Plate

Plate is the list of what your sources have put on you. Nobody assigns anything here and nothing is typed in: an item opens because a message asks something of you, or you asked something of someone, or someone promised you something, or a source records a commitment or a deadline — and it closes when a later source shows the thing was done, or when you say so. Every row keeps its receipts.

It is its own page (Plate in the sidebar) because it is not a day: an ask from last Tuesday is still on your plate on Friday. Today shows the three that need you most in the strip near the top of the page; Plate shows all of them.

What opens an item

A small model reads each new source once — once for everyone who can read it, never once per person — and says what it creates between the people on it: who asks whom for what, who promises whom, what two of them agreed. Your plate is then worked out from that without another model call: what you were asked, what you asked, what was promised to you. Someone copied on the same mail who is not part of it gets nothing on their plate, and nothing of yours — not your other mail, not your plate, not your rules — is ever part of what the model reads about a shared source. The question it answers is: does this put something on someone's plate? It is not given a list of kinds of mail to accept or refuse — a list is always one case short, and a receipt written in a friendly voice would pass it. It is told what an obligation is, and reasons from the signals the message carries:

• A specific person — or someone acting on a matter you are party to: an application you made, an introduction someone made for you, a case, a deal — asks you to do, decide, send, confirm, book or answer something concrete; or commits to do something for you; or you ask that of someone, or send an organisation a form, a claim or a request you are now waiting on.
• It is written to you as an individual, or by you to one party — not broadcast to a list, not generated by a system, not a template sent to everyone.
• The thing is concrete enough that you could act on it — or wait for it — and know when it is done. Choosing or confirming a time from a link counts: you have to pick.
• Nothing counts merely because a sentence is imperative, says "please" or ends in a question mark. A receipt, a notice, a status update, an offer, a reminder of something already arranged, a hope or a courtesy is not an ask however it is phrased, and a sentence quoted or forwarded from someone else is not the sender's ask. A question asked and answered in the conversation itself — "walk me through how you'd handle it", answered there and then in the meeting — is talk, not an item; only what is still owed afterwards ("send me your write-up by Friday") is. A personal or social matter between you — the link to a restaurant for tonight, photos, whether someone is free this evening, a drink after work — is never an item, whether it was asked or promised: it stays in your sources, where Ask finds it, but not on your plate. Anything about the work you share is work, even over a meal ("book dinner with the investors so we can close the round"). Nor is an ask that applies only if something else happens first ("if you speak with Ravi, ask him…"). A target said only as a question and answered with a hope ("funding by the end of the year?" — "hopefully") is not agreed. What the sender wrote themselves still counts wherever it sits: under the quote, between its paragraphs, or beneath something they forwarded ("alex, can you send it over to Ola?").

The signals it reasons from are who wrote to whom (a named person, or a role, a brand or a system address), whether the message replies in a conversation you are already part of — an earlier message in the thread that you sent or were copied on, which is how a reply to an introduction a colleague made differs from a stranger's first approach; a message that itself says you already have something with the sender, such as a trial, an account or an order, is not a first approach either — whether you are the sole addressee, whether the text is written to one person, and what would count as the thing being done. Nothing it says is trusted on its own:

