Glossary

How often two labelers give the same label to the same record. The labelers can be two AI models, or a model and a person. Agreement is published with every frame and register figure.

How many distinct people posted about a subject, and the subject's share of each platform's posts, compared with its normal share on that platform.

The share of counted records that have an identifiable author. Article sources (news and papers) have none by definition, and deleted accounts have none, so they count toward records but not toward people.

A subject's normal share on a platform, calculated from earlier weeks. It never includes the week being measured.

What differs from the previous stored reading of the same subject, usually the previous week's, under the same method.

The facts a figure rests on: its population, coverage, author coverage, n, runs and agreement. They appear beside the figure, not in a footnote.

What AIDRAN captured from each source on each day, with backfills, gaps and outages marked.

Quoted records, each linked to the original: for every frame, and as a fixed sample of up to 10 records behind every number.

A value read directly from a record's own data, such as its platform, thread or date.

A recurring way people talk about a subject, with the stance they take. Frames are found by reading every post first and then looking for the patterns across them. They give a Reading's writer the substance and point of view to write from, and a Reading may also list them, each figure shown only where it clears its agreement bar.

A label given by a rule, an AI model or a person, such as a frame or a subject link. Every judgment names the run that produced it.

A one-line finding generated from a reading's stored figures. It carries its figure and conditions and links to the evidence. Once stored, it isn't edited; a correction is a new reading.

The number of records or people a figure is based on, always with its unit.

Where the conversation happened: which platforms, and which communities within them.

Exactly which records a figure counts: the platforms, the week they were published, and what was excluded (for example spam, or records not about AI).

Any item AIDRAN captured: a social post, a comment, a paper or an article. A count of posts says what it combines, for example a Reddit count of thread-starting posts and their comments.

A measurement of one subject over one week. It covers attention, place, frames, register and change, with the evidence and conditions behind every figure. Once published, a reading's figures never change.

The page where a reading is published. It can carry a dated notice when the reading is superseded or corrected. The figures on it stay as they were.

The emotional tone a record is written in, taken from a fixed list. Examples include analytical, anxious, skeptical and defiant. Register is reported within each frame, so a reading shows the tone each frame is expressed in.

The weekly publication. A release covers posts published during one week, has a number (for example 2026-W38), and contains the summary table and a reading for each subject in it. Until it's published, its page shows the week settling.

One labeling pass over a defined set of records, by a rule, an AI model or a person. It is recorded with the labeler, the model version or prompt, and the date.

What a reading is about: a company, lab, model, product, person, policy, event or claim.

The subjects in a release, ranked by how many distinct people posted about each one on a platform. Each platform has its own ranking, because people can't be added across platforms. Every count shows its author coverage.