Evidence

Evidence is the flowss studio for reading a body of literature and turning it into a living, citation-grounded evidence graph. You ask a research question, bring in the papers (pasted abstracts, PDFs, a search on OpenAlex, or a reference-manager export), and Evidence extracts every relationship each paper asserts, checks that each one is backed by a verbatim quote from that paper, and draws what the literature says as a map of concepts and links. On top of that map it rates certainty with GRADE, pools effects, finds contradictions and gaps, answers questions with clickable citations, and exports a report with a Methods section and a PRISMA flow. You open it at /atlas. This page is a complete tour of the studio: every view, panel, menu and shortcut. The systematic-review workflow itself (importing references, screening, de-duplication, extraction, verification, risk of bias, GRADE and living updates) is covered in depth in Screening, extraction and synthesis.

At a glance

What Evidence does, and what it refuses to do

The atomic unit in Evidence is a claim: one paper asserting that one concept relates to another (for example, sleep deprivation decreases working memory), together with the exact sentence from the paper that says so. Claims are extracted by an AI model, but nothing the model writes is trusted on its word:

• Every claim must carry a quote that can be found, verbatim, in the source text. Claims whose quote cannot be found are dropped before they reach the graph.
• The quote is read in its sentence. If the sentence negates or qualifies the quote (a "not" that the quote left out), or mentions neither of the claim's concepts, the claim is dropped.
• Numbers the model attached (effect sizes, sample sizes, arm-level counts) are kept only if they actually appear in the source text. Numbers that do not are set aside and never pooled, graded or exported as data.
• Everything after extraction is deterministic: the concept graph, contradiction detection, gaps, GRADE certainty, pooled estimates, PRISMA counts and reports are computed by rules you can inspect, not by a model.

Evidence also says plainly what it did not do. A run that ran out of time before it opened every source is marked as partial on every screen. A certainty rating made without a risk-of-bias assessment says it is an upper bound. A publication-bias test with too few studies says "not tested" rather than "no bias". When something is missing, the studio tells you what and why, and what to do next.

Plans, sign-in and limits

A few allowances apply to everyone who can open Evidence:

• Synthesis cost. On hosted AI, a synthesis or re-run is charged one point per screened-in source the run may open, up to 10 points per run, and only after extraction has delivered. A run where extraction failed on every source is not charged. The charge is the same whether or not strict verification is on. With your own key, points are not used, but each request counts toward the daily ceiling for your own key.
• Ask and Suggested searches. One point per message you send to the flowss Evidence Agent, and one point per Suggest run. An Ask answer is charged once its first words reach you; an answer that fails before anything arrives is not charged.
• Starter reviews and the evidence graph JSON. Opening a starter review or re-importing an exported graph runs no AI and costs nothing.
• Exports and causal analysis. Report, bibliography and CSV exports, and the causal analysis on the Analyze panel, count toward a daily allowance of 60 requests per account. The allowance is the same on every plan, and Send to Evidence from other studios draws on it too. The JSON export and the graph images are made in your browser and do not count.
• Literature lookups. Each Find papers search, each DOI lookup, and each batch of up to 50 DOIs looked up during a reference import counts as one lookup. The allowance resets at midnight UTC.

See Plans and what they include and AI points and limits for the full picture.

The screen

Evidence uses the same island layout as every flowss studio: the review fills the screen, and the controls float over it in five places.

The title is the saved project's name, or, for an unsaved review, your research question shortened to a phrase. With no question yet it reads Untitled review.

Screen, Analyze and Lenses read the synthesis, so before you have one they stay on the rail but are disabled, with the reason in their tooltip (for example, "Synthesize, import or load the example first — the analysis reads the graph"). Screen is available earlier if you have imported records or opened a saved project, because screening reads the review's records rather than the graph.

On a phone (narrower than 768 pixels), the Sources flyout and the inspector open as bottom sheets over the review, one at a time: opening one closes the other. The colour legend starts hidden on a phone; turn it on from the zoom percentage menu.

Your first review

The quickest way to see Evidence work is the built-in example. Its three studies are fictional and labelled as such in their titles, so they cannot be mistaken for real citations. The example is sources, not a finished graph: synthesizing it uses AI (your own key on Starter, or points on Plus and above). To see a finished review with no AI at all, open a starter review instead (see Starter reviews).

1. Open /atlas. On an empty review the canvas shows Add sources, then synthesize ("Every claim is checked against its source quote before it enters the graph.") with a Load the example synthesis button, four starter chips (PRISMA review, GRADE summary, Living review, Rapid review) and two links, Add sources and All starter reviews.
2. Click Load the example synthesis (also on ⋮). The question "Does sleep deprivation impair working memory, and what role does cortisol play?" is filled in, three example abstracts are added, the note "Example dataset loaded — three fictional abstracts for trying the pipeline, not real publications." appears, and the Sources flyout opens on them.
3. Click Synthesize evidence graph at the foot of the Sources flyout. While it runs, the canvas reads Extracting claims….
4. When the graph appears, press ⌘. (Ctrl+.) to open the inspector on Inspect: a summary, the certainty distribution, key findings, contradictions and gaps. The example is built to show a contradiction (one study finds an effect, another finds none) and a mediation chain through cortisol.
5. Click any link on the graph to see its certainty, its studies, its pooled effect and the quotes that back it. Click a concept to see everything connected to it.