• The ask it names must be a sentence that appears verbatim in the source. That sentence becomes the row's title — "Please supply a phone number for your order." — never the subject line, with one exception: a form or request you sent whose body only says what is attached may be titled by its subject line, verified the same way.
• The direction must match the envelope: an ask of you, or a promise to you, comes from a message addressed to you; an ask you sent comes from a message you wrote. Every address your own mailbox answers to counts as you — the aliases you send as, not only the address you sign in with — so mail to the address you applied for something with is on your plate like any other. They come from four places, in this order: every address verified on your account with your sign-in provider, which already proved you control it; the addresses your connected mailbox sends as, read from its own settings; the Google account your own Drive or Calendar is signed in to (it marks mail as written to you, and never proves the address is yours to anyone else); and any you add under Tune Today → Your other addresses, for the case none can state — an address that reaches you by forwarding from somewhere else is in no send-as list and on no account. Never from who a message happened to be delivered to: a group address your whole team reads is not you. Saying an address is yours changes what is called written to you, never what you can read.
• A message with no address behind it — a channel message — is written to you when it calls you by your handle (@ and the part of your address before the @), or by a name of yours nobody else the workspace knows shares — colleagues, and the people its connected apps know by name though they have no account here: your full name, or your first name when nobody else has it. With two Alexes, "Alex, can you…?" is neither's; "Alex Morgan, can you…?" is still Alex Morgan's (written as a mention, "@Alex Morgan", too), and never Alex Chen's because they share a first name.
• A message in a direct message or a small channel — 8 people or fewer, as the chat app itself lists its members, kept current as people join and leave — that is said to the channel as a whole is written to everyone in it: "I'll send the deck tonight" there is a promise to each of the others, and "can someone book the venue?" an ask of each. A message that names or @-mentions someone is that person's alone — a name in it means the one member of that channel who has it, even if someone elsewhere in the workspace shares it, and someone who isn't in the channel is never read as the channel. The sender's own message is never "to" them. Being on a channel's member list never lets anyone see it: only people who can read the channel ever get anything from it.
• Who sent a chat message is known by the chat app's own id for them, linked to your account when you connect it — not by their name. A message you sent in Slack is yours on your plate ("you asked"), and your own reply in its thread closes an ask made of you; somebody who happens to share your name is never taken for you, and the same holds for a commitment or deadline recorded from a chat message. An @-mention is known the same way: "@Dana Lee" in Slack is whoever Slack's mention points at, linked to their account — never a colleague who shares the name. A name written without an @-mention, or in a message that records no mentions — sent before Louvain recorded them, a bot's mention, one past the fiftieth in a message, or a name too long to record — is read by the name rule above.
• A due date is kept only when the model also quotes the words it read it from ("by Friday", "before the 20th"), those words are in the source, and the date is on or after the message. The phrase is shown beside the date so you can see where it came from. A due date is a day, not an instant: something due by Friday reads as late only once Friday has ended in your browser's time zone, never during it. The stored value remains a calendar-day marker; travelling does not rewrite the deadline. API clients can supply timeZone; without it, your account's time zone is used (the one your browser last reported), and UTC only when it has never reported one. The same day is used by the AI-assistant tool and the due reminder email; the Slack plate post uses the zone on your Slack profile. A moment with an explicit offset remains an absolute instant.
• Only a certain judgement opens a row. A row that should not exist costs more than one that was missed; a missed ask is still on Today and in Search.
• Mail from a no-reply, notifications, billing or payments address never opens an ask or a promise, whatever it says — and whatever your rules say — with two exceptions. A service asking you to take a step only you can take on your own account — verify your identity, confirm your email address or finish setting up an account, sign its agreement, fix a payment method or renew something before it stops — is on your plate as something to do; a step the account already works without (a setting you may change, an extra you may add) is not. And a notification that carries a named person's own request of you — a comment that mentions you and asks you something, a task assigned to you with the assigner's request in their words — is that person's ask, shown as from them. A receipt, a newsletter, "your statement is ready", a sign-in alert, an announcement, an offer, a password reset or sign-in code you asked for yourself, and a notification with no one's request in it (a document shared with you, an assignment or review request on its own, a status change, a build result, a daily summary) never is. In Gmail, neither does mail its sender declares as a mailing or an automated notice: a mailing list's headers, an automatic-reply marker, or Gmail filing it under Updates or Promotions — unless it asks one of those steps on your own account or relays a person's request. A colleague's mail is never read this way, even when it comes through a Google Group, and neither is a reply in a conversation you took part in. Outlook mail is not yet read for these declarations. Mail Gmail files under Updates or Promotions is kept for search, and unless it reads as a receipt or a notice about your own account, nothing is learned from it: a date it says something of yours lapses does not reach the plate as a deadline. A failed payment or an expiring certificate in a receipt-like notice still can.

A promise made to you — "we will give you our feedback as soon as we can", written to you by the person who will do it — opens under Waiting on others, beside the asks you sent, because you are waiting on it in the same way. The row shows who it is from, and it closes when that person writes again in the thread, or when a later source says the thing arrived.

Tune what counts

Your plate has a second set of words beside your Today priorities: what should and should not count for you, in your own words — "recruiters never", "anything from my immigration adviser always", "forms I send to government count as waiting on them". Tune what counts on the plate's lead card opens the same dialog as Tune Today, on that field. Louvain reads your words once, when you save them, and shows you what it understood underneath the field (Louvain reads this as: …) — if it read a rule wrongly, rewrite it. If it could not read them at all, the field says so and Louvain tries again on a later check. A rule can keep a kind of item off your plate (who it is from, what kind of sender, a first approach from someone you were not already talking to), or exempt something from your other rules ("anything from my adviser always"). It can never put something on your plate that the model did not find, make a receipt, a notice or a broadcast into an ask, or add words to a source — every check above applies whatever the rules say. Saving rules changes your next plate check and nothing else; Today's ranking is not redone.

Commitments and deadlines your sources record can open as rows in To do, but only when the same model read says the fact is your obligation. The test is what happens if you do nothing: when something of yours then lapses, expires, stops working, is missed, renews or charges you, it is yours — whoever or whatever stated the date, a colleague, an automated notice or a bulk mailing. A certificate expiring, a payment falling due, a registration that must be renewed and a form you must file are all yours. The doing has to be yours, though: work you have asked someone else for and are waiting on is not a record of yours, and neither is a statement that something already happened, someone else's deadline, an invitation and its time, a launch or release date, or the closing date of an offer, a rate or an event you have not taken up. Where the thing happens by itself and asks nothing of you — a plan that renews on its own — it does not open; where the automatic thing has failed and now needs you, it does. The fact's own words must appear in the source, and only a certain judgement opens it. Some facts never reach the model at all: nothing from a calendar event (an event's date is when it happens, not when something is due — a birthday is not a deadline), nothing from a recurring event, and no deadline more than six months out. When a recorded fact sits in the same thread as an ask, the ask carries the date and there is one row, not two.

