic collection of later papers.
More reliable extraction edits and validation
Editing and validation
Saving preserves source evidence, identifies your edits and clearly reports unsuccessful saves. Validate study saves edits first, and leaving with unsaved changes asks for confirmation.
Accepting source values fills the form, review decisions are submitted reliably, and rejected values stay excluded during study validation.
Qualitative findings are no longer duplicated on repeated saves. New extraction runs can start after earlier ones finish.
Status and accessibility
Loading errors are distinguished from missing evidence, and each study shows its own status, confidence and review reasons.
Ratio previews label uncertainty correctly. Clearer wording, keyboard-accessible evidence and status controls, labelled fields and announced progress make the page easier to use.
Review and remove papers collected by the companion
Collection history
Full text and Settings show which papers the companion collected, where they came from and when. You can also see whether they support later review work.
Removing collected papers
Remove one or more companion-collected papers without touching your own uploads. The confirmation explains any affected text, tables and source references.
Extracted values remain, but removed page references are not restored by fetching the paper again. Removal requires permission to configure the review; the companion will not collect removed papers again, though you can upload them yourself.
Publisher pauses
Temporary pauses show their reason and end time, and collection resumes automatically.
Open evidence directly in its source
Source viewer
Evidence links open PDFs and images, readable text from PubMed Central and trial registries, or download cards for supplementary files. Large PDFs start displaying sooner.
Open source quotes from Extraction and Risk of bias. Recovered page locations are labelled, and Open in new tab and Download remain available.
Highlights and excerpts
Where available, the in-app PDF viewer highlights the cited passage. Missing, ambiguous or differently located quotes are clearly identified; reviewers can mark the correct area with Recrop.
Source-page excerpts can appear beside values, in appraisal evidence and in Word exports. Availability depends on a known location and the document’s licence.
If the in-app viewer cannot open a document, the browser viewer is offered instead.
Ask why a study was screened out
Review assistant
Ask about a study by title, PMID or DOI to see its recorded screening decision, confidence, criterion-level reasons and supporting abstract passages.
Find records by screening status and title keywords, or ask which questions your review uses. Answers reflect the saved assessment and framework, not a new screening decision.
Ask the setup assistant about an uploaded paper
Configure
Attach a PDF review, protocol or search appendix and ask about its outcomes, methods, eligibility criteria or search strategy once it is ready.
Answers cite pages and identify information the paper does not report. Suggested configuration changes are listed for your approval before anything is updated.
More usable assistant panels
Review assistant
Suggestions and wide tables scroll within narrow panels, and the support button no longer covers Send. Tables can still be expanded for a full-width view.
Source-backed corrections to individual values
Extraction checks
A quality review can correct an individual number against its existing source quote without changing other values, groups or units.
Unclear cases come to you as suggestions. Corrections require follow-up checks, and validated or manually edited values are preserved.
Clearer quality-review findings
Review summaries
Findings and proposed changes use readable formatting, concise previews and recognisable evidence references.
The review shows each round of corrections and checks, separating changes already applied from those needing your decision and prioritising unsuccessful checks.
Discover more related papers
Citation chasing
Discover can suggest papers linked to a single seed, while using the number of connections to rank results. Similar-papers search works with the default settings.
Results identify papers already in your review. If a seed cannot be found, the warning names it and explains the resulting gap in citation coverage.
Find quality-review results from the overview
Quality review
The overview links to completed quality-review findings even when no changes were proposed. Results list automatic corrections and their follow-up checks.
Extraction progress
Completed runs distinguish successful extraction from outcomes awaiting review. Follow-up extractions are clearly labelled, and full-text readiness messages appear only when relevant.
Keep each study’s reported effect measure
Extraction
Extracted estimates record the effect measure and label used in the paper. Differences from your configured measure are flagged in either direction.
Clearly labelled protective effects are retained even without treatment-group counts. Removed or unidentifiable estimates have a recorded explanation.
Retain hazard ratios when measures differ
Extraction
Reported hazard ratios and confidence intervals are kept when an outcome expects a different measure, with a warning for review.
Confirm the correct effect measure or configure a time-to-event outcome before relying on the pooled result. Estimates belonging to another outcome are excluded with an explanation.
Screen for the comparison a study actually analyses
Eligibility
Screening distinguishes an analysed comparison from one merely mentioned in an abstract.
Ask the Configure assistant to require the analysed comparison for inclusion. Studies that only mention it are excluded; unclear cases wait for your review.
Eligibility criteria are phrased as answerable questions, with suggestions to improve vague topic labels.
Clearer screening when evidence is limited
Screening
When no abstract is available, screening can use the title and indexed publication type, clearly marking the decision as title-based for confirmation.
Contradictory reasoning goes to review. Unreadable records show unknown criteria, and the original AI suggestion is distinguished from the applied decision.
Duplicate detection
Citations without titles are no longer merged solely because they share a publication year.
Reliable conversations across reviews
Review assistant
The review-selection list stays compact so conversations remain visible. Saved conversations load reliably and offer a retry if loading fails.