To start your own review instead, type your question into the prompt at the bottom and press Enter. With no sources yet, the prompt's button reads Find papers: Evidence opens the Sources flyout and searches OpenAlex with your question. Add the papers you want, then press Synthesize evidence graph.

Tip: Two example questions appear as chips above the prompt until you have sources ("Does fasting improve insulin sensitivity?" and "Does sleep loss impair memory?"). Clicking one puts it in the prompt for you to edit or submit. The × beside them (Hide suggestions) hides them for good in this browser.

Warning: Load the example synthesis replaces the research question and the whole source list without asking. Save or export a review you care about before you load the example over it.

Starter reviews

A starter review is a complete, ready-made review: a framed question, a corpus with its screening decisions, claims that quote their source verbatim (with arm-level counts where a trial reports them) and, where the method calls for it, RoB 2 or ROBINS-I appraisals. It is synthesized in your browser by the same deterministic engine every edit uses, so it opens instantly, needs no AI key and costs no points. Every study in every starter is fictional: each source title ends "(fictional)", and the footnote reads "Every study in these starters is fictional — invented abstracts and numbers for learning the workflow, not real publications."

Open the gallery from ⋮ Starter reviews…, from All starter reviews on the empty canvas, or from one of the four starter chips there. In ⌘K, each starter is also a command of its own: type "Open starter" or a word such as PRISMA, GRADE or living.

The Starter reviews flyout groups the starters by category and has a search field (Search — PRISMA, GRADE, living, scoping…) that matches titles, descriptions, questions, categories, frameworks and the question's elements. Each row shows the title, a one-line description, and a line such as "PICO · 6 studies · 40 records". Question and methods expands the question, its framework elements (for example Population, Intervention, Comparator, Outcomes) and the methods notes, so you can check the protocol before you open it. A search with no match reads "No starter matches … Try PRISMA, GRADE, living, scoping or rapid."

The gallery holds 18 starters:

Picking a starter replaces the open review, exactly as New review does, and asks first when the open review exists only in your browser. Once it opens, a note beside the corpus says what loaded (for example "Starter loaded — "Systematic review and meta-analysis (PRISMA 2020)": 40 records, 6 included studies."), and a message suggests what to try next. A starter is an ordinary unsaved review: you can screen it, appraise it, edit numbers, export it, or Save as project to keep it. Its included studies are recorded as already read, so a later re-run in a saved copy reads only records you screen in afterwards.

The research-question prompt

The prompt at the bottom of the screen is where you ask the review's question. Its placeholder reads Ask a research question…, or New research question… once a graph exists. Press Enter to submit and Shift+Enter for a new line. ⌘/ (Ctrl+/) puts the cursor in it from anywhere.

What the button does depends on the review:

The AI settings button on the prompt opens one setting: Strict verification ("A second pass by the flowss Evidence Agent checks each quote supports its claim."). It applies to the next synthesis or re-run. See Verification for what it adds.

Errors appear just above the prompt, with a Dismiss button. They stay until you dismiss them or start the next action. Notes that are news rather than failures (a diagram arriving from another studio, a review restored from this browser, a re-run's coverage report, "Marked 3 duplicates.", a partial-review export warning) float at the top centre of the canvas and stay until you dismiss them.

The Sources flyout

Sources on the rail opens the review's corpus in a 360-pixel flyout beside the rail. You can also open it from ⌘K (Question & sources) or the empty canvas's Add sources link, and it opens by itself when you ask a question with nothing to read yet. The rail button's badge counts sources that have text. Escape, a click outside, or Close sources closes it. Keep open (the pin in its header) keeps it beside the graph for a long session of adding papers: a pinned flyout ignores outside clicks and Escape, closes only from its close button or the rail button, and reopens on your next visit when the window is at least 1,024 pixels wide.

From top to bottom, the flyout holds:

• Research question. The question the next run asks (placeholder "e.g. Does intermittent fasting improve insulin sensitivity?"). It is the same question as the prompt's; edit it in either place.
• Find papers — topic or DOIs and Find. Search OpenAlex by topic, or paste one or more DOIs. Results appear in a list beneath; click a result to add it. Find is disabled while the field is empty.
• Sources (N), where N counts every card, with two import buttons: PDF and References (which reads Importing… while it works).
• The import note, which reports what the last import brought in and what it could not.
• One card per source, each with a title field (Source N — title), a text box (Paste the abstract / full text…) and a remove button (Remove). The last remaining card cannot be removed.
• Add source, which adds a blank card for pasting.
• At the foot, Synthesize evidence graph for an unsaved review, or for a saved project Check for new papers (when the find field holds a query) or Re-run with new sources (when it does not). After a re-run, a green Re-synthesis: line under the button says what changed.