What closes one

A recorded fact's row does not track a sentence; it tracks the obligation — what kind of thing it is (a deadline, an open commitment) and who and what it is about. Whether it still stands is worked out afresh from everything you can read, every time your plate is checked, and only a closing event closes it: a source saying the commitment was fulfilled or withdrawn, the deadline met, the work completed. Three things follow, and each is deliberate:

• A source going quiet closes nothing. A deadline nobody has mentioned for a year is still a deadline. Age is not a contradiction, so an obligation whose newest evidence is old still stands, and the row says when it was last stated.
• An ending you cannot read never closes your row. What stands is worked out from the sources you can read, so something stated in a room you cannot see ends nothing for you.
• Reading a source again does not churn your plate. Re-reading a mailbox — after a configuration change, or when an operator rebuilds — produces new readings of the same sentences. Because a row tracks the obligation and not the reading, it stays as it is rather than closing as "superseded" and reopening as a duplicate beside itself.

Several sources naming one due date are one row, not one each. A promise is different: it is the promise that message carries, so two people owing you the same thing are two rows.

Any row also closes when a later source you can read says the thing was done — a Slack message in another room saying you sent the number, a note that the order was cancelled — in whatever words it uses: "sorted the parts thing with the moto guys" can settle a parts order it never names, and a confirmation that names the interview you were asked to book settles the booking ask. A message saying the step was taken and only someone else's answer is pending — "application for incorporation submitted, should hear back in a few days" — is offered rather than closed, and so is a shared file. The order things arrive in does not matter: an item that turns up late is still read against what was said after it. Three things are checked rather than trusted: the quoted sentence really is in that source, the source came after the item, and who wrote it. A list of what is still to do — action items, next steps, reminders — never closes anything, whoever posts it: only a message saying the thing happened does. But a line of it that gives one of your open items a date dates it: Tod's "Order the oak boards – Fri 9 Oct" under his name sets that day on the oak-boards item, and its history says Dated by Tod. Only you or the item's other person can date it — anyone else's date is ignored — the date's words must be in the message and on or after the day it was sent, and the latest date said wins. The other person's copy of the same item takes the same date, if they can read that message, and its history names who dated it. The day is read in your account's time zone, so a message sent at 8pm in Los Angeles is still that day there. A week or a weekend — "this week", "next week", "by the end of the week", "by the weekend" — is due by the Sunday that ends it, on your calendar (weeks run Monday to Sunday; said on a Sunday, "this week" is the week ahead): the item shows the words until that Sunday is over and is late only after it, whether the words were in a message, typed by you or set by an AI assistant. A day said with them wins: "by Friday next week" is that Friday. When the model is certain and you said it — your own message, your own words in a meeting — your row closes with the quote as its receipt.

Your plate updates without you opening it. When a meeting's transcript or an email that holds an ask, a promise or an agreement arrives and names you (by an address your account is known by) — or one already there is read again and found to hold a new one — your plate is checked again in the background, with that source, so the day's reminders and digest include it. If your plate was checked in the last ten minutes, it is checked once more when those ten minutes are up, so a burst of mail checks it once and the last arrival is never left waiting for you to look. Chat messages still reach your plate the next time it is checked.

Everything new you can read is checked against everything still open on your plate. Once your plate has been checked, it keeps catching up in the background: if you haven't opened Louvain for a day it is checked anyway, within your workspace's daily allowance for this, so a close doesn't wait for your next visit. Each time your plate refreshes, Louvain reads your open items beside every message, meeting and newly shared file you can open that arrived since it last looked (a file again when it is renamed, or edited on a later day — not for every save) — in a DM, a channel, any conversation — and asks which of them each one completes: "application for incorporation submitted" is offered on your ask that Tod get the company incorporated, a file Tod shares or updates that is plainly the shortlist he promised — by its name or its opening pages — is offered on that promise (a file is only ever offered, never closes a row on its own), and "Company name Harbourline Labs Ltd" decides the open question of what to call it (the row reads Decided in a later source). Only what you can read is ever checked, and the close it suggests is shown with that message's own words. Someone who isn't part of an item never closes your row: what they say is only offered, without their name — a display name is not who someone is — and a bare "Done!" from them counts for nothing. A reply is read with the message that started its thread, so "Done!" answers what was asked there.

The one who does it says it's done, so it's done. When the person you're waiting on says it's done — Tod replies "Done!" to your ask, or writes elsewhere that he sent it — his row and yours both close, with his words as the receipt. When the other person says something you owe is done — you promised Tod the deck and he says he has it — yours stays open with "Tod says this is done — is that right?", his words quoted, and two buttons: Confirm closes it, Not yet keeps it open (and that message isn't offered again). The same goes when Tod marks his own row of the same request Done in Louvain: yours shows "Tod marked this done — is that right?", and if he undoes it, the question goes away. If you both said it's done, both rows close. This holds whenever your row appears: if Louvain reads the conversation again later and your row comes back, it is closed by what Tod already said, or shows his Done, just as it would have at the time — as long as you can read the message that settled it. One exception: if the person who asked takes it back ("never mind, I sorted it") or says they have it, the row they asked of you closes — it was theirs to ask — and someone backing out of a promise to you closes it with their words. A reply in the thread of a promise works the same way — "Tod replied — is this done?" — because answering isn't keeping it; a reply from the person you asked still answers your ask. When the model is only likely — or when the sender is someone the conversation does not include (in Slack, an account other than the one that asked or the ones it @-mentioned), or an automated mail — the row stays open with the quote shown (its first two lines; tap it for the rest) and two buttons: Yes, close it or Keep open (Close or Keep on a phone). A stranger replying to the thread can suggest a row is done; only you can agree. Either way you can reopen. If the other person's own copy of the same item was closed by a message you can read too, yours asks "Tod's copy of this was closed — is it done?".