Broader search coverage with clearer checks
Search strategy
PubMed and Cochrane CENTRAL searches allow more flexible wording around multi-word concepts. Strategies consider both sides of treatment comparisons and older terminology where relevant.
OpenAlex searches combine synonyms correctly and respect your date and language settings.
Strategy checks identify omitted terms and test retrieval of known-relevant papers. Broad search mode also relaxes query restrictions.
Reach every notification
Notifications
The notification menu scrolls through your full backlog, including older unread messages.
Check review settings before saving
Configure
Changes from forms and the setup assistant are checked for completeness and consistency. Problems that cannot be corrected automatically are explained before the settings are saved.
Pick up saved conversations across reviews
Cross-review conversations
Reopen, delete or start conversations from the Conversations menu. Reopening restores the selected reviews, and changing the selection identifies the earlier context.
Replies continue when you close the page, with completed answers available on return. Suggested follow-up questions appear after lookups.
Each conversation can cover up to 50 reviews. Very frequent questions may require a short wait.
Analysis assistant
The overview and Analysis share a saved conversation and show replies when background work completes.
Temporary analysis or assistant failures are explained without blocking the rest of the page.
Configure appraisal instruments and export to Word
Risk of bias
Choose instruments, published variants and required details in Configure. Apply them by study design or across all studies, with warnings for uncovered designs and unverified variants.
Multi-instrument runs show separate progress. Runs exceeding your allowance are stopped before work begins.
The 2011 Cochrane tool supports outcome-specific blinding and incomplete-data assessments.
Word exports
Reports include instrument-specific judgements, ratings and assessment status in a risk-of-bias appendix.
Concordance results include study-by-rater and agreement tables, with Word and spreadsheet downloads.
More risk-of-bias appraisal instruments
Appraisal
Alongside RoB 2.0, choose Jadad, the 2011 Cochrane risk-of-bias tool, Downs & Black or ROBINS-I. Each retains its published domains and scoring scale.
AI judgements require human validation before influencing certainty where the instrument has no scoring algorithm.
Comparison and concordance
Use separate instrument tabs and import published assessments from spreadsheets.
Compare assessments with a study-by-rater matrix, percent agreement and Cohen’s kappa with confidence intervals, with spreadsheet downloads.
A clearer, more accessible extraction workspace
Evidence and navigation
Extraction summaries use plain language, label unspecified timepoints and combine repeated quotes while showing every value they support.
Clearer action labels, summaries and contextual explanations help you see what to do next and why an action is unavailable.
Theme and accessibility
Forms, study lists and progress panels follow your theme. Keyboard navigation, labelled fields and spoken status information improve screen-reader access.
See whether a source quote was verified
Evidence checks
Quotes distinguish confirmed, approximate, missing and unchecked matches. A source that cannot be checked is not labelled as a missing quote.
You can confirm whether an excerpt appears in the paper and share that decision with other reviewers. Marking a quote missing does not change or delete its extracted value.
Clearer assistant answers and progress
Review assistants
Setup and synthesis replies support headings, lists and tables. Progress shows running, completed and unsuccessful steps while keeping results visible.
Copy replies with one click or regenerate the latest synthesis answer. Messages that fail to send return to the box for another attempt.
Charts, themes and coding summaries have clearer titles and actions. Supporting evidence cites studies by author and year, and follow-up questions take less space.
Clearer source verification in Data Lineage
Evidence
Data Lineage marks excerpts that cannot be confirmed in their cited document, including values extracted earlier.
Newly extracted values identify their source document and page where known. Uncertain locations are stated plainly.
More reliable custom R analyses
Custom R analysis
Completed custom analyses return usable results more reliably. Unsuccessful runs show an explanation so you can decide how to proceed.
Time spent on unsuccessful runs now counts towards review usage and spending limits.
More cautious automatic screening
Eligibility decisions
Records with contradictory reasoning are held for review. Screening better distinguishes uncertain details from a clear mismatch with the population you specified.
Citation-only records are recognised as missing an abstract. Other sources are checked before leaving them for review.
Confident exclusions for study design or scope are applied consistently across review types; uncertain cases remain for you to decide.
Secondary literature
Papers summarising other studies are labelled Secondary literature and excluded from primary-study screening, helping avoid repeated counting of the same evidence. You can restore any you need.
Subgroup and follow-up reports of a trial’s own participants remain eligible. Unclear publication types are assessed against your criteria.
Existing decisions stay until the records are screened again.
Preview a search strategy’s relevance
Search
See the share of a small sample that matches your review question for each source, with common reasons for mismatches.
Use the roughly 15-record sample to spot large problems, not to rank similar strategies. Unavailable checks say “not measured”, and results for an edited query are marked as belonging to its earlier version.
Clearer extraction limits and recovery
Extraction
When your allowance is exhausted, extraction stops with a clear explanation. Completed studies keep their data, and unfinished studies remain ready for another run.
Restart when your allowance resets or your plan changes. Allowances are counted correctly, and stalled runs no longer block future attempts.
