The Agora Almanac

Section 07 / Methodology

How the almanac is computed

Every figure on this site is produced by program from the public archive of The Philosophy Forum. This page sets out the rules: what is counted, what is removed first, how the comparisons are made, and where the method runs out.

Short version: quotations are removed before anyone's words are counted, replies are traced by comment number, and no post text is published.

Source material

The almanac is computed from the public archive of The Philosophy Forum at archive.thephilosophyforum.com. The archive offers each member's posts as a plain-text export. 4,099 such exports, one for every account with at least one post, were read; together they hold 1,012,079 posts dated from 20 October 2015 to 25 February 2026.

Nothing else is used. No private messages, e-mail addresses or logs are involved.

Posts, threads, comments and members

Post
Any single contribution: either the opening post of a thread or a comment within one. There are 1,012,079.
Thread
A discussion with a title. There are 13,923. For 13,770 of them the opening post is in the archive; the other 153 were opened by accounts that no longer exist, so their comments are counted but their opener is unknown.
Comment
A post that is not an opening post. There are 998,309. A thread's size is its number of comments, and the members "commenting" in a thread are the distinct authors of those comments.
Member
An account with at least one post. There are 4,083, each with a profile. A further 16 accounts are placeholders for deleted members; see deleted and renamed accounts.

Cleaning text before counting

Posts are written in a markup that includes quotation blocks. Before any words are counted, each post is reduced to what its author wrote:

  • quoted passages are removed, including quotations nested inside quotations;
  • formatting tags, images and links are removed;
  • mentions of other members by handle are removed, since addressing someone is not using a word.

What remains is the member's own text. It runs to 103,799,093 words in all, an average of 103 per post. Word counts on member and thread pages are counts of this cleaned text. Post lengths are measured the same way.

Words, phrases and rates

Text is lower-cased and split into words. A possessive ending is folded into its base word, so that "Kant's" counts as "kant". A fixed list of multi-word terms, such as "free will", "hard problem" and "language game", is joined so that each counts as one unit whether written with a space or a hyphen.

Common function words are discarded using a standard English stop-word list, extended with conversational filler and forum residue. Words that are members' own handles are also discarded wherever the text usually capitalises them, so that word lists show ideas and not addressees.

The tracked vocabulary is the 4,500 most frequent remaining words that at least twelve different members used, together with the curated names and concepts described below. 4,705 of these have a page. A handful are left out of the published vocabulary: file-name residue from image uploads, common profanity, and variant spellings used fewer than twenty-five times.

Because the forum was busier in some years than others, most figures are given as a rate: uses per 10,000 words written in the same period (per 100,000 for thinkers, whose names are rarer). A rate lets a quiet quarter be compared fairly with a busy one.

Signature and distinctive words

Several pages list the words that set one body of text apart: a member's signature words, a thread's distinctive words, the words of a month or a year. All use the same measure, the log-odds ratio with an informative Dirichlet prior described by Monroe, Colaresi and Quinn in "Fightin' Words" (Political Analysis, 2008). It compares how often a word is used in the text of interest with how often it is used everywhere else, and discounts differences that rest on only a few uses. Lists are ranked by the resulting z-score.

For periods of time two further conditions apply, so that one member's hobby horse cannot define a year. A word can sign a year only if at least eight members used it that year and no single member supplied more than 35% of its uses; for a month the threshold is five members.

Member word lists leave out profanity, insults and a short list of charged terms. A tally of such words attached to a person's name is easily misread as a description of the person, which it is not. For the same reason the pages for those words report forum-wide figures only.

Subject families

The nine subject families on the topics page are hand-made lists of keywords. A family's count is the sum of its keywords' counts. The lists were chosen to be unambiguous, not complete, so a family shows the direction of a subject over time better than its absolute size. The keywords are:

Mind
consciousness, conscious, qualia, mental, brain, brains, neuroscience, subjective, subjectivity, phenomenal, awareness, physicalism, dualism, panpsychism, hard problem, cognition.
God
god, gods, religion, religions, religious, faith, theism, theist, theists, atheism, atheist, atheists, christianity, christian, christians, jesus, bible, church, islam, muslim, muslims, scripture, divine, theology, afterlife.
Ethics
moral, morality, morals, ethics, ethical, virtue, virtues, evil, duty, obligation, obligations, justice, injustice, utilitarianism, utilitarian, consequentialism, deontology, deontological.
Politics
politics, political, government, democracy, democratic, capitalism, capitalist, socialism, socialist, communism, communist, trump, biden, election, elections, republicans, democrats, liberals, conservatives, voters, fascism.
War
war, wars, russia, russian, putin, ukraine, ukrainian, nato, israel, israeli, gaza, hamas, palestinian, palestinians, military, invasion.
Physics
physics, quantum, universe, particle, particles, relativity, cosmology, spacetime, entropy, photon, electron, einstein, big bang, general relativity, many worlds.
Logic
logic, logical, semantics, semantic, proposition, propositions, predicate, sentence, sentences, grammar, syntax, inference, valid, validity, axiom, axioms, theorem.
Machines
ai, chatgpt, llm, llms, gpt, computer, computers, algorithm, algorithms, artificial intelligence, machine learning, robot, robots, chatbot.
Pandemic
covid, pandemic, vaccine, vaccines, vaccination, vaccinated, virus, lockdown, lockdowns, masks.

