Linked results connects a manuscript number or statement to a named calculation over local data. When the data changes, you review what should change in the paper. It preserves ordinary LaTeX rather than inserting hidden result macros.
Define a calculation
Open Research → Linked results. Choose a CSV, TSV or JSON-array dataset in the workspace and define a named result. Inspect the source columns and select any equality filters. Multiple filters are combined with AND.
Choose a single cell, mean, median, sum, minimum, maximum, sample standard deviation or row count. A single-cell calculation requires exactly one matching row. Missing or nonnumeric measurements cause visible errors rather than silently reducing the sample.
The preview shows matching rows, rounding, units and exact LaTeX output. Percentage formatting explicitly multiplies by 100: data stored as 0.87 becomes 87%, while data already stored as 87 should not use that conversion.
Example: select rows where method is ours and dataset is test, calculate the mean of accuracy, choose percentage display, and inspect the row count to confirm that the intended seeds were included.
Connect a number or statement
Select the existing number in source or Visual, then choose Link result. A unique source passage can also be pasted into the workspace.
A numeric connection can preview an exact old/new replacement.
A statement connection creates a reminder to reconsider the interpretation when data changes.
Linking does not edit manuscript text. Keep the selection specific and verify the surrounding context. A number repeated many times needs a reliable anchor, not a guessed global replacement.
Generated tables are maintained through their figure/table workflow instead of arbitrary source replacement. Related Figure Studio items can be shown when they share datasets, including joins; shared input does not imply an identical calculation.
Review changed data
The visible workspace checks sources and also offers Refresh. Changed connections are grouped by file and manuscript section. Inspect the changed input and resulting values before accepting.
Data changes still require acknowledgement when rounding leaves the printed value unchanged. An identical displayed mean can conceal changed sample size or variability. Statement connections require author review; the app does not rewrite a scientific conclusion automatically.
When the input of a statement changes, its card shows the linked value before and after. If the sentence compares values ("higher than", "lower than", "exceeds"…), the card says so, and when the sentence names another group from the same data (for example Baseline next to Guided) it computes that group with the same calculation and shows both numbers. If the sentence's direction no longer matches the numbers, the statement is marked May no longer hold: it is left out of Select all, and it can only be accepted after you tick I checked this statement still holds. The check reads plain comparative wording; it cannot judge claims such as "performs better", where the meaning depends on the metric, so read those yourself.
If the passage changed or became ambiguous, reconnect it from the editor. Do not force an update onto a different occurrence simply because it has the same number.
Apply and verify updates
Select numeric changes to accept and statements to acknowledge. Acceptance requires a named manuscript checkpoint, rechecks source/data fingerprints and refuses dirty buffers or stale previews.
After successful updates, inspect related figures and tables separately. Refresh their assets if needed; frozen items stay pinned and externally edited outputs remain protected. Recompile explicitly to inspect the final PDF. Applying a result update and compiling the manuscript are separate operations.
History records the calculation, data fingerprint, row count, scope, displayed result and accepted passages. Its version link opens the checkpoint against the working draft. The latest 30 acceptance records remain in .versofold.results.json with definitions and anchors.
Recovery and scope
If an operation is interrupted, reopening can recover from its saved journal. Recovery checks exact known states and refuses to overwrite newer external edits. Preserve the project and inspect the message rather than manually deleting the journal.
Current limits include 100 named results, 100 connections per result, 16 MB data sources and 2 MB manuscript files. Calculations use explicit equality scopes rather than arbitrary executable formulas. This feature does not infer scientific conclusions or scrape arbitrary PDF numbers.
Related: Figure Studio, remote data, versions.