Evidence and measurements
Figures and tables
Tables are read from their cells, plots are measured against their own axes, and every value read from a figure is reviewed by a person.
Tables
A table in a digital PDF is read cell by cell, and each value keeps its table number and page. This is the strongest evidence link the platform produces, which is why a publisher PDF is better than a scan: a scan gives only a picture of a table.
How a plot is read
For each figure, the platform finds the panels, identifies the chart type and calibrates the axes from the tick labels it can read. Bars, markers and curves are then measured against those axes, and each series is matched to the legend.
Every number read this way is a candidate marked as requiring human review. Values read from figures always go to a reviewer, and one joins the verified dataset only after a person has compared it with the figure and approved it.
When no number is given
A plot that does not support a number gets no number. The platform records why and moves on; this is a valid result, not a failure.
- An axis with fewer than two readable ticks, a broken axis, or a scale (linear or log) that cannot be established.
- Two y axes on one panel, because the series cannot be assigned to an axis with certainty.
- Heat maps, spectra, micrographs and diagrams, which are recognised but not measured.
- A figure that is split across pages or cannot be located reliably on the page.
- Marks that match no legend entry keep an "unknown series" label, and an error bar keeps no statistic unless the caption states SD, SE or a confidence interval.
What the confidence means
A figure value carries a confidence and an uncertainty range based on how cleanly the axes and marks were read. These have not been checked against expert readings of published figures, so no accuracy figure for plot digitisation is claimed, and a value read from a figure stays marked as such in the dataset and its exports.
Reading a figure yourself
Where no number is given, or you do not trust the reading, read the figure in a tool you trust and enter the values through the measurement import, recording the figure number and panel. These values are reviewed in the same way.
- Skip a figure whose values are also in a table or in the text; use those instead.
- Take the axis units from the figure itself.
- A log axis read as linear is the most common mistake, and quality checks cannot catch it. Check it yourself.
Related
Extracting values
How candidate values are read from approved documents, what each one carries, and why a candidate is not yet data.
Reading the structure of a PDF (parsing)
What parsing produces, why a person checks it before anything else uses it, and what the warnings mean.
Reviewing extracted values
How to read a candidate value, the four review decisions, and who can give final approval.
Confidence, thresholds and abstention
What a confidence number means, how thresholds decide who looks at what, and why the platform sometimes gives no answer.