Pasting text

Click Add source, give the card a title, and paste the abstract or full text into the text box. A pasted source has no screening status, so it is treated as an included study and is read by the next synthesis. A card that has a title or a DOI but no text warns you, because a synthesis skips it: "No text on this source — add the PDF or paste the text. Synthesis skips a source with no text." For a record that came with a DOI the warning reads "No abstract upstream for this DOI — add the PDF or paste the text. Synthesis skips a source with no text."

Uploading PDFs

Click PDF (or ⋮ Import ▸ Import PDFs…) and choose up to 10 PDF files at once. The text is read in your browser; the files themselves are not uploaded. Each PDF becomes a source titled with its file name (without .pdf), holding the text of its first 40 pages. Pages beyond the 40th are not read, and the import note does not say when a PDF was cut short, so for a long document paste the sections that report the findings. The import note then says what happened, for example "2 of 3 PDFs added · 1 with no selectable text — a scanned PDF? Paste the text instead (scan.pdf)." The note names every file that was not added, and why:

If nothing in the batch could be added, an error says so (for example "No selectable text found in that PDF — a scanned PDF? Paste the text instead. Nothing was added.").

Finding papers on OpenAlex

Type a topic and click Find (or press Enter). A topic search returns up to 8 works from OpenAlex, and only works that carry an abstract are ranked. Paste one or more DOIs instead (bare 10.xxxx/... DOIs or doi.org links, separated by spaces or commas) and each DOI is resolved to its record, whether or not OpenAlex holds an abstract for it. One lookup resolves at most 50 DOIs; split a longer list.

Each result shows its title and year. Two badges may appear:

• no abstract: the record has no abstract upstream. Adding it brings in the citation but no text to extract claims from; add the PDF or paste the abstract afterwards.
• retracted: OpenAlex marks the work retracted. You can add it for the record, but the synthesis leaves its claims out. The source card then reads "Retracted, according to OpenAlex. It stays in the corpus for the record; the synthesis leaves its claims out."

Click a result to add it. A paper already in the list is not added twice. If nothing comes back, the list reads "Nothing found. A topic search only ranks records that carry an abstract; a DOI resolves whatever OpenAlex holds, abstract or not." If OpenAlex cannot be reached, the error ends "You can paste source text manually."

The date of the last topic search is remembered with the query. It is printed in the exported Methods section as the date the search was last run, and kept with the project when you save. If you edit the query afterwards, the date no longer applies and is left out.

The find buttons on the Gaps list and on Suggested searches to run use this same search, and they put their query into the Find papers field. In a saved project that matters: the button at the foot of the flyout follows the field (see Keeping a review living), so put the saved query back, or clear the field, before you press it.

Importing references

References (or ⋮ Import ▸ Import references…) imports a search export from PubMed, Embase, Scopus, Zotero or EndNote as RIS, BibTeX or PubMed .nbib. Every record becomes a source awaiting screening, abstracts are filled in by DOI where the file has none, and the note reports exactly what came in. Imported records must be screened in before a synthesis reads them. The whole process, its limits and its messages are described in Importing references.

Synthesizing

Click Synthesize evidence graph in the Sources flyout, Synthesize on the prompt, or Synthesize on the empty canvas once sources and a question are in (it reads Ready to synthesize). A synthesis:

1. Reads only the sources that have text and are included studies. Pasted text and PDFs count as included; imported records count only once screened in.
2. Sends each one to the flowss Evidence Agent's extractor, a few at a time, within a budget of about 55 seconds.
3. Verifies every claim against its source text (and, with Strict verification on, asks a second AI pass whether the quote supports the claim).
4. Builds the concept graph, detects contradictions and gaps, rates certainty and lays out the result.

When it finishes, the graph replaces whatever was on screen, the inspector is switched to its Inspect tab (open it with ⌘.), and any earlier Ask conversation is cleared. A synthesis always starts a new, unsaved review; to add papers to a saved project, use its re-run button instead (see Keeping a review living).

Warning: A project saved from imported records has no graph yet, so the canvas still offers Synthesize under Ready to synthesize. Pressing it starts a separate, unsaved review and detaches the screen from the project. To build the project's first graph, screen records in and use Re-run with new sources at the foot of the Sources flyout.

Before anything is sent, Evidence checks two things and refuses by name if either fails:

If the run returns an empty graph, Evidence tells you which of several different things happened, because each has a different fix. These messages are described in the Troubleshooting table below.

The full detail of extraction, verification and how a partial run is reported is on Screening, extraction and synthesis.

Reading the evidence graph

Graph (G) is the concept map of the synthesis: every verified relationship drawn as a link between two concepts.