Searches respect your protocol limits
Search
Publication dates, excluded terms and required terms apply to generated PubMed, Europe PMC and Cochrane CENTRAL queries, including revised searches.
Qualitative study filters are preserved. Pinned known-relevant papers remain included even outside the date range.
Strategy checks report differences from the plan and confirm that each selected source has a query.
A guided launch for automated reviews
Search setup
Build a missing search strategy directly from the workflow launcher, review it on Search and return to launch after locking it.
Search plans save more reliably and include richer synonym lists. Missing approvals and unsuccessful runs have clear explanations.
Automatic strategy drafting (rolling out)
Where available, the automated pairwise workflow drafts and checks a strategy, estimates study volume and pauses for your approval before searching.
Lock the strategy to continue or decline to refine it first. Notifications let you know when a decision is waiting, and a relevance assessment accompanies the draft.
Catch up on analysis after returning to the page
Analysis
Returning to an analysis shows its current progress and brings in completed replies and results automatically.
Interrupted runs and pending approvals are clearly labelled. The sidebar assistant shows the same progress and completion notices.
Custom R analysis (early access)
Analysis
Ask the statistics assistant for custom R methods beyond the standard options, using a fixed copy of your dataset. Enable custom R analysis in the workspace; each run requires your approval.
Results are labelled User-defined and include reproducibility details. They are not used in certainty assessments or generated reports.
Longer analyses can run in the background with progress shown in chat. Save an analysis as a recipe to repeat it on the same data.
Saved setup conversations and approval-led capture
Configure
Setup conversations retain both your messages and the assistant’s replies when you reload or return.
Browser companion
Proposed capture actions wait for Approve. Declined actions are marked as skipped.
Unlock a protocol to make amendments
Protocol changes
Unlock a protocol from Configure with a short reason, make your amendments and lock it again. Earlier locked versions and the reason for reopening are preserved.
Available to team admins, lead reviewers and owners of personal reviews.
Broader ecology and evolution synthesis
Effect measures and models
Use response ratios, standardised mean differences, correlations, coefficient of variation ratios, variability ratios or pre-computed effects. Results appear on each measure’s natural scale.
Multilevel models account for related effects within studies, and moderators can help explain variation. Results identify independence assumptions where the information needed to model dependence is missing.
If a moderator model cannot be fitted, the unadjusted result is retained.
Scope and reporting
Ecology reviews support PRISMA-EcoEvo, ROSES and systematic maps, with narrative certainty. Phylogenetic meta-analysis and moderator interactions remain out of scope.
Availability is expanding gradually while validation continues.
Pool risk differences from event counts
Analysis
Risk-difference outcomes use reported events and totals, including studies with a zero-event arm. Results use the absolute scale, with 0 indicating no effect.
Previously skipped outcomes may pool on their next run. Living reviews may report a one-time change at their next update.
More accurate usage and budget tracking
Usage
One statistical analysis request counts as one run. Search, screening, extraction, appraisal and synthesis activity are reflected more completely in usage.
Budget warnings more accurately reflect the work performed.
Clearer guidance for advanced reviews
Setup and review
Network analysis guidance explains readiness, treatment definitions, connectivity and interpretation. Ecology guidance covers supported frameworks, effect measures and early-access limits.
Search and full-text guidance explains coverage, citation chasing, uploads and browser capture, with a reminder to screen retrieved studies.
Automation pauses for protocol changes, reports remain drafts, and appraisal, certainty and synthesis interpretations stay reviewer-owned.
See the results of approved review changes
Review decisions
Approved changes can take effect within the same automated review run.
Out-of-date appraisals can be retired with approval while preserving their history. Protocol deviations record what changed, who approved it and when.
One place for proposed protocol changes
Review proposals
Review suggestions with plain-language reasons and before-and-after previews, then approve or reject them.
The page identifies pending decisions and who can approve them.
An automated second check of your review
Quality review
Check evidence completeness, out-of-date appraisals, source-backed values and outcome consistency.
Findings come to reviewers, and changes to the protocol or evidence set still require approval.
Clearer qualitative review and reporting
Qualitative synthesis
Coding agreement leads with a measure suited to sparse codes, while retaining familiar statistics for context.
Clearer review cards, study flow, codebook coverage and re-run behaviour help distinguish and maintain qualitative and quantitative reviews.
End-to-end qualitative and mixed-methods reviews
Qualitative synthesis
Move from question setup through screening, appraisal, coding, themes, CERQual confidence ratings and Summary of Qualitative Findings.
AI can draft supporting work; themes, interpretations, appraisal decisions and confidence ratings remain reviewer-owned.
Easier monitoring of living reviews
Living reviews
See update status, pending studies, alert settings and evidence changes more clearly.
Unsuccessful updates offer recovery actions, and alerts link to the relevant update.
More connected appraisal, collaboration and reporting
Review workflow
Reviewer-validated risk-of-bias and certainty judgements flow more consistently into reports.
Teams have clearer roles, assignments, notifications, conflict handling and sign-offs.
Broader search coverage, citation chasing, reference-manager handoff, supplementary files and manual uploads support everyday review work.