Confidence, thresholds and abstention

Confidence shows where to look

The number attached to a screening suggestion, a vocabulary match or an extracted value says how sure the platform was. It is useful for ordering work and deciding what a person must look at. It is not a measured probability of being right.

This is why review shows the evidence beside the value. A reviewer approves the evidence, not the confidence number.

Thresholds decide who looks

The study configuration sets the thresholds for its project. They decide how much is checked individually; they never change a value, a label or a stored decision.

Two rules cannot be changed in any project. A value whose evidence could not be confirmed is never accepted automatically, and a value read from a figure always needs individual approval by a person.

Abstention: "I cannot tell"

Abstention means no value was produced because the evidence did not support one: an unclear unit, a figure axis that could not be calibrated, a table header that could not be read, or a range where a single value was needed. It is shown as work waiting for a person, not as a failure.

• An abstained extraction creates no candidate, so nothing enters the dataset with a made-up value.
• An abstained figure panel is listed with the reason.
• An abstained unit conversion keeps the reported value and unit and flags the case.
• An abstained vocabulary match leaves the mention unmatched and visible.

What this costs

Declining to guess means some values in most projects need a decision from a person. Conversions between concentration bases need a density many papers do not report, and a figure without a readable axis cannot be measured safely.

The alternative is a dataset that looks complete but contains numbers nobody can defend. A visible gap is worth more than a hidden mistake.