Concepts

Each concept is a card showing its name, its kind, how many times it is mentioned and how many links it has. A concept's accent colour shows its kind, and its card and name are drawn larger the more connected it is.

Drag concepts to arrange the map. On a saved project, an owner's or editor's arrangement is saved to the project automatically, so collaborators see the same layout; on an unsaved review, and for viewers, positions hold for the session.

Links

A link's label is its relationship (for example increases, no effect on, correlates with). When more than one study backs it, the label adds the count: decreases ×3. Studies are distinct sources, not claims: a paper that states the same thing twice counts once.

A dashed link rests on a single study: dropping that source would remove the link. The legend at the bottom left names five of these colours (increase / cause, decrease / prevent, correlation, no effect, contradiction), and shows a single-study sample when any dashed link is on screen. While a concept is selected, links that do not touch it are faded.

Press C (or Colour by certainty in the zoom menu) to colour links by their GRADE certainty instead: green for High, blue for Moderate, amber for Low, red for Very low. The legend switches with it.

Evidence recognises twelve relationship types: increases, decreases, causes, prevents, treats, correlates with, associated with, no effect on, moderates, supports, contradicts and part of.

Selecting, focusing and the selection toolbar

Click a concept or link to select it. A small toolbar appears just above it:

• Focus shows only the selected concept and its direct neighbours. Press F again, or clear the selection with Escape, to see the whole graph.
• Source quotes opens Inspect scrolled to the link's Backing claims, so the quotes are the first thing you read.
• Appraise its studies opens the Screen panel at its Risk of bias section.
• Ask the flowss Evidence Agent about it opens Ask with a question filled in, such as "What does the evidence say about cortisol?" or "How strong is the evidence that sleep deprivation decreases working memory?".
• Analyze as the exposure or as the outcome fills that side of the causal analysis on Analyze.

Double-clicking a concept or link selects it and opens the inspector. ⌥F10 (Alt+F10) moves keyboard focus to the selection toolbar.

The zoom menu: filters and display

Click the zoom percentage on the view island for the graph's view options:

The − and + buttons zoom out and in. The same menu also holds Zoom to fit, Zoom to 100%, Full screen and Hide interface (⌥Z); on a phone it adds Zoom in and Zoom out. The graph binds no zoom keys of its own; ⌘+ and ⌘− remain your browser's. The Pipeline view zooms from the same controls.

Lenses

Lenses on the rail redraws the graph and the table:

The other views

Table

Table (T) shows the synthesis as rows. When the table is on screen, the view island shows a Table switch with Evidence and Summary.

The Evidence table has the columns From, Relationship, To, Studies, Effect and Certainty. Click any heading except Effect to sort by it (click again to reverse); it starts sorted by Studies, most first. An amber warning sign beside a relationship means it is contradicted. An amber Effect means the studies' effects are heterogeneous. Click a row, or focus it and press Enter or Space, to select that relationship and open it in the inspector.

The Summary of findings groups relationships by outcome. Each outcome has its own block with Determinant, Relationship, Effect, Studies, Participants, Risk of bias and Certainty, sorted from highest certainty down. A Risk of bias cell is amber when the domain has not been assessed, or has been assessed for only part of the evidence; read the footnotes beneath the table for what that means for the rating. Rows are selectable in the same way.

Both tables read the same single certainty rating per relationship that the graph and the inspector use, so the four always agree. See How certainty is rated.

Pipeline

Pipeline (P) shows how the synthesis on screen was produced, as six connected stages. Each shows a status colour (grey idle, violet running, green done, red when the run was partial) and a short statistic.

When a run's time budget ran out before it opened every source, Ingest, Extract and Verify turn red and say so, for example "12 of 50 sources read · 38 never opened". The later stages keep their colour, because what they show is true of the sources that were read. Before any synthesis, a line reads "Stages fill in as the synthesis runs — add sources and Synthesize, or load the example." The pipeline zooms from the view island like the graph.

Entities

Entities (E) lists the concepts in the review as entities. It has two scopes:

• This project lists every entity with its links (outgoing and incoming), its claims and its sources, marks the most connected as hub, and shows a count line ("N entities · N hubs · N isolated · avg N links"). Search with Search entities or their relations…. Expand an entity to see its relationships; the counts there are claims, not studies. Selecting an entity opens it in the inspector, and you can focus it on the graph.
• Across workspace shows Workspace entities: canonical concepts shared across all your Evidence projects. On a saved project, Link this project's concepts proposes which of this project's concepts are the same as entities you already have, each marked exact, alias, fuzzy (with a confidence) or new entity. Only exact matches and new entities are ticked for you; tick the rest yourself after reading them, then Apply selected. Click an entity to see its relationships across projects, the projects it Appears in, and any Cross-project disagreements (projects that disagree on how it relates to another concept). Unlink removes one project's link to the entity without changing that project's concepts or claims. An unsaved review shows "Save this project (top bar) to link its concepts across your workspace."

