Publishing, provenance and compliance in Lab

A research figure is finished when a reader can trust it, a checklist is satisfied and a journal will accept the file. The Lab has a pane for each of those jobs. Provenance shows where every number on the figure came from, and which ones a person typed with no source. Reporting standards scores the document against CONSORT, PRISMA and the other checklists that govern it, item by item, with repairs you can preview and apply. Publish walks you through four steps, from choosing a journal and measure to building a submission archive that holds the figure, its source, its data and a preflight report. Revision history keeps the last fifty snapshots of each document so you can always go back. This page documents every control, state and message in those panes, the Lab's export commands, and what each one can and cannot prove.

At a glance

The Lab is a Starter workstation, so every feature on this page is available on Starter, Plus, Ultra and Enterprise. None of them uses AI or AI points. In particular, the Publish pane's Export LaTeX is generated in your browser, unlike the AI-based TikZ conversion in the main Studio. See Plans and what they include.

Opening the panes

The panes open in the column on the right of the canvas, one at a time, alongside the Advisor, Comments and the flowss Lab Agent — Lab methodologist. Opening one closes whichever was there, and also puts the Block properties drawer away. Each has its own header with a count and a close button, and shows the open document's title so you always know whose record you are reading.

On Windows and Linux, read ⌘ as Ctrl and ⇧ as Shift.

Provenance

The Provenance pane reads the figure's passport: a record of every number the document declares, where each came from, and which numbers each derived value was computed from. It leads with the distinction that matters most to a reader: a number read from data or computed in the document looks identical on the figure to a number someone typed, but only the first can be checked against anything.

The pane is read-only. Clicking any value with a location puts the cursor on that line of the source. The header counts the numbers recorded ("42 numbers").

The six origins

Every number has one origin. The pane uses a short chip label for each:

A source note is reproduced verbatim. The Lab has never opened the table, paper or file it names.

The header

The top of the pane shows:

• A headline, such as "12 of 40 numbers came from data or a named method; 25 were typed by a person."
• Split chips with counts: from the kernel, from data, derived here, typed, cited, typed, no source, and no origin when there are any.
• A caution whenever anything was typed: "A typed number and a computed one look identical on the figure. The difference is that only one of them can be checked against anything."
• Traceable to data, a fraction and a percentage, for example "14 / 40 35%", with the sentence "14 of 40 numbers can be traced to something outside the document." The percentage is rounded down, so a document one number short never reads 100%. An empty record scores 0, not 100%.
• The rule: a value counts as sourced only when it names a dataset or carries a source note, or, if it was derived, when every number underneath it does. One typed number at the bottom of a chain makes the whole chain unsourced.
• The caveat: the score measures whether an origin was recorded, not whether the figure is right. It is not a score to maximise.

The five tabs

The tab strip holds Numbers, Derived, Methods and Data, each with a count, and Limits.

Numbers

Every number, grouped by origin from strongest to weakest. Each group shows its chip, title, count and a one-line explanation. Each row shows the number's path (for example randomised.n), its value, and its location ("line 19"), or "no span" when it has none. Click the row's description to expand it:

Derived

Every derived value with its expression, each input with its origin chip and value, a restatement of the arithmetic, and a note on whether the chain traces to data. An input that is not recorded is highlighted. If the restated arithmetic disagrees with the stored value, it is shown in red with a warning icon. When nothing is derived, the tab says so and suggests writing n: = treat.n + control.n so the compiler keeps totals honest.

Methods

The statistical methods recorded for this figure, with the kernel version. Each row shows the method's name and id, its module, the version it was recorded at, the question the module answers, the formula and the citation. Badges flag a method not in this build's registry, one with no method record, and a version that differs from this build. When the registry has no citation, the row says so rather than inventing one. produced lists the values the method produced; click one to jump to it. When nothing is recorded, the tab says "No statistical method is recorded for this figure." and explains that this means none was recorded, not that none was computed.

Data

The datasets recorded for the figure: name, rows and columns, column names, a short content hash or no content hash recorded, and which values read them. A dataset that a value names but nothing recorded is listed as cited, not recorded. With no data at all the tab says "No data is attached to this figure. Bind a dataset and the numbers read out of it stop being claims and start being readings." Bind rows on the Data bench; see The Data and Statistics benches.

Limits

Three sections, in this order:

• What this record cannot tell you. A recorded source is not a checked source. A traced number is not a correct number. A dataset hash detects change, not truth, and is not proof against a hostile author. A unit that parsed is not the right unit. Anything nobody recorded is invisible.
• What the passport found. The passport's own findings, such as datasets without a content hash.
• Numbers that name no source. Every unsourced path, each a jump to its line.