A suggestion you leave alone closes itself after two days. If a row has shown you that it looks settled (or that someone says it's done) for 48 hours, counted from when Louvain first showed it to you — not from when the message was sent — and you haven't tapped Keep open or Not yet, Louvain closes it for you, with that message as the receipt. It shows under Done this week as Closed for you: looked settled, with Undo: Undo reopens it and that suggestion is never offered or closed again. A row you snoozed waits until 48 hours after it wakes. A row you handed to someone or that someone handed to you, or that someone only replied to (a reply isn't proof it was done), is never closed this way, nor one whose message you can no longer open. When the other person's copy of the same item was closed this way for them, your row is not closed by it and doesn't name them: it shows at most the plain Looks settled.

Chat that isn't threaded counts too. In a DM or channel, a message posted on its own rather than as a thread reply ("it's in", "submitted") is read with the earlier message in the same chat that it continues, so it can settle the ask it answers the same way a thread reply does. Which message it continues is decided by your workspace's decision model (Workspace settings → Fast judgement → Unthreaded chat: what a message continues, on unless switched off; on a workspace that answers its decisions on this deployment's own model rather than the hosted one, it records what it would link and links nothing yet); the link only gives the message its context. It never closes anything by itself, and it only ever links messages in the same chat, so nobody sees anything they couldn't already read.

A later meeting in the same place, meaning the same notetaker's recordings that the same people can read, can close an earlier meeting's ask the same way. It looks back at up to five meetings from the last 30 days: "I sent the parts list yesterday" closes the ask to send it.

The order things arrive in doesn't change this. An ask and its reply that arrive together close on the same refresh, for both of you: the person who asked and the person asked. A reply only closes something asked before it was written, so it never closes a newer ask in the same thread.

An item that reaches your plate late — a meeting your notetaker delivered hours after it was held — is checked once against what arrived in between, so something already done before it appeared still closes it (or is offered to you as above).

And each closes when you say so — Done, or Not mine — from the row's menu. Snooze hides one for a day, until next week, or until a day you pick.

Said in more than one place

The same ask is often said more than once: in a mail, then again in Slack, then in a meeting. Your plate shows it as one row, and the row says how many times it was said ("Said 3 times"). Opening the row lists every other place it was said that you can open; a place you can't open isn't counted.

Two rows are only treated as one when they are about the same work and nothing says two different people are involved. Louvain tells who people are by their address or their linked account, never by a name: when it knows both sides and they differ, the rows stay apart however alike the words. Where a source doesn't say who someone is — a speaker the meeting didn't link to anyone, a list heading — the same work said alike is still checked, and Louvain decides whether it is the same thing. It is shown both rows with their people named as your plate names them, so a meeting's "Speaker 2" linked to someone reads as that person. A reminder is only matched with the ask it repeats when they were said within a month of each other. An agreement is matched with the ask or the promise of the same work, and so is a question left open with the ask or promise that settles it ("whether to loosen the export limits or drop the feature", then "fix the limit check or pull it"). The row then shows the ask or promise, so it stays on your to-do list. The same words asked of two different people stay two rows. So does the same weekly report asked for in two different weeks. When Louvain isn't sure two asks are the same, or can't get an answer, it keeps them apart. When a later message words the work differently, Louvain checks the two again rather than relying on its earlier answer. Marking the row Done or Not mine, snoozing it, or reopening it applies to every place it was said. A reply to any of them closes it, as long as the reply came after the ask it answers.

Said again, it moves on. When a later meeting or message picks up something already on your plate — the same work by the same person, its next step, or an update to it ("research the trade fairs", then "review the Milan and Cologne fairs") — Louvain reads the two together and makes them one row. The row reads as it was said last, with "first said" and the day it was first said beside it, and takes the latest due date anyone gave it. Something related but separate, or the same words about someone else's work, stays its own row. A date given later stays, even when the first place it was said is read again.

A new ask in the same mail or chat conversation is a new row of its own, never folded in behind the earlier one. A reminder sent after the first ask was already done is a new open ask, too.

A row said in several places lists the others under Evidence, as Also said in. If Louvain put two different things together, open the row and choose Not the same next to the place that doesn't belong. It goes back to being its own row. Louvain never puts those two together for you again, and doesn't ask about them again.

A mail and a Slack message can only be matched when the person behind them is a member of your workspace, with their Slack account linked.