If an entity is also linked in projects you no longer have access to, it says so ("+N unavailable"), and its cross-project view covers only the projects you can still read.

The inspector drawer

The inspector opens at the right when you ask for it: from a tab, ⌘., the selection toolbar, a double-click on the graph, or a row in a table. It opens 384 pixels wide; drag its left edge to resize it, and it remembers the width. You can pin it open. Its tabs are Inspect, Ask, Team and Runs; Screen and Analyze open in the same drawer as a single panel titled Screening or Analyze. Before a synthesis exists, asking for the inspector shows "Synthesize, import or load the example first — the inspector reads the graph." (except Screen, as described above).

Badges on the tabs count real things only: Ask shows the number of messages, Screen the records awaiting screening, Team the members once there is more than one, and Runs the number of runs.

Inspect: the overview

With nothing selected, Inspect summarises the whole review:

• The synthesis summary, then the certainty distribution as coloured dots with counts.
• Evidence base: the study designs behind the claims (for example "3 RCT · 2 cohort").
• Status chips. A rose N/M sources read chip comes first when a run was partial. Then N/M grounded, N rejected (claims removed by strict verification), a rose N unverified (claims a strict pass could not check), a robustness chip (robust, N load-bearing or robustness not assessed) and the year span. The robustness and year chips open Analyze.
• Key findings: the five strongest relationships by number of studies, then certainty, each with its certainty dot. Click one to inspect it.
• Next steps: up to six recommendations drawn from the analyses, such as "Resolve conflict: A → B", "Corroborate A → B", "Investigate A ↔ B" or "Update the evidence base". Most are clickable and select what they are about.
• Tidy concepts: Likely duplicates among the concepts (for example an acronym of another concept, or a spelling variant), each with a Merge button. Merges you have made are listed above with an Undo this merge button. Merging relabels every claim and recomputes the graph; on a saved project the merge is saved to the project.
• Most connected: the six best-connected concepts.
• Contradictions: each pair of concepts where studies disagree, with their positions, and beneath it how the disagreement reconciles: whether a moderator explains it (study design, population or publication year) or it stays open. Click one to inspect it.
• Gaps: up to eight pairs of concepts that are both studied but never studied together, each with a find button that searches OpenAlex for papers linking them.

Inspect: a concept

Selecting a concept shows its name, kind, mentions and every relationship it takes part in; click one to inspect that link. Trace mechanism finds chains of evidence from this concept to another: choose the target in to concept… and up to eight paths of at most four steps are listed, each with its length and its weakest link in studies. If none exists, it reads "No evidence path within 4 steps."

Inspect: a link

Selecting a link shows the full evidence behind one relationship: its certainty pill, its number of studies, a single study — fragile flag where it applies, the pooled effect with a forest plot, publication-bias checks, robustness, the absolute effect, a subgroup comparison by study design, the reasons for its certainty rating, and every Backing claim with its quote. Reviewers with edit rights can Edit numbers on a claim. Two How this was calculated buttons open a full record of the rating and of the pooled estimate, which you can download as reproducible JSON. All of this is documented in Pooled effects and the link inspector.

Ask: the flowss Evidence Agent

Ask is where you talk to the flowss Evidence Agent, a multi-turn agent grounded in this graph's verified claims and the analyses derived from them. It can summarise the evidence, draft narrative for a review, explain contradictions and suggest searches.

• With an empty conversation, Try offers ready-made questions: "Summarize the current state of the evidence", "Draft a narrative summary for a review", "Explain the contradictions and which side is better supported" (when there are contradictions), "Suggest literature searches to fill the gaps" (when there are gaps) and "What are the most important findings, and how certain are they?".
• Type in Ask anything about this evidence… (later Ask a follow-up… (⌘↵)) and press ⌘Enter (Ctrl+Enter), or click the send button (Send (⌘↵)). Answers stream in; Stop generating stops one, and the answer is marked stopped.
• Answers cite findings inline as [n] ("Jump to this finding"). Click a citation to select that finding's link on the graph. A bracketed number that names no finding is marked as unresolvable instead of being drawn as a live citation; hover it for the reason. If a citation ever fails to resolve when clicked, a notice says it names no finding and should be treated as unsupported.
• Every answer is checked after it arrives. If it cites a number that names no finding, a Citation check line is added under it ("[7] refers to no finding in this evidence base — treat it with caution."), and an answer that cites nothing at all is marked "this answer cites no finding in this evidence base — treat it as unsupported."
• The flowss Evidence Agent is told to use only the findings for empirical claims, never to invent a citation or a finding, and to say plainly when the evidence does not cover something. Search and next-step suggestions may go beyond the findings, labelled as suggestions. With no findings at all, the answer is "The current evidence base doesn't address this."
• Clear (shown when the conversation has messages and nothing is generating) empties the conversation. On a saved project the conversation is saved automatically with the project, for every member.
• Suggested searches to run appears when the review has gaps, weak evidence or contradictions. Suggest (then Refresh) asks for concrete literature searches with a one-line rationale each, and each has a find button that runs it on OpenAlex. "No suggestions right now." means the model proposed none.

