Symposium

The app opens with empty chairs, on purpose — nobody is chosen for you. Three ways to fill them, and they are different offers rather than one offer three times.

Two to five chairs. You can seat someone mid-conversation with draw up a chair — he arrives into the argument in progress and answers that, not a fresh question.

Sixty curated arguments is a good default and a poor ceiling. When a table turns out to be the one you wanted, keep this table saves who was in it and the questions you want to ask them — during the argument, not only after it, because the evening a table proves itself is the evening you will remember to save it.

Your tables can never shadow or delete a curated one: they are kept apart in the app's own storage and carry their own identifiers. Keeping the same men twice revises the table you already have rather than leaving you two rows you cannot tell apart. A table that names somebody no longer in the roster is withheld from the menu rather than laid short.

Type it, or hold the microphone and speak it. Open the sentence with a name and it goes to that man first — “Marcus, does the fort hold?” — and the rest of the table decides for itself whether to join him.

Everyone at the table is scored on how strongly your question touches what they care about. The strongest reaction leads. The man most opposed to him speaks second — friction is the design, not a side effect. A third joins only with a genuinely different angle.

Each contribution arrives in two or three turns: a short opening that states his answer, then the explanation. That is not a style choice — every turn is spoken the moment it is written, so a short opening reaches you while the rest is still being composed.

Four things you can ask for that a plain question will not produce.

Tap your own portrait and you have asked to speak. The man talking finishes his thought, nobody new begins, and the label under you goes waiting → your turn. Holding the microphone does the same thing and silences the room at once.

For a sitting you cannot attend to — running, driving, cooking. Choose let them talk by themselves before you ask, then ask however you like. The table takes it from there and the toast closes it. Nothing needs tapping again.

It quickens near the close, whatever it started as, so the last round is the end of a movement rather than a sudden change of manner. And it stops early when the argument runs dry — a twenty-minute setting ending at eleven is the table having finished, not a fault. take the floor ends it at the next round boundary.

toast closes the sitting. The table finishes what it was already saying first — the button says letting them finish… while it drains — and then raises a glass.

Carrying one on rebuilds what the philosophers were told: the question, the rounds in the order they happened, and every interjection of yours between them. What is not rebuilt is how strongly each man reacted — those numbers decide who leads a round, they are spent by the time it is over, and inventing them would read as evidence of something. A sitting merely reopened stays a recording, and the screen says which you are looking at.

The sliders icon in the top bar → Kept sittings. Everything you kept is there, newest first, each row showing when it was, how many lines it ran to, whether it was toasted, and who was at it. Reachable between sittings and from the toast — not mid-argument, because opening one replaces what is on screen.

Both land in Files → On My iPhone → Symposium, so an evening can be moved to a computer, backed up, or handed to somebody. Carrying one on needs the archive, so keep it if you think you might want the argument again — keep the whole sitting writes both, and save the voices writes the audio separately.

What they remember

They have met you before

Each sitting is recorded as a gist and read back at the next one, so a philosopher can say “when you were last here you doubted…” — and will not reuse an image or an opening he has already spent on you.

A steelman is recorded as a case put, never as a view held. He argued it because you asked. The same is true of a movement — see below.

How hard they press

Cordial, candid, adversarial

One dial in the settings menu, and a different question from whether anybody's mind changes. A man can be immovable and courteous, or open to persuasion and merciless.

Never in the opening round, whichever way it is set — two men are stating positions and there is nothing yet to press on. Change it mid-argument and it takes effect at the next round.

When somebody gives ground

They can be moved

A table where nobody ever moves is not a symposium; it is five lectures taking turns. So something said in the room — by another philosopher, or by you — can change what a man says next, and the turn is marked when it does.

It is never remembered as his position. He granted something to somebody on one evening; that happened, and you may refer to it. What does not survive is the reading that he now believes it — a philosopher rewritten one concession at a time would drift out of himself over a season of sittings, and nobody would see it happening. It is also rare on purpose: a man who concedes something every round is not open-minded, he is furniture.

The rules they cannot break

Why the citations can be trusted

Every integrity rule here is enforced in code, not merely requested in a prompt. A turn that breaks one is dropped before it can be spoken.

Thirty-five philosophers can be quoted exactly today, across 121 bundled texts: Aristotle, Augustine, Bacon, Berkeley, Comte, Confucius, Darwin, Descartes, Diogenes, Dostoevsky, Douglass, Einstein, Epictetus, Hobbes, Hume, James, Julian, Kant, Leibniz, Locke, Marcus Aurelius, Marx, Mill, Nietzsche, Plato, Pythagoras, Rousseau, Russell, Schopenhauer, Seneca, Socrates, Spinoza, Spooner, Tolstoy and Whitehead. For everyone else, quotation is refused rather than guessed — and you can change that yourself for any of them, by adding a source and fetching it.

