The categories answer a navigation question

Quantum Observatory’s topic labels help a reader narrow a large corpus. They do not attempt to place every work into one definitive academic discipline. A paper about compiling a decoder for a superconducting surface-code experiment can reasonably match software, error correction, and hardware. Removing two of those labels would make at least two relevant searches worse.

The classifier uses visible keyword dictionaries against titles and available abstracts. This makes the rule inspectable and repeatable. It also means the label reflects language in the metadata, not a full expert reading of the paper. A missed synonym can produce no label, while a phrase used in a different context can produce an imperfect match.

Why percentages can exceed one hundred

For each topic, the weekly share is the number of research works carrying that label divided by all research works in the same comparable week. Because the numerator sets overlap, adding the topic shares double-counts multi-topic work. A sum above one hundred percent is therefore expected and should not be presented as an error or a market allocation.

The same issue affects raw counts. Eight topic totals cannot be summed to recover the total number of works. The correct denominator is the separately grouped research-work total. Topic bars are best read one at a time or compared with the same topic in another period after coverage checks.

A rising label has a narrow meaning

The interface can mark a topic as gaining share when its proportion in a complete week is above its average proportion across the previous four comparable weeks and at least five works match it in the selected week. This identifies a change in the observatory’s tagged corpus. It does not establish that the topic is scientifically more important, commercially larger, or receiving more funding.

Vocabulary can also change. If a new term enters common use before the dictionary is updated, activity can be understated. If the dictionary changes, historical records are reclassified during a rebuild. Reproducible analysis must therefore record the classification version as well as the data snapshot.

  • Compare a topic with its own prior share, not with an exclusive-category assumption.
  • Open several matching records to check whether the phrase is being used as expected.
  • Treat an untagged result as unclassified, not irrelevant.
  • Record the topic dictionary version when publishing a derived chart.
  • Do not infer quality, impact, or funding from tag frequency.

Where human review belongs

Rules are useful for a browseable weekly index because they apply consistently and can be inspected. They are not sufficient for an editorial claim such as one research approach replacing another. That claim requires reading the relevant works, defining what counts as an approach, and checking whether source coverage or terminology changed.

The observatory separates these jobs. Automated tags organize the research library; editorial analysis states its selection criteria and limitations. Readers can follow every item to its source and can report a misclassification. This makes disagreement actionable without disguising a lightweight discovery tool as a scientific taxonomy.