What a row says

The line you read is one imperative sentence the judge writes beside the ask — "Send Jasmine a copy of your current passport". On a row you are waiting on it is the bare action ("Look into the trade account"), because the list it sits in already says you are waiting. It is checked the way every answer is: every name, figure and date in it must be in the source or on its envelope, and a line that carries one the source does not is dropped, in which case the row reads the ask verbatim. A figure has to be one the source states as a number: the "5" in a meeting timestamp such as "[0:05]" does not support "5 investors". What counts as a name is read from the sentence's grammar — never from capital letters — and a line whose names could not be read is not shown either: the row reads the ask verbatim. The verbatim ask stays under the line in quotes.

Click the line to open the item (on Today's strip, it opens the source): the source itself in the same document viewer the rest of the product uses, on the version the row cites. There is no excerpt on the row: two lines of a mail are almost always its greeting, and the point of a row is the ask, not the salutation.

An ask bound to a moment — "running 10 minutes late, does that still work for you?" — shows Needed by the time it names instead of a due chip, when the words are in the source, the time is within two days of it, and the time was resolved with its time zone (a time with no zone would move with wherever the server runs, so it is not kept).

A link in a quote or an excerpt that carries a sign-in or access token — a sign-out link in a new-device notice, a password reset, a signed download — is shown as the site it points to with the token removed. The source itself is unchanged.

The page

The plate is a to-do list. One line at the top says how many things are to do and when the plate was last checked (Check now checks again). Under it are two lists:

• To do — everything you act on: what someone asked of you, what you promised, what you agreed with someone, and a decision left open. Rows you starred come first; then by due date, with undated rows after (newest first among rows without a date).
• Waiting on others — what someone promised you and what you asked of someone. Folded until you open it.

Five ways to read To do. A switch on the top line chooses how: List (the default), By when — the same rows under Late, Today, This week (the next six days), Later and No date, each heading shown only when it has rows — or Matrix, which sets when a row is due against whether you starred it: Do first (starred, late or due today), Do soon (due now, not starred), Schedule (starred, not due yet) and The rest. Every cell is drawn even when it is empty. "Today" is today in your own time zone. Who groups To do and Waiting on others by the person or organisation each row is with — one group per person Louvain has identified, however a message wrote their name, so two people who share a name stay apart. Rows naming someone Louvain hasn't identified are under Not identified, and rows with nobody on the other side under No one. By category groups both lists by your own categories, a row in two categories under both, and rows in none under Not sorted. In the other views Waiting on others stays one list, because when it happens is up to someone else. Your choice is remembered in this browser.

Star what matters. The ☆ at the end of a row (or Mark important in its ⋯) marks it as important; tap again to remove it. Louvain never decides on its own what is important — only a category rule you switch on stars for you; an AI assistant connected to your account can star or unstar a row when you ask it to, and the row then says so ("Starred by an AI agent (Claude) for you") until you change it yourself or hand it to someone. On your Plate page, Undo beside that line puts back what was there before an AI assistant changed the date, the wording or the star, while the item still shows the assistant's change. A starred row is yours: when Louvain reads its source again it keeps the row and its star, a row said twice and shown as one is starred if either was, and a star stays on a row you mark done and comes back if you reopen it. A star one of your category rules set is different: it keeps nothing, so a row Louvain would otherwise retire goes, star and all. A star never holds a row open: when a later source settles it, or the other side marks it done, it closes or is offered to you exactly as an unstarred row is.

Your own categories

Sort your plate your way: Legal, Fundraising, Hiring, Needs signature — whatever you work by. The tag button on the plate's top line opens Your categories. Write each one as you would explain it to an assistant: a name, one sentence saying what belongs in it ("Anything someone is waiting on me to sign"), and, if you like, up to five examples. Or start from the ready-made ones it offers (Legal, Fundraising, Hiring, Sales, Finance and Needs signature) and change their words. You can have up to twelve.

Louvain then sorts each item on your plate into the categories it fits, and shows them as tags beside who the item is with (on a line of their own when they do not fit), usually within a few minutes of a new item or a change to a category's words. It tags an item only when it is confident, so an item can be in none ("Not sorted"); a wrong tag is worse than a missing one. Tap a tag's ✕ to take the item out of that category, or ＋ to put it in one (on a computer, the ＋ shows when you point at or tab into the item). Your choice always wins: Louvain never puts back a tag you removed, nor takes off one you added, however often it sorts again, and your choice stays with the item when Louvain reads its source again.

Star what fits. Switch it on for a category and every item Louvain or you put in it is starred, once — so "anything in Needs signature is important" puts your signatures in Do first and Schedule in the Matrix. The ☆'s label says the rule starred it. If you un-star an item, no rule stars it again, whichever rule and whenever you made it; any star or un-star of your own replaces the rule's. A rule's star doesn't keep an item on your plate: when Louvain reads its source again and finds nothing for you, it goes, as an unstarred item would.