Your own material

Giving one of them a source

On the table screen, give one of them a source → the philosopher's name. Then a title, the link, optionally a note on where in it or why it matters, and whether it is a talk or video or something written — and Add it to his library. The sheet lists what he can already cite underneath, so you are not adding a second copy of a book he has.

He may cite it from the next round on, and the control is reachable mid-argument, which is usually when you want it: the moment to hand somebody a source is the moment he has said something it bears on. Swipe a source in that sheet to take it back out.

Video transcripts work well, and are worth knowing about: paste the link to a transcript rather than to the video, and the talk becomes as quotable as a book.

This is the one place the closed library opens, and it is worth being plain about the trade. The researched library is a promise the app makes: somebody checked that the man wrote the book and that the title is the title. A source you add carries no such promise — what it carries instead is an address, which is a stronger guarantee than most footnotes offer, and the reason it can be let in at all.

The roster

A hundred and sixty-one, in twelve schools

Grouped by where a man stands rather than when he lived. The groups exist to make a menu navigable, not to settle historiography.

Forty-four of them are alive. Those carry a standing disclosure: a living person can be got wrong in public, by name, and be around to see it — so the screen says plainly that this is a construction from published work and that he has not been consulted.

Two Simones, three Thomases

Calling a man by name

Where a given name is shared, everyone who shares it takes a surname initial — Karl M and Karl P, Simone W and Simone B, Thomas H, Thomas A and Thomas K.

The bare name reaches nobody, deliberately. Guessing between two men is worse than asking again, and you would hear the wrong one answer with no sign that anything had gone wrong.

Spelling is no help. Georg Cantor and George Berkeley are one name across a table, and dictation picks a spelling for you — so neither given name reaches anybody while both are seated, and each is addressed by surname. A name a letter or two out still finds its man (Larkin reaches Larken Rose), but never when it could find two.

Voices

Who sounds like whom

The phone's own voices work with no setup and no cost. Cloud voices are better and are billed per character; the server holds the key and enforces a daily budget, so the system voices take over rather than the table going quiet.

Give your name and they will use it — sparingly, once or twice an evening, at the moment a man turns to you. A name in every turn is the register of a hold queue and would undo the thing it was added for.

When something goes wrong

What the messages mean

The same table, in a browser

The web table

Everything in this manual so far describes the app on your phone. There is a second way in: a web page, at a URL you can hand to somebody else. The philosophers, the citation audit and the rules of a round are the same code — the server and the phone share one definition of who these men are, so nothing drifts between them.

Two differences are worth knowing before you open it.

Sixty-five men, in thirteen schools. The web table seats the dead only — five per field — while the phone carries all a hundred and fifty-nine. The living are held back until their voices are a settled question. The web also has a school the phone does not: Waking up, for Watts, Dyer, de Mello, Abhinavagupta and Yogananda.

The floor is the whole page. There are no sheets and no toast: a sitting is a transcript you can read, save, and open again.

One thing the web shows that the phone does not. When a philosopher returns a turn with no words in it, the phone records the fact and says nothing; the web prints it — "reacted 0/10 but returned no words, so he stayed silent this round". It is not a fault you need to do anything about. It is the audit being visible.

Whose credit pays

Your own key, and nobody else's

Open the settings wheel and paste an Anthropic key. Without one the table cannot answer — the roster, the curated tables and the portraits all work, but nobody speaks.

The key is kept in your browser and sent with each question. The server uses it for that one round and never writes it down. So when you send the link to a friend, he pastes his own key and his rounds bill to him. Yours are never spent by anybody else.

An ElevenLabs key is optional and works the same way. Without it the philosophers are read aloud in the browser's own voices, which is free and worse. With it they are cast from your own voice library, exactly as on the phone.

Forget my keys clears both from the browser. Worth pressing on a machine that is not yours.

The voice casting carried over from the phone is offered as a starting point and is read-only on a shared table. Voice ids belong to whichever ElevenLabs account made them, so your changes stay in your own browser rather than overwriting somebody else's cast list with ids his account has never heard of.

The sidebar

Laying a table on the web

A curated argument lays one of the twenty-seven tables whose men are all on the web roster, and puts its opening question in the box.

Seat me a table draws four strangers — with one rule: the second man never shares the first man's school. Four at random produced tables that agreed with themselves.

An evening with… draws one man and asks his own question for you. One chair is the only arrangement where somebody gets to be long: no floor to yield, nobody waiting. It remembers the last twenty it drew, so an evening feels like meeting a person rather than reshuffling six names.

Clicking anyone on the roster draws up a chair for him, mid-argument included — the conversation carries on with him in it. Clear the table ends the sitting and empties the floor.

The moves

Pushing the conversation

Hear one of them out. One man, at length, with nobody interrupting — his own account of what he thinks, what drove him there, what he is against, and what he never settled.