Thinkers

The thinkers table counts a curated list of 168 names. Each name is counted through a short list of word forms, for example "kant", "kantian", "kantianism" and "kantians"; every thinker's page lists the forms that are summed.

A word form is counted whatever it referred to. A surname shared with someone else, or with a place, is therefore counted in full, and where that matters the page says so. A few names can only be counted as a full phrase ("William James", "David Lewis") and are undercounted as a result. Three names on the original list are omitted altogether because the word form mostly means something else on this forum: "scotus", which is usually the United States Supreme Court, and "boethius" and "paine", which are also the handles of prolific members.

A mention is not an endorsement. The count cannot tell agreement from attack.

Replies and the network

When a member quotes another or uses the reply button, the forum inserts a tag naming the member answered and, usually, the number of the comment answered. The almanac resolves each tag to an author by the comment number first, because numbers survive changes of username, and by name only when no number is given.

Of 1,217,480 tags, 1,173,050 (96.4%) were attributed to a known author. 28,187 point at accounts that no longer exist and 16,243 are members quoting themselves; both are left out.

The network is built from the 400 most prolific members. Two members are tied if they exchanged at least three replies. Clusters are communities found by the Louvain method on those reply counts: groups whose members answer one another more than they answer outsiders. Each cluster's name is editorial shorthand chosen to summarise the words its members use unusually often; the clustering itself uses reply counts alone. Belonging to a cluster says whom a member talked to, not what they think.

Related threads, related words and closest vocabulary

Threads on similar ground share distinctive vocabulary. Each thread is represented by up to twelve of its most distinctive tracked words; two threads are related if they share at least two, weighted so that rare words count for more than common ones.

Words that keep company are tracked words that are distinctive of the same threads, in at least three of them.

Closest vocabulary, shown for members with at least 50 posts, compares members' word profiles by cosine similarity. It finds people who write about the same things. It does not find people who agree.

Dates and part periods

All dates and times are in UTC, as the archive records them. A "day" is a UTC day.

The archive begins on 20 October 2015 and ends on 25 February 2026. October 2015 and February 2026 are therefore part months, and 2015 and 2026 are part years. They are marked as such, and rankings of "full" years or months leave them out.

Deleted and renamed accounts

Some members deleted their accounts before the archive was made. The archive handles them in three ways, and so does the almanac.

  • 16 placeholder accounts stand in for deleted members. The largest, "Deleted User", merges many people; its posts survive only as a notice that the text was removed. Placeholder posts count toward forum totals and thread sizes, but placeholders have no profile, appear in no ranking and are not part of the network. They account for 13,033 posts.
  • Some deleted accounts left no posts at all. Replies addressed to them cannot be attributed and are the "tags left out" on the network page.
  • 153 threads lost their opening post in this way. They are listed with an unknown opener and dated by their first surviving comment.

Members who changed their handle appear under the name the archive now shows. Replies are matched by comment number, so earlier exchanges under an old name are still credited to the right person.

Names and characters

Member names and thread titles are taken from the archive. They are shown here in plain Latin letters: accented letters lose their accents, typographic quotation marks and dashes become plain ones, and characters outside the Latin alphabet are dropped. Where this changes a member's name, the profile says so. One member whose name could not be represented at all is listed by member number.

What is deliberately left out

  • Post text. The published data holds counts, member names, thread titles and single words. No passage of anyone's writing is reproduced.
  • Hour-of-day profiles of individuals. The forum-wide pattern is shown on the timeline; per-member patterns are not published.
  • Per-member tallies of profanity, insults and charged terms. See above.
  • Anything about identity. No attempt is made to link a handle to a person.

Checks

The analysis was implemented twice, independently, in two programming languages. The two implementations agree exactly on the number of posts (1,012,079), the number of words (103,799,093), the size of every thread, the totals for every member, and the count of every one of the tracked words.

Limitations

  • The archive is the only source. Whatever it omits, the almanac omits.
  • Counting words does not measure meaning. Irony, quotation without quotation marks, and words with several senses are all counted at face value.
  • Keyword families and the list of thinkers are hand-made and reflect editorial choices.
  • Small numbers are noisy. A signature word for a member with a few dozen posts is weaker evidence than one for a member with thousands.
  • Volume is not merit. Nothing in the almanac measures the quality of an argument.