Your categories are yours alone: nobody else sees their names or which items are in them, a teammate you hand an item to sees their own categories on it, not yours, and a category never changes who can read anything. Deleting a category takes its tags off your plate. An AI assistant connected to your account can list your categories, filter your plate by one, put an item in or out of one, and add a new category for you when you ask; each is labelled as the assistant's, and it never changes a tag you set or removed yourself (Connecting an AI assistant). Which categories an item fits is decided by your workspace's decision model (Workspace settings → Fast judgement → Sort my plate into my own categories, on unless switched off; Advisory sorts nothing for this one), which is sent the item's line, its source's title and your category words. Where nothing can answer it — the workspace has no hosted decision key and the deployment runs no decision model whose answers act — or an administrator switched it off, Your categories and the By category view say so, and only the tags you add yourself are on the plate.

The list holds still while you read it. While the page is open, a check that finds something new puts the new row where it belongs, marked briefly, and a closed row drops out; rows already on the page keep their places. Check now or reloading the page puts everything back in order.

Something agreed is To do when you, or both of you, do the work, and Waiting on others when only the other side does — "Tod will get the LOI by the end of the month", agreed, is Tod's. An agreement read before 8 October 2026 is To do.

Which list a row is in is decided by Louvain from who owes what, never from the wording. Above the lists, one line says what they hold: how many are overdue, due this week, on you and closed today, beside a ring of today's share closed. Each row is the task, with the face of who it is with and what kind of item it is in a few words beside their name (Asked you · Tod Pretorius, You promised · Tod Pretorius, Promised you, You asked, Agreed, Open question); a row someone handed on to you, declined or sent on keeps its own words ("Sam Reid asked"). At the row's right are when it is due, as a pill, and the mark of the app it came from, which opens the source. A date the source gave shows as the date, with the source's own words ("by the end of the month") when you point at it; words without a date ("eventually") show as those words; a passed date reads in red, in place ("2 days late"). Pointing at a row, or tabbing into it, also shows Set due date and Snooze until tomorrow beside the pill. In Matrix, where each cell is half the page, a row keeps its title's whole width: the due pill and the app's mark sit under the title, and the date and snooze are in the row's ⋯. The List view has no Overdue or This week sections; By when has them: Late, Today, This week, Later and No date, each with its count and its rows together on one card, a late or today heading in red or amber. A sender that is an app or an organisation is named, never said to have asked or promised anything. A person is named the same way on every row — a member by their name in your workspace, someone Slack knows by the name Slack gives them — whether a message called them "Tod" or "Tod Pretorius"; where a message doesn't say who someone is, the row keeps the name it used.

Tick the box to mark it done (on a Waiting row, that they did it, which the row then says when it closes); a note at the bottom of the page says so, with Undo, which reopens it. The row's ⋯ holds Snooze until tomorrow, Snooze until next week, Set due date… (a day picker, with Clear once it has one), Edit wording… (the item opens with its line ready to type over), Snooze until… (a day you pick, from tomorrow to a year out), Not mine, Ask someone to take this and Discuss; on a phone it opens as a sheet from the bottom. Every one of them works from the keyboard. Discuss opens the item's source with its discussion beside it and the item's words marked, so you can ask a teammate about it: @-mention them, and only people who can read that source can be mentioned or see the thread (see Discussing a connected document). A snoozed row leaves the list and is counted in one quiet line under To do. When Louvain thinks a later source settles a row, a line under it says so and quotes it, with Confirm and Not yet.

Done this week is a third folded list, with the receipt on every row: who or what closed it, when, and the sentence that did.

The strip on Today and the plate post in Slack (/louvain plate) show the same To do list: the first three things to do, starred ones first, and a count of what you are waiting on.

By email, if you ask. Turn on Weekly digest and due reminders (Account settings → Account) and your plate comes to you on Mondays, with a reminder on the morning an item is due, in your own time zone — read as you when it is sent, so only what you can read then (see Emails).

Click a row to open it. The item says who it is with and which way it runs (you are waiting on them, you promised them, they asked you), when it was said, any date it carries and its status. Evidence shows the words it was opened on with what was said around them: in a recorded meeting, that turn and the turns before and after it; in a message, the paragraph. Open in source takes you to exactly that place in the document, highlighted. Since then lists what has come in since that may relate, from the sources you can read, labelled Possibly related. Only the source that closed the item, or one Louvain suggests may settle it, is ever called that. The item's actions are below it (Done, Nudge or Reply, Snooze, Ask about this, Not a commitment), and History shows when it was said, later activity in its thread, a suggestion, a snooze and how it closed. Nudges are not recorded, so they are not listed there.

The item is yours to edit. While it is open:

• Its wording. The pencil beside the line edits it in place (Enter saves, Escape puts it back). Your words become the row's line everywhere you read your plate — the list, Today, the digest, Ask and your AI assistants — and the source's own words stay under Evidence. Use Louvain's wording puts its line back.
• Its due date. The date beside Due sets the day it is due, or clears it for no due date, whatever the source said. It is a day where you are, late only once that day has ended, and History says Dated by you.
• Snooze hides it for 24 hours, until next week, or until a day you pick (within a year); it comes back as that day starts where you are.