The flowss Evidence Agent sees at most 200 findings per question. On a larger review it is given the 200 most relevant to your question, is told it holds only part of the evidence base, and every answer drawn from a subset ends with a note saying so ("Answered from N of M findings — the ones most relevant to this question…"). It reads the most recent 24 messages of the conversation, each up to 8,000 characters. A review with more than 5,000 findings cannot be sent to the assistant ("Too many claims (>5000).").

Analyze

Analyze on the rail opens three analyses:

• Causal analysis. Choose an exposure… and an outcome… concept and click Analyze. Evidence reads the causal relationships of the included studies as a directed acyclic graph and reports the confounders and a back-door adjustment set (do-calculus): a summary sentence, Adjust for: the concepts to adjust for, and advice. Analyze as the exposure and Analyze as the outcome on a concept's selection toolbar fill these in for you.
• Robustness. A leave-one-out check: would the conclusion survive dropping any single study? It gives a verdict and lists up to eight sources as load-bearing, supporting or redundant, with the links each one is the Sole support for and the contradictions it drives. With only one source it says robustness cannot be assessed; with more than 400 sources the full check is not run and the summary says so.
• Timeline. A summary sentence, claims per publication year as a small bar chart (when the claims span more than one year), up to five Evolving findings where the relationship changed over time (marked overturned where it reversed), and up to five Contested across years where the disagreement spans the corpus without resolving. Click any of these to inspect the link. Studies without a publication year are counted as undated.

Screen

Screen opens the PRISMA screening panel and the Risk of bias appraisal for each study. Before a project is saved it shows Screening (PRISMA) with a Save as project button. Everything on it is documented in The Screen panel and Risk of bias.

Team

On a saved project, Team lists the Members (you included) and their roles. Only the owner sees the controls: an Invite by email field with an Invite button (or press Enter), a role menu on each member (owner, editor or viewer) and a Remove button. Invitations are added as editor; change the role afterwards if needed. Members who are not owners see each role as a label. The person you invite is added straight away, with no invitation to accept, and the project then appears in their list. Before saving, Team offers Save as project.

Runs

Runs is the dated Run history of a saved project: each run's trigger (create for the save, manual for a re-run with your sources, query for a check for new papers), when it ran, what it changed and how many claims it added. A project with no runs reads "No runs yet. Re-running the synthesis against new sources records a dated entry here." Below it, N extraction note(s) expands to up to 20 per-source notes, such as a source whose extraction failed or whose text was too long to read in full; after a re-run, that run's notes come first. ⋮ Runs… opens this tab too.

Undo

The view island's Undo button reverses the last curation action and names it, for example Undo concept merge or Undo screening of "Smith 2021". It holds the last five actions, newest first, and is emptied by New review. Undo covers concept merges (and removing one) and screening status changes; it does not cover a synthesis, an import, a de-duplication, a vote, a risk-of-bias judgement, a change to an exclusion reason alone or an edit to a claim's numbers.

Saving and projects

Kept in your browser

Every review is saved in your browser as you work, shortly after each change, and restored on your next visit with a "Restored your last review from this browser." message. This keeps the question, the sources, the graph, the claims, publication years and risk-of-bias judgements. If the review is too big to keep with its full source texts, Evidence keeps it without them and tells you the papers' text will need re-adding. If it cannot be kept at all, a note asks you to use Save or Export so it is not lost on reload. When you arrive through a project link or a Send to Evidence hand-off, that takes the screen instead, and the browser copy is left untouched for your next plain visit.

The browser copy keeps the review's content, not which project was open. After a reload, open a saved project again from the title's ▾ list, ⋮ Projects… or its link to keep working in the project (and to re-run, screen or share it).

Save as project

Save as project stores the review on your account. You need to be signed in. You can save as soon as you have a synthesis, or straight after importing records, so that you can screen them before anything is read:

• With a synthesis, Save as project is on ⋮, in the Share popover (Save to share a link), and on the Screen, Team and Runs panels.
• With imported records but no synthesis yet, it is on the Screen panel.

A project is named after the research question (its first 80 characters). With no question it is named Untitled synthesis, or Untitled review for records saved before any synthesis. A project can hold up to 5,000 findings, and one save can carry up to 8 MB.

Saving stores the claims, the sources with their screening status and bibliographic details, the saved search query and the date it was run, which sources the synthesis actually read, and your risk-of-bias judgements. If some judgements could not be stored, a notice says how many and asks you to re-record them on the Screen panel. A saved project gains a link, a team, screening votes, a run history and living updates.