The footer

The footer identifies the record: the figure id (or "(unnamed figure)"), the view, a short document hash, a short checksum and the generator. Any integrity problems are listed in amber. The last line says what the checksum is for: "The checksum pairs this figure with the data it was built from. It is change detection, not a signature — it proves nothing against anybody who wants to deceive you."

If the document does not compile, the pane shows no record at all, because a passport describes one specific version of a figure and showing the last one that worked would describe a document you are no longer looking at. Fix the source and the record returns.

Raising traceability

1. Read your rows on the Data bench and derive a dataset block, so the document declares the data it was built from.
2. Annotate numbers read from the data with source { dataset: …, column: …, rows: … }.
3. Annotate numbers copied from a paper or report with source "Table 2" (or similar).
4. Replace hand-added totals with derived expressions, so they cannot drift from their parts.
5. Open Limits and work through Numbers that name no source.

The Advisor turns this into a worklist: its Say where the numbers came from band proposes Record where N of M numbers came from, Read these numbers from data instead of typing them, or Declare the dataset in the document.

Reporting standards

The Reporting standards pane scores the document against the reporting checklists that govern it. The Lab knows 33 standards: CONSORT 2010 and its cluster, non-inferiority and pilot extensions, SPIRIT 2013, PRISMA 2020 and its flow diagram, PRISMA-S, PRISMA-ScR, MOOSE, STROBE and its cohort, case-control and cross-sectional variants, STARD 2015, TRIPOD 2015, TRIPOD+AI, RECORD, ARRIVE 2.0, CARE 2013, SRQR 2014, COREQ, TIDieR, CHEERS, SQUIRE, an ML-evaluation figure standard, model-card disclosure, RoB 2, ROBINS-I, QUADAS-2, ROBIS, AMSTAR 2 and GRADE. The standards are chosen from the document itself: declare a study design, arms, a cohort chain with exclusions or an economic model, and the standards that apply appear. For the catalogue and the Standards bench's scaffolds, see Reporting standards.

The summary

The header counts the standards that apply ("3 apply"), or shows not scored when the document does not compile. A row of chips lists each applicable standard, most relevant first, with its score. Pick one to see:

• Its name, version and year, and the score as two integers, met over applicable (for example "31 / 37").
• A headline: "31 of 37 applicable items fully reported."
• A bar and chips with the four statuses: Reported, Partial, Unknown and Missing, plus N excluded for items that do not apply (shown as N/A on the rows). Excluded items are in neither half of the fraction.
• A note when items are Unknown: "N items the engine cannot decide from the figure — counted as not reported, and yours to confirm by hand."

There is deliberately no single percentage, because many manuscript items cannot be decided from a figure and a confident percentage would hide that.

The three tabs

Clicking an item's header jumps to its line in the source. If an item failed while being checked, it says "This item failed while running, so the verdict above is not a judgement — check it by hand."

Previewing and applying a fix

Many items ship a repair.

1. On To fix, click Show the fix (or Fix available on the Checklist tab).
2. Read the preview: a description, the lines it would add (prefixed +), any it would remove (prefixed -), and a note. The preview is produced by recompiling your document with the change.
3. Click Apply this fix. A toast confirms it (with the fix's own note, or "Fix applied") and offers Undo; ⌘Z works too.

A fix writes placeholders, never data, and the item stays outstanding until you replace them. If the recompiled change was refused, the preview says "The engine recompiled this change and refused it, so it is not offered." and there is no apply button.

Starting from a compliant example

When the document does not follow a standard, the foot of the pane offers Start from a compliant example with one or more starter documents, and the Doesn't apply tab offers Start from this standard buttons. Loading one replaces the open document without asking first. The toast "Template loaded — <name>" offers Undo (⌘Z works too), and any rows bound on the Data bench are released, because they described the study that was there before.

Starter documents contain example values from fictional studies and say so in a notice comment near the top; see The "Example data" stamp.

Publish

The Publish pane takes a finished figure to a journal in four steps. A stepper across the top shows each step's number, short name (Target, Export, Preflight, Bundle) and state, and the header counts finished steps ("2 / 4"). Each step's section shows its title, a state chip, a one-line status and, at the bottom, a Limit line stating what that step cannot do.

The pane works on the view currently shown on the canvas. Your chosen destination and measure stay chosen while the Lab is open, even after you close the pane or switch documents; reloading the page clears them.