Louvain never overwrites what you set: reading the source again, a later message that rewords or redates the item, and the same thing said somewhere else and put together with it all keep your wording and your date. Your date and wording are yours alone. A meeting's follow-up to everyone there, and an answer Louvain gives in a shared discussion, use the source's words and the dates the sources state. AI assistants connected to your account can mark an item done or snoozed, snooze it to a day, set or clear its date and reword it when you ask them to; the row says what an assistant changed ("Dated and reworded by an AI agent (Claude) for you"), the assistant can undo its own latest change, and anything you change yourself clears the label.

Add an item

The line at the top of To do takes anything you need to remember: type it and press Enter. A date in what you type becomes its due date — "send the lease renewal on Friday" is due Friday, read in your own time zone — and a reminder email goes on its day like any other item's, if you have reminders on. A past or vague date ("in a few weeks") gives no due date; set or change one from the item's view, where you can also reword it. The words stay exactly as you typed them, and a name in them is never read as a person: to say who it is for or with, pick them from For or with beside the line (your workspace's members). The item then shows them like any other, on your plate only — they are not told and nothing appears on theirs. While the plate is focused on a client, the line says for Acme beside it, and the item you add is for that client: it shows under the focus and under their name in Who. Remove that with its ✕ to add the item for no one, or to pick a colleague instead. When nothing open is with the client yet, the empty plate offers Add a to-do for Acme too.

An item you add is yours alone and closes only when you close it, with Done or Not mine (a snooze hides it for a while). No source settles it, because none opened it — so no later message closes or redates it, and it never shows another person anything. Its view says Added by you where a source's words would be. It is on your list, your digest and reminders, Today's strip and Slack's /louvain plate, and your own Ask lists it ("you added this"); an answer Louvain gives in a shared discussion leaves it out. AI assistants connected to your account read it (with no source), and can add one for you when you ask: it says Added by an AI agent (Claude) for you, the assistant can take it off again while nobody has changed, handed on, done or snoozed it, and it can mark done the items it added — never one you added yourself or another assistant added, which only you close.

Ask someone to take an item

Ask someone to take this in a row's ⋯ hands the item to a teammate. Pick who, check the line they will see (the row's own, which you can reword) and add a note if you like. They see that line and your note, not the source: no quote, no title, no sender, no thread. If they can already open the source themselves, they see it as evidence too, and only for as long as they can.

On their plate it is theirs from the moment it arrives, in To do, saying who asked. They can snooze it, reword it, star it and set its date like any item. On yours it moves to Waiting on others and says who is doing it.

• Their Done closes yours, and your row says who did it. Reopening yours ends the handover and reopens nothing of theirs.
• They can decline, with a reason or none: it comes back to your To do saying so.
• Each of you is told: they get an email when it is handed to them (or a Slack message, if they chose Slack for it), and you get one when they decline it or do it. See Emails.
• You can take it back from your row's ⋯ while they still have it.
• If the source settles it first (a reply in the thread you can read, or your own Done), theirs closes too. They are told it was settled, not the source that settled it. An AI assistant marking your row done does not close theirs: they are still doing it.
• If they leave the workspace, what they had not done comes back to you. If you leave, what you handed on stays with them.

One person holds an item at a time, and they cannot hand it on again. An item you added yourself can be handed on the same way. AI assistants connected to your account cannot hand an item on. One can suggest it: your row then says who it suggests asking, Review opens Ask someone to take this with its suggested teammate, line and note for you to change and send, and Discard throws it away. Nothing is handed on, and nobody is told, until you send it. Your teammate's AI assistants can mark theirs done or snoozed, and that alone does not close yours. An administrator can turn this off for the workspace (Settings → Workspace → Asking someone to take an item); items already handed on stay where they are.

Focus on a client

If your admin has turned on the client picker (Settings → Workspace; it's off until they do), its icon — a compass, first on the right of the top bar — narrows the plate, Today, Search and Briefs together to one organisation or person Louvain already shows you. It appears only when there is someone you can choose. Open it for a short list (organisations first) or type to find one; the plate then shows only the rows that are with them, and the icon is lit with their initials until you clear it with the ✕ in its list. Each page it narrows says so under its heading ("Showing what's with Acme"), with a ✕ beside the name that shows everything again. When nothing open is with them, the plate says so once, in the list's place, with Show everything. The client is part of the page's address and stays with you from page to page — this browser remembers it until you clear it — so Back undoes it and a link you send carries it: whoever opens it sees only their own plate, narrowed the same way, if their workspace shows the picker. A page that doesn't narrow by client yet says so in the list. With the picker off, a link's client is ignored and nothing is narrowed. On Briefs, and in Today's and the plate's coming meetings, you see the meetings with them: a meeting is with Acme when someone invited to it is, by the same rule as a row below, never because its title names them. A link to another meeting on that page still opens Briefs, says so, and offers Show all briefs.