The status badge beside the title shows how the review is kept: Saved (On your account), Saving while a save or re-run is in progress, Saved in this browser only, or Not saved (Save or export it). Signed out, the shell's own row says the review is kept in this browser and offers Sign in. Open the badge (or press ⇧⌘M) for the full status: The evidence lists problems in the evidence (claims not found in their source, unreconciled contradictions, very-low-certainty findings, or certainty not yet rated), Next says the next step, and Graph counts the concepts and links on screen. The status also carries the review's readiness level, one of Draft, Sound, Reviewed or Publishable. See Readiness.

Opening, renaming, duplicating and deleting

• Rename is available to owners and editors, and Delete to the owner only. Rename in the Saved projects list asks for the new name in a "Rename this review" prompt; names are kept to 200 characters, and an empty name is refused.
• The Saved projects list shows Rename and Duplicate to owners and editors only; the title's ▾ list offers Duplicate and Delete on every review, and refuses a delete by a non-owner with "Only the owner of "name" can delete it."
• Duplicate creates a copy you own named "… (copy)", and says what it carried: the corpus, the findings and the risk-of-bias records, with the concept merges and the saved layout, but not the run history, the screening votes or the Ask thread, because those belong to the project they happened in. Making a copy needs Starter or above, like any new project.
• Delete asks first. From the Saved projects list the question is "Delete "name"? Its sources, findings, screening decisions and run history are removed for everyone it is shared with. This cannot be undone."; from the title's list it is "Delete "name"? This cannot be undone." Deleting the open project detaches the review on screen from it.
• New review (also on ⋮) starts an empty question, one blank source and no graph. If the current review exists only in your browser, it asks first ("Start a new review? This one is kept only in this browser — save it as a project or export it first to keep it."), because starting over replaces it.

Saved projects also appear on your Library board, and open straight into the review from there. The Library learns the list from this device: it is refreshed each time you open Evidence signed in, so on a new device open Evidence once before looking for your reviews in the Library.

Sharing and collaboration

Share is the filled button at the top right. It offers:

• Copy link on a saved project. It copies the project's address (/atlas?project=<id>), and only members of the project can open it; anyone else who follows it is told "Not found". The message reads "Link copied — members of this project can open it". A signed-out member who follows the link is asked to sign in and then lands on the project.
• Save to share a link on an unsaved review, which saves it as a project first.
• Send to…, which hands the graph to another studio (see below).

To give someone access, add them on the Team tab. They need an existing flowss account; otherwise the invite fails with "No flowss account with that email — ask them to sign up first." Inviting people and promoting a viewer to editor need Plus or above on the inviting owner's account; demoting someone, handing ownership to an existing editor, or removing someone never needs a plan. Invitations are metered: an account can send up to 100 a day, and one address can be invited up to 8 times a day across flowss.

A project must always keep at least one owner, so the last owner cannot be demoted or removed: transfer ownership first.

Note: Live presence, voice and screen sharing are temporarily unavailable in private projects, including Evidence projects. Saved projects, screening and the Team tab still work, and members see each other's saved changes when they next open the project. The ⌘K entries for joining voice and screen sharing say so.

See Sharing and permissions and Teams and organisations.

Exporting

Exports are on ⋮ Export ▸ and in ⌘K (type "Export"). They appear once a synthesis exists.

The report opens with the question, a summary, the certainty distribution and the evidence base, then an auto-drafted Methods section (search, screening, extraction and verification, risk of bias, GRADE). Then come the key findings (up to 25 relationships), contradictions and their reconciliation, gaps, the evidence table, causal identification (a back-door analysis of the strongest relationship), robustness, temporal trends, limitations, recommendations, the GRADE summary of findings, the PRISMA flow with an excluded-studies table (when any record was excluded), and the sources. The footer states whether every claim in the file is grounded in a verbatim quote. Sections with nothing to say are left out. The HTML and Word reports include up to four forest plots of the best-evidenced relationships (with funnel plots where a relationship has ten or more studies), a risk-of-bias traffic-light figure and the PRISMA 2020 flow diagram. See What goes into the Methods and PRISMA.

The image exports draw the graph as it is on screen, with its current filters and colours; from another view, Evidence shows the graph first and then exports it. CSV cells that could be read as spreadsheet formulas are neutralised.

Exports are made from what is on screen, not from the server's copy of a project, so they also include sources you have added or imported since you saved. The Methods section prints the Find papers query in the field at the moment you export, and the inter-rater agreement of the open project.

Warning: If the run behind the graph was partial (its time budget ran out before every source was opened), an export completes but warns you: the file's PRISMA counts and bibliography describe your whole source list, including papers no claim came from. Re-run to cover them before you circulate it.

To bring an exported review back, use ⋮ Import ▸ Import an evidence graph (.json)…. The file replaces the open review entirely (question, sources, claims and graph) and no AI runs. It is detached from any project that was open: its run history, team, screening votes and Ask thread do not carry over. Files exported before the product's rename are accepted too. A file that is not a valid evidence graph is refused with "That file isn't a flowss Evidence graph."

See Exporting your work for exports across flowss.