Step 1: Choose a destination and a measure

1. Type in Search journals — nature, plos, ieee…. Up to eight matches show, each with its publisher, "row unconfirmed" (or the date it was confirmed), and its single / double column widths in millimetres (a dash where none is captured). "N more match — narrow the search." appears when there are more.
2. Click a destination. The chosen row shows its publisher and how its numbers were sourced ("published guideline", "derived from template" or "needs confirmation"). Change brings the search back.
3. Choose a measure: single column, 1.5 column, double column or maximum width, each labelled with its width in millimetres. A measure the destination does not publish is greyed out, and its tooltip says so. The measure starts at single column.

The section then shows:

• Recommended: a measure, with the reason. With a destination chosen, it is the measure that uses the least page area while still fitting the destination's height limit (for a destination that states a width band, the nearest permitted width). If no measure fits the height, it says the figure has to lose height, by restacking panels or splitting it.
• What fitting to this measure does. Before-and-after lines for the Measure (with the scale production will apply, such as "Everything below is multiplied by 0.62"), the Height against the destination's ceiling, and type, stroke and raster lines where they can be measured.
• The destination's own findings, each with a severity (error, warn, info), a rule name, what was observed, the requirement and the fix. In the Lab you typically see format-not-accepted (an error, see below), a column-width warning when the figure is drawn wider than the measure and will be scaled down, max-height (an error) when the scaled figure is taller than the destination allows, and width-unspecified (info) when the row has no width for the chosen measure.
• A provenance note for the destination's row, with an Author guidelines link.

The destination table holds 30 rows:

Warning: The table is a convenience list captured from published author guidelines, not a live feed. No row has yet been confirmed against the publisher's current page from inside the app, so every row is marked "row unconfirmed" and the step shows: "These numbers have not been confirmed against the publisher's current author guidelines from inside this app — open the guidelines link and check the width, resolution and format list before you submit." Where a publisher states no resolution floors, the table uses the standard print convention (300 dpi halftone, 600 dpi combination, 1200 dpi line art); the checklist's resolution item shows the floors in force for your destination. If your journal is not listed, pick the closest family and confirm the numbers yourself.

The Lab measures the figure's drawn width and height from the rendered view. It does not supply type sizes or stroke widths to this step, so the fit lines show Type sizes and Stroke widths as "none measured"; the type and stroke checks run on the artwork itself at step 3. The Lab's figure format is SVG, which no destination in the table lists, so after you choose a destination step 1 reports a format-not-accepted error and stays blocked. Convert the exported SVG to an accepted format (usually PDF or EPS) in your vector tool, or submit a PNG from step 2 where the destination accepts PNG, and treat this finding as a reminder.

Step 2: Export the artwork

This step's status reports what has actually been written, not what could be. In the Lab it turns done once you export the SVG of the artwork currently on the canvas. Exporting LaTeX or a PNG does not change the step's state, and editing the figure after exporting returns the step to not started, because the file you saved no longer matches the artwork. While LaTeX is available but not yet written, the status says so: "available, not written: no file exists until you export one."

The raster card's title states the width and destination, for example "Raster at 89 mm (Nature, single)"; the measure appears under its short name (single, oneAndHalf, double or max), and with no destination the title ends "(drawn size)". When the destination publishes no width for the selected measure, the card says "… publishes no 1.5 column width, so there is no printed size to rasterise at. Pick another measure." Before anything is rendered it says "Nothing measured yet — render a view and the printed width appears here." A PNG file is named like glyphscript-flow-89mm-300dpi.png, its background follows the theme, and the toast reports its pixel size and printed width ("PNG exported — 1051×700 px, 89 mm at 300 dpi"). If the browser cannot draw a canvas that large (more than 16,000 pixels on a side or 200 million pixels in all), the export is reduced and a second toast says so.

About the LaTeX:

• It is generated in your browser, with no AI and no points.
• The status line shows the figure kind, the view, the number of coordinates and the mode (snippet). The file holds the picture alone (a tikzpicture or a tabular) with no preamble and no figure float, so load TikZ, and pgfplots or booktabs where the figure needs them, in your own document.
• Notes about what the conversion did are listed under the status, and the download toast counts them ("LaTeX downloaded — 2 notes in the panel").
• It supports these view renderers: CONSORT and PRISMA flows, funnels, experiment flows, forest and meta-analysis plots, bar, point and line charts (including OKR charts), ablation, FMEA, cap and OKR/KPI tables, Gantt, roadmap and timeline views, C4 container views, and multi-panel composites. Any other view is refused, for example "TikZ refused this view (declined renderer)." or "TikZ refused this view (unmapped renderer).", with what would be lost and what to do instead, and Export LaTeX is not offered. In a multi-panel composite, a panel TikZ cannot express becomes a labelled gap rather than refusing the whole plate.

