Glossary

Each question that the extension asks, and what each answer means. The model gives a probability for each answer.

The text sent to the model has more examples. It is in src/decide/questions.ts.

Answers

What kind of numbers are in the column?

measure: An amount that you can compare between rows: a price, a count, a rate, a size, or a duration.
identifier: A number that is a name: an ID, a code, a version, or a jersey number.
time: A year, a season, a date, an age, or a time of day.
rank: A position in a list: 1st, 2nd, 3rd.
bins: A bracket of values, such as ages 20–25 or 65+. Shaded at its midpoint, except in a first column that names the rows.
record: Two counts together, such as a win-loss record (12-4) or a score (35-10).
threshold: A limit that tells which rule or price applies, such as ≤ 272K tokens.

How do the labels in the column relate to each other?

ordered levels: Grades in a fixed order: sizes, tiers, severities, ratings.
unordered kinds: Names, kinds, places, or groups with no order.
workflow states: Steps in a process: open, in progress, done.
options: Alternatives. No option is more of anything: default, custom.
scores or records: Scores or records written as two numbers: 35-10, 12-4.
names or versions: Names or versions of products or models: iPhone 14, iPhone 15.

Is a higher value better or worse?

higher is better: A reader prefers larger values: home runs, battery life, revenue.
lower is better: A reader prefers smaller values: price, time taken, errors.
neutral: No direction is better: height, population, year.

Are ratios more important than differences?

yes: Yes. Use a log scale.
no: No. Use a linear scale.

Which of two labels is the lower grade?

a: The first label is lower.
b: The second label is lower.

Is this row a total?

yes: The row is a total, an average, or another summary of the other rows.
no: The row is one entry.

Do these columns in different tables measure the same quantity?

yes: Yes, in the same units. They get one scale.
no: No. Each table keeps its own scale.

Do you compare values down each column or across each row?

columns: Down each column. Each column is one measure.
rows: Across each row. Each row is one measure.

The table description

Each request includes a short description of the table. Practical Data makes it from the cells of the table: the page itself does not supply it. The model reads the description, then answers the questions. The description has these fields.

page_title: The title of the page.
table_title: The caption of the table, or the nearest heading above it.
rows_are: The first column where each value is different, with examples. It tells what each row is: a player, a model, a country.
columns: One entry for each column that the model is asked about. Code does not send text columns.
header: The column header, as on the page.
kind: number or label. Code decides this from the cells.
unit: The unit that code found in the cells, such as $, %, or km. Not sent when there is no unit.
examples: Five values from the column: the smallest, the largest, and three between them.
values: For a label column: each different label, in the order that it first appears.
spread: How wide the range of values is, in words. Code calculates it from the numbers.
notes: Hints about the form of the values. Code adds a note only when it applies.

Possible values of `spread`

values are within a factor of 2 of each other: The largest value is less than 2 times the smallest.
largest value is under 10 times the smallest: From 2 times to 10 times.
values span about N orders of magnitude (largest is ~X times the smallest): 10 times or more. Only then does the model get the log-scale question.
includes zero or negative values: A ratio is not possible, thus code does not calculate one.

Possible values of `notes`

the values are exactly 1, 2, 3, … in row order: The column counts the rows. It is usually a rank or a row number.
every value is a whole number between 1800 and 2100: The values can be years.
the values are spans of calendar years, such as 1950-1954: Most values are two years joined by a dash.
most values are written as bounds with ≤, <, ≥, >, or + rather than exact amounts: Half of the values or more are limits, not exact amounts.
most values are written as two numbers joined by a dash, such as 20-25: Half of the values or more are ranges. Code puts a range at its midpoint.
the values are win or loss streaks, counted as wins minus losses: W3 is 3, L2 is -2: Every value is a streak such as W3 or 2L, and the column has both wins and losses or its header says Strk or Streak.
values are written like clock readings (such as 2:41); they could be durations or times of day: Half of the values or more look like clock times.