Let one question me. He asks; you answer. Most of the work here is restraint — left alone the model writes a question with the answer folded into it, which is a lecture wearing a question mark.

Steelman. Two choices: who argues, then whose case. The advocate puts the other man's view as well as it can be put, takes one aside of his own, and then the man whose case it is says whether it was put fairly. A steelman nobody is allowed to judge is a performance.

Let them talk. Nobody prompts; the table carries on under its own momentum. An empty round is the honest end of it.

Seat an opponent. Somebody on record arguing with the man you pick, and the quarrel is named — Schopenhauer, at "What the will wants". When the record has nothing, the offer is billed as what it is: a draw from another school.

Keep this table. Remembers the cast between visits. One table, not a library of them.

Typing a name works too. "Marcus, wouldn't this make you a slave to your own habits?" is an address, not a question to the room, and the man named answers first whether or not the scoring would have picked him.

A hand goes up

Who is waiting to speak

When one philosopher turns to another by name, the man addressed gets a badge on his coin carrying his place in the queue. Two hands up is an order you can honour rather than a guess.

The hand comes down when he speaks. If a round ends with his hand still up, the question is still owed him — click his coin and he takes it up.

Clicking a raised coin takes his hand rather than excusing him. His portrait is marked because he is waiting; a tap that removed him from the table would be the opposite of what the badge invites. Coins without a badge still remove.

The wheel beside the table is thinking keeps turning between turns, not just before the first one. A philosopher may answer and the next may still be a whole request away.

Taking it with you

Saving a sitting

The transcript, as text. Built from what is on the floor rather than from the bare history, so the citations and the audit's own notes come with it. A transcript that quietly dropped what the audit found would be the wrong document for this app to produce.

The sitting, to reopen. A file, not a slot on a server. The server keeps nothing about anybody, which is what makes the URL safe to hand out — and a sitting you can mail to someone is a sitting that survives the table being cleared.

The sitting, as audio. One file of the spoken lines.

A reopened sitting is spoken again. Reopening brings back the words, not the voices — those were made in a browser that has since forgotten them. So saving audio for one speaks whatever this browser never spoke, in the transcript's order, and tells you what that will cost in characters before it starts. It used to hand over whatever happened to be in the buffer, which for a reopened sitting meant a twenty-five minute evening saved as three and a half seconds — a file, not an error, which is the worse of the two.

Audio needs an ElevenLabs key, and says so plainly when there is none. The browser's own voices cannot be captured — they render straight to the speakers and no interface hands you the sound — so rather than save a file full of silent gaps and let you discover it later, the button explains itself.

Before it is filmed

Changing what somebody said

Your own words you may fix. On the web, hover your question and press edit; on the phone, hold it and choose Fix a typo. Both mend the words on screen and in the transcript together, so the film is voiced and captioned from what you meant rather than from what the dictation heard. It is worth doing before anything is voiced: a slip left in is spoken by ElevenLabs, animated, and burnt into the subtitles.

Theirs you may not. A philosopher's turn cannot be retyped where it sits, on the web or on the phone, and that is deliberate: a line altered in place would leave no trace of having been altered, and the whole point of the table is that what a man said can be checked against what he wrote. The host is not a philosopher — he cites nothing, quotes nothing, and no audit runs on him — which is why the pencil is drawn on his lines and on nobody else's. The words live in the saved sitting, and editing that file is the edit.

Edit before you record, not after. The text drives the audio, the audio is cut into the clips, and the subtitles are timed against the finished film. A word changed after any of that has happened means doing all three again.

One limit, and it is the only one. Rewording what a man says about his own life is ordinary editing. Putting a book in his hands that he did not write is not — it breaks the single promise the table makes, and it breaks it invisibly, because a citation nobody checks reads exactly like one that holds. If an edit touches a work, verify the work.

From an evening to a film

Making an episode of it

A sitting can be cut into a film — the philosophers animated, speaking their own lines in their own voices, with subtitles and a works-cited card. The cutting is done in Anima Studio, a separate app on the desk; Symposium's part is to hand it an evening it can read. Which evening, and which app saved it, decides the first step and nothing else.

From there the two paths are one. The lines are voiced one file per line, the long ones cut at their own silences into pieces short enough to animate, and each piece becomes a shot of whoever is speaking. Anima Studio rotates through the eight framings each philosopher has, so a man with fourteen lines is not fourteen identical shots. The opening and closing sequences are kept between episodes and place themselves.

A line the round could not parse is written as (malformed) and has no words to caption — the audio may exist, voiced by hand afterwards, but the only copy of what was said is in the host's head. Those, and anything cut in by hand at the end, are supplied in tools/caption-fixups.json and the film tools take them as read.