Step 3: Preflight the rendered artwork

Click Run preflight (later Re-run). Preflight scans the rendered SVG at its final printed size against the destination's numbers. It runs on your click, so "not started" is a real state, and the result is discarded the moment the artwork, the destination or the measure changes. With no destination chosen it still runs the checks that need no published number, such as script tags and low-resolution images.

The verdict comes in one of three forms, shown exactly as the checker states it:

The verdict box also counts blockers, warnings and advisories. Findings are listed blockers first, then warnings, then advisories, each group headed with its count and its meaning:

Each finding shows its position and check id, its message, the requirement where one applies, the fix, and where in the file it was found. Findings a tool could fix mechanically carry a tool badge. The filter buttons All N, Fixable by tool N and Needs the author N split the list, with a sentence such as "4 of 7 findings can be fixed mechanically. The other 3 need a decision only the author can make — every one of them changes what the figure says, not how it looks." When a filter half is empty it says "Nothing in this half of the split." The Lab does not apply these fixes for you; the split tells you which are cosmetic and which change what the figure claims.

What this scan could not measure lists the scan's own limits on every run, clean or not: transforms on ancestors are not composed, inheritance is approximated, clip paths, masks and reused elements are not resolved, text is not measured for overflow or collision, and a pass means these checks passed, not every check that exists.

Note: The verdict counts preflight's own findings only. When step 1 raised destination errors, step 3's Limit line says those errors are outside this verdict and outside the archive's preflight.txt, so "no blockers" here is not a statement about them.

Preflight ignores the "Example data — not real" stamp when measuring, so it does not count the stamp's text as part of your figure.

Step 4: Assemble the submission bundle

Click Build submission bundle. The Lab writes one ZIP, named like glyph-<title>-<yyyymmdd>-<code>.zip, and downloads it. A green box names the file and lists any warnings; a red box headed "The bundle was not written." gives the reason. If preflight has not been run, the step warns first: "Preflight has not been run, so the archive would carry no verdict. Run step 3 first — the bundle is the artefact a co-author trusts." You can still build it, and preflight.txt will then say "Preflight: NOT RUN."

Checklist · N expands the chosen destination's submission checklist: draw at final width, stay inside the height limit, export an accepted format, set the colour mode, check raster resolution at final size, check the smallest type, check the thinnest stroke, embed or outline the fonts, letter the panels to house style (when the destination has one), name the files in citation order, and confirm against the current guidelines. Items marked advice are recommendations rather than requirements.

What a bundle built in the Lab contains:

The verdict written into preflight.txt is exactly the one the pane showed. CHECKSUMS.txt lists every file except itself and manifest.json, so anyone receiving the archive can verify that nothing changed after you built it. The dates inside the archive are the moment you click. The file name's date is the UTC day, and its code is derived from the title, that date and the source, so two builds of an unchanged document on the same day get the same file name. The bundle is titled after the document, or "Untitled figure" when it has no title.

Bundles built from the Lab carry no caption, author list or licence, so expect these warnings, which are accurate: "No caption was supplied; a figure without a caption is not submittable.", "No authors were supplied, so CITATION.cff names a placeholder and the bundle cannot be attributed." and "No licence was chosen, so this bundle has no LICENSE.txt and a reader has no stated permission to reuse it." Every Lab bundle also notes "No DOI is recorded, so the citation files point at no persistent identifier." Without rows on the Data bench you also see "No data tables were supplied, so this bundle contains no data/ directory…". With rows, expect "data/<name>.csv: N column(s) have no description…" and, for numeric columns, "… have no unit…", plus a note when blank cells were written as the column's missing-value code. If the document does not compile, there is no provenance.json and a warning says so. Add the caption, authors and licence in your submission system, or edit the README before you share the archive.

Bundle limits: 200,000 rows and 512 columns per data table, 96 MB per file, 256 MB per archive and 4,096 files.

Exporting outside the Publish pane

The ⋮ menu's Export submenu works on the view on the canvas:

PNG backgrounds follow the theme: white in the light theme, the Lab's dark ground in the dark theme. Copying an image needs a secure (https) page and a browser that supports image clipboard writes. For every format across the platform, see Exporting your work.