Moving work between studios

Send to… in the Share popover hands the current picture to another studio as a diagram:

Studios that render the diagram natively open it directly; others show a convert chip and a preview of the conversion (Open in …) before you open it. A conversion uses a built-in converter where one exists and AI otherwise, which costs a point on hosted AI. The PRISMA flow's boxes are observed counts, so a conversion is checked: if any count changed, it is refused with "The conversion changed the PRISMA counts, so it was not sent — open the flow in Studio instead, which renders it as it is.", and a conversion that passes carries a note that every count appears but which box holds each was not checked. Sending the PRISMA flow to Lab writes it straight from the counts, with no AI and no points.

Diagrams can also come into Evidence. Send to… in another studio offers Evidence: the diagram's concepts and relationships are written out as a starting source (titled "Imported", then the engine's name and its numbers of concepts and relationships), and a question is proposed from its first concepts, such as "What is the evidence for the relationships between A, B, C?". A diagram whose engine has no structural reader is sent as the first 8,000 characters of its source text instead, with the question "What does the evidence say about this model?". The hand-off replaces the open review's question and sources. The note reads "Imported a diagram as a starting source — review the question, then Synthesize. Add papers to ground it." A diagram is your hypothesis, not a publication, so its own claims are never accepted as evidence: add papers, and the synthesis tests the hypothesis against them. See Moving work between studios.

Keyboard shortcuts

Single-key shortcuts work only while no button, field, table row or menu has focus; Escape works everywhere. On Windows and Linux read ⌘ as Ctrl and ⌥ as Alt.

The full list across flowss is on Keyboard shortcuts.

Tips

• Open a starter review (⋮ Starter reviews…) to see every panel populated, including PRISMA counts, risk of bias and a GRADE summary of findings, before you bring your own papers. It costs nothing and needs no AI key.
• Use Load the example synthesis to watch extraction and verification run on three short abstracts.
• Use DOIs in Find papers when you already know the papers: a DOI resolves even when OpenAlex has no abstract, so you get the citation and can add the PDF for text.
• Turn on Colour by certainty and Certainty at least: Moderate to see only the relationships you would stake a conclusion on.
• Hide single-study links strips out everything that has not been replicated; dashed links show the same thing without hiding them.
• Before trusting a link, use Source quotes: the quote is what the link stands on.
• Record risk of bias for each study from the link's Appraise its studies button. Until every study is judged, every certainty rating is an upper bound.
• Save as a project early on a real review: screening, the team, living re-runs and the Methods search date all need a saved project.
• Keep an Export evidence graph (.json) of important reviews. It re-imports without any AI.

Limits and known constraints

• Evidence needs a signed-in account on Starter or above. Starter uses your own AI key for extraction and Ask; hosted AI starts on Plus.
• One synthesis reads at most 50 screened-in sources with text. Each source contributes at most its first 24,000 characters to extraction; the rest is not read, and an extraction note says so.
• A synthesis has a budget of about 55 seconds. A large corpus may be read only in part; Evidence marks such a run as partial everywhere and you can re-run to cover the rest.
• PDFs are read in your browser, 10 per batch, up to the first 40 pages of each, and must have a text layer. Scanned PDFs need OCR or pasted text.
• A topic search returns up to 8 works with abstracts, and one DOI lookup resolves up to 50 DOIs. Full texts behind paywalls are not fetched; abstracts are used.
• Indirectness, one of GRADE's five domains, is not assessed from claims and is omitted.
• The flowss Evidence Agent sees at most 200 findings per question and the most recent 24 messages, and refuses a review with more than 5,000 findings.
• A saved project holds up to 5,000 findings; exports, Ask and saves each carry at most 8 MB.
• Live presence, voice and screen sharing are temporarily unavailable in Evidence projects.
• Inviting collaborators needs Plus or above, and invitees need an existing flowss account. There is no invitation to accept: an invited person is added at once.
• The inspector's tabs (Inspect, Ask, Team, Runs) and Analyze need a graph. Before one exists, only the Screen panel opens, and only for imported records or a saved project.
• A saved project does not send its sources' text back to the studio when you reopen it later. To re-read records, import the same file again or re-add the PDFs or text (see What a synthesis reads).
• The readiness level shown beside a project in the Library is remembered on this device only.
• The browser copy of an unsaved review is limited in size; very large reviews are kept without their source texts, or not at all. Save or export them.

Troubleshooting

For problems with screening, re-runs and imports, see Troubleshooting on the screening page. For everything else, see Troubleshooting and FAQ.

Related pages

• Screening, extraction and synthesis
• The template gallery
• Lab and Reporting standards for PRISMA and other reporting checklists
• Figure analysis and statistics
• AI in flowss, AI points and limits and Bring your own AI key
• Your library and Moving work between studios
• Sharing and permissions and Teams and organisations
• Exporting your work
• Plans and what they include
• Keyboard shortcuts
• The work model and The continuum