Something for them to work from

Adding information to the table

Giving one of them a source hands one man a link, and the asymmetry is the point: a paper put in front of Dennett that Teresa has not seen is a move in an argument. Add information, on the web table, is the other thing — an article, a passage, an excerpt that the whole table has read, so the evening argues against something rather than from memory.

Give it a name — it is cited by that name — say where it came from, and then paste the words, fetch a page, or open a file. Read what lands in the box, cut what you do not want, and put it on the table. Every philosopher seated may refer to it, disagree with it, and quote it, and every quotation is checked against those words before the turn is heard. Click a piece's name afterwards to read it again.

Where it all lives

Three places, and losing any one of them loses nothing.

Only on a table you run yourself. The words are kept on the server, so this exists only when SYMPOSIUM_SAVE_DIR is set — the same gate the save buttons use, and for the same reason: a server holding writable state for whoever has the URL is a gift to a stranger. On the deployed table the panel says so instead. SYMPOSIUM_LIBRARY_DIR moves the library if you want it somewhere other than beside your sittings.

The ten steps, in four movements

Making an episode with one tool

tools/make-episode.js runs the sequence below for you, in four phases. With no phase named it reads the folders and says where the episode stands and what the next step would be — nothing is remembered in a state file, because a state file is one more thing to fall out of step with the folders it describes.

It does not cut the sitting, and will not. Choosing which lines make the film, and in what order, is the editorial act — episode 003 is 26 lines of 73 — and dangling.js exists because selection breaks conversations invisibly. Cut it yourself, then hand the result over.

The same evening, from archive to upload

An episode, in order

The section above is what happens; this is what you type. Every tool shows what it would do and writes nothing until --go, so the first run of each is free to be wrong. Anything touching a film's timeline reads ANIMA_PROJECT — set it once per terminal and the rest of the evening knows which episode it is working on.

One archive, in the shape the film tools read. Which command depends on where the evening was recorded, and on whether it happened in one go.

A morning's sitting runs longer than a film should. Write the edit list by hand — movements with titles, and the line numbers each keeps, in the order they should be heard — and the tool does selection and order and nothing else. A number used twice, or one past the end, stops the cut.

The report is deliberately noisy — it reports suspicion, not proof. Reading six lines beats re-cutting a film.

It looks for somewhere each cited work can actually be read and saves only what it could verify. What it cannot verify it says so about, rather than guessing a link that later turns out to point at nothing.

One file per line, evened out so four ElevenLabs voices sit at the same loudness, and lines over 28.8 seconds cut at their own quietest moment into parts the animator can take. Run it once without --go first: it prints the character count, which is the bill.

A key lives in ~/.elevenlabs-key, or in ELEVENLABS_API_KEY — and an environment variable that is set wins over the file, which is worth remembering on the second attempt. Never paste a key into a chat window.

A philosopher who never appears on screen needs no sheet. John of the Cross is discussed for a whole episode and is not at the table.

In Anima Studio: Voice & Dub → the sixth card, whole sitting → reel. Point it at the episode's Audio-Chunks folder, fill the sheets, and generate. It skips any line already shot, so a second pass costs only what is missing. It appends — clips made in a later pass land at the end of the timeline, after the outro, rather than where they belong.

That reads the timeline against chunks.json and names four kinds of trouble: a line clipped twice, a line with no clip at all, clips out of order, and a clip whose length does not match the audio it claims to be. It infers the reel's lead-in from the clips themselves rather than assuming it, because assuming it once produced forty-seven faults that were not there.

If anything is out of order — and after a second pass it will be — quit Anima Studio and sort it. The clips know where they belong; each carries the line it was made from, so this is sorting rather than guessing.

Quit Anima Studio first — it holds the project in memory and writes its own copy back over yours the moment anything changes.

Every second turn per speaker, the first turn always shown, and the host named once at the top and then left alone — he is the voice the viewer has been listening to since the first second, and naming him four times is a caption on the wrong thing. --every 3 thins it further; --pos upper moves it off the lower third.

The opening and closing sequences are kept between episodes and are dragged in from Bookends. After a film is finished, cut the next episode's copies from it:

Export from Anima Studio, then the three commands that turn a render into a published episode. The subtitles are not transcribed — every shot knows which line it was made from, so only the timing is measured.

Ask for the YouTube Studio video details and the whole block comes back in one piece: the title as Symposium 003 — Teresa of Ávila, Daniel Dennett…: "the question", and a description opening with the website, then the chapter timestamps, then the works referred to, then the full bibliography.

Upload your own subtitles rather than letting the automatic captions stand: they mishear every name at the table. The video file itself cannot be replaced after upload — a re-cut is a new video.

One thing to know

The roster ships with the app

Philosophers, curated tables and bundled texts are compiled into the app itself. There is no database and nothing downloads in the background. Anyone added since your last install is not there until you build again — which is worth remembering the day a search comes back empty for a man you know exists.