The "Example data" stamp

The Lab's starter documents, including the reporting-standard starters, contain realistic values from fictional studies. Each carries a notice comment near the top, such as "// Every string below is EXAMPLE text from a fictional trial: replace it." While a document still contains a notice with the words "EXAMPLE text from" followed by "fictional" (with or without "a" or "an" between them, and with EXAMPLE in capitals), the canvas and every export (SVG, PNG, the views ZIP and the bundle's figure) carry a faint diagonal "EXAMPLE DATA" watermark and an amber "Example data — not real" label in the bottom-right corner, and the bundle's preflight.txt opens with "Example data — not real: these values are an example that ships with the platform, not data from a study." The LaTeX from Export LaTeX carries the same warning as its first line, a comment that begins % Example data -- not real:, and the export shows an example-data note. Replace the example values with your own, then delete the notice comment; the stamp disappears at the next render.

Revision history

The Lab saves a snapshot of the open document about a second and a half after you stop typing and keeps the most recent 50 per document. Snapshots are stored in this browser and survive reloads. The Revisions pane's header shows how many are kept ("12 / 50").

Each row shows:

• How long ago it was taken ("just now", "5m ago", "3d ago" and so on), and a current badge on the snapshot that matches the editor.
• A name tag, when the snapshot has one.
• Its first line, and a summary such as "24 lines · +3 lines · line 7: arm control {". When the line count did not change it reads "edited · line 7: …"; an identical snapshot reads "no change"; the oldest reads "first snapshot · N lines".
• A What changed button (not on the oldest snapshot), which reveals the block-level changes since the previous snapshot, such as arm arm_b: n 240 → 235, or "No block changed — formatting or comments only." Click Hide changes to fold it again.

The current snapshot has no Restore button, because restoring it would change nothing.

To go back:

1. Click Restore on a row.
2. Confirm: "Replace the editor's source with this snapshot? Undo brings your current source straight back." Click Restore or Cancel.
3. The pane closes, and a toast says "Snapshot restored" and offers Undo (⌘Z works too).

To delete a document's history, click the bin icon in the header ("Clear revision history"). The pane asks "Delete all N saved snapshots for this document? This can't be undone."; click Delete history (or Cancel). A toast confirms "Revision history cleared for this document". This cannot be undone. Clearing history is also the quickest way to free local storage when the Data bench reports that storage is full. For history across the platform, see Version history and undo.

Tips

• Work the Advisor's bands in order. "Make it compile", then "Meet the standard", then "Say where the numbers came from" is the order in which problems stop a figure being accepted.
• Choose the destination early. Step 1 works before the figure is finished and tells you which measure to draw at, which saves a rescale later.
• Re-run preflight after every change. A verdict is discarded as soon as the artwork changes, so a stale one can never be mistaken for the current one.
• Bind your data before building the bundle. With rows on the Data bench, the bundle carries them and a data dictionary, and the data-availability statement says so.
• Use source notes generously. source "Table 2" turns a typed number into a cited one and lifts the traceability score honestly.
• Treat "Unknown" items as yours. The standards pane counts them as not reported because only you can confirm them.

Limits and known constraints

• Kernel origins and methods are not recorded in the Lab yet. Numbers written by the Statistics and Data benches are stored as plain values, so the Provenance pane shows them as Typed unless you add a source note, the Kernel group is empty, and the Methods tab reads "No statistical method is recorded for this figure." The method sentence and citation are still shown on the Statistics bench when you compute, so copy them into a note if you need them recorded.
• Bound datasets carry no content hash. The Data tab lists bound datasets with no content hash recorded, and the passport finds that "a later edit to the data would go unnoticed".
• The destination table is not live. No row has been confirmed from inside the app; always check the guidelines link.
• SVG is not an accepted journal format. Step 1 stays blocked with format-not-accepted for every destination. Convert the SVG, or use a PNG where accepted.
• Type and stroke sizes are not measured at step 1. Those checks run at step 3 on the artwork.
• Only an SVG export completes step 2. Exporting LaTeX or a PNG from the pane does not change the step's state.
• The destination and measure are not saved. Reloading the Lab clears them.
• Preflight reads the SVG text, not a rendering. See "What this scan could not measure" on every run.
• The Lab does not apply preflight fixes. The Fixable by tool split is guidance only.
• LaTeX is a snippet. No preamble or float; and only the renderers listed under step 2 are supported.
• Bundles carry no caption, authors or licence when built from the Lab.
• Revision history and bound rows are stored in this browser. They do not follow you to another device.
• Reporting standards can only decide items visible in the figure's model; manuscript items often land in Unknown.

Troubleshooting

Related pages

• Lab, the overview of the studio
• The Data and Statistics benches, including the Advisor
• The FlowScript language, for source notes and derived values
• Lab views and renderers
• Reporting standards
• Exporting your work
• Version history and undo
• Privacy and your data
• Science and the Science verification model
• Plans and what they include