Quality checks

Fixed rules, not an opinion

Quality checks apply fixed rules to extracted values. They do not judge whether a value is plausible science. They check that the value is complete, that its evidence is in place, and that its unit can be converted safely.

A value cannot reach the verified dataset by any route that skips these checks.

Three levels

Every flag has a name that says what to check.

• Error: something is broken, such as missing evidence, a value that is not a number, an unknown unit, a unit of the wrong kind, range limits the wrong way round, or a property outside the project scope.
• Warning: a decision is needed, such as no vocabulary match, missing required conditions, an unstated concentration basis, a conversion that needs a person, a percentage above one hundred, or a possible duplicate.
• Info: an observation that never holds a value back, such as a percentage change recorded where a measured value was expected.

Warnings hold a value back too

Open errors and open warnings both stop a value from joining the verified dataset. No setting accepts warnings automatically.

To resolve a warning, either fix the value or record why it is right as it stands. Both are kept in the record.

Two flags that are not errors

An unusual value is not an error. A reported value is never rejected for being unusual, since unusual values are often the point of a paper.

A possible duplicate is not an error either. Whether two values describe the same measurement is a judgement about the experiment, and a person makes it.

The usual case is a sentence in the Results that repeats a table cell. Both are read, and the flag says where the other reading is. Keep one and reject the other so the value is not counted twice, then run the checks again; if they really are two measurements, resolve the flag with a note saying so.