A row is with Acme when the person on the other side of it is, by identifier: an address at a domain that is Acme's (your workspace's verified domain, a domain a structured source such as your CRM states for the account, or the domain of an organisation your workspace corresponds with, in at least two messages or meetings — never a free-mail domain like gmail.com, which is everybody's, and never a domain your colleagues' own addresses are on — the ones they sign in with, or that a source links to them; an address a colleague's mailbox sends as, or that their own Google Drive or Calendar is signed in to, is never a correspondent itself, but its domain stays a candidate unless it is your workspace's own verified domain), a colleague whose own verified address is, or someone a source you can read says works for Acme or is a contact there — so a client who writes from Gmail is placed only when something you can read says where they work. Nothing is matched by a name, and a client never shows you anything you could not already see: a private record saying Bob works at Acme files Bob's rows under Acme only for people who can read that record.

An item you add while focused on Acme is with Acme because you picked it, never because of its words, and so is one an agent adds for Acme when you ask. It is with Acme for as long as Louvain shows you Acme, or names Acme's person through something you can read; if you can no longer see them, the item stays on your plate and names no one.

Which organisations you can pick. Louvain identifies an organisation by its email domain once someone in your workspace has corresponded with it: written to an address there, been in a meeting with someone there, or written in the same mail thread as them. Mail that only arrives from a domain (a tool's notifications, a newsletter) never makes it one, nor does a free-mail domain or your own. It is named by its domain (kauri.example), never by a guessed name, and you see it only if you can read a message or meeting in which someone in your workspace corresponded with them (the mail they sent there, the meeting, or the reply in the same mailbox's thread): mail that only arrived from them shows you nothing. A domain is the one registered under the public list of web suffixes (kauri.example, and for an address at x.eu.com it is x.eu.com, never eu.com). A name that Louvain hasn't matched to anything it has identified isn't offered as a client.

Some rows name nobody by an identifier (a meeting where the speaker wasn't recognised, a source that names no one, a row with several people on the other side). A client can't place those, and one line under the header says how many there are; Show them lists them below under their own heading. Showing them doesn't assign them to the client.

A chat id of someone outside your workspace doesn't place a row yet; a colleague's chat id does, because it is theirs.

Recorded meetings

A meeting recorded by a notetaker (Fireflies, Gong, Zoom, Fathom, Granola, Otter, tl;dv) or a Teams transcript can open a row, when you took part in it. Two kinds only:

• Waiting on: someone else in the meeting committed to something for you or your side ("I'll send you the revised quote on Friday"). The row is with the person who said it.
• Needs you: someone else asked you, in the meeting, to do or send something.

A list of action items or next steps posted with each person's name over their lines — "Tod" then "• Order the new oak boards – Fri 9 Oct" — opens a row for every line still to do: Tod's own lines are his promises — on his plate, and on yours when the list names you as the other side — and the lines under your name are his asks of you. Whatever the list's wording, a line under someone else's name in a list you posted is what you asked of them — never a promise made to you. However long the list, each person's lines are counted separately, so one person's long section never crowds out another's. A line saying something is already done, an agenda of topics, a recap and a line under no one's name open nothing.

A chat message that carries on from an earlier one in the same channel without being a reply to it — "Read the recommendations and shortlist" posted a minute after "Here's the doc on our ICP" — is read with the message it continues, so its row says what it is about ("Read Tod's supplier survey and shortlist") rather than leaving you to work it out. Louvain links the two only within one channel, so nothing from a conversation you can't see is ever used. Open the row and its Evidence shows those earlier messages under Earlier messages Louvain linked it to (the last three, as Louvain read them), above the words the row was opened on. When Louvain links a message only after it has already been read for your plate, that message is read again with what it continues, once, for messages from the last two days (a burst of links in one channel, as when old chat is first connected, re-reads nothing).

A meeting can open several rows — one for each distinct thing someone committed to or asked of you, however many there are; the same thing said again later in the call is the same row, and the same work said again in other words ("start outreach to companies", then "start outreach after validating the list") is put together as one, as it is across conversations. So is the same work at a later stage (looking into the options, then sending what was found); doing something with what came of it (signing, submitting) is a row of its own. Talk is read strictly: a row needs something you could check was done, named in that turn or in the turn it answers ("I'll look into the export licence application"). A turn that only points back ("give me that at some point", "I'll try to get that done soon"), a hedge ("should be fine once you try it"), or someone describing how they'll go about their own work opens nothing; a separate thing they'll hand you does, even when it comes out of that work. A message or a turn that asks you for one thing and promises you another opens both. When Louvain reads a meeting again (after it changes how it reads, or when the notetaker sends the transcript again), the same thing quoted in other words — or with more or less of the sentence around it — stays the same row, keeping what you did with it. Reading the same message or transcript again never rewords an open row: it keeps its line, whether it is yours or theirs, and its due date, and only adds a date the first reading missed and the source states. New words come only from a later message or a new transcript — or when a new reading finds that one row held two separate asks: the row keeps the first ask's words, and the other ask gets its own row. A row the new reading does not find goes (after a new transcript, only when its words are gone from it), unless you changed it yourself (done, not yours, snoozed, reopened, a date or wording you set, or a suggestion you confirmed or kept open) or an AI assistant marked it for you. And something done during the call ("I've just shared the deck with you") closes the row it settles, as does a target revised later in the same call ("let's make that five by December") replace the one it revises. On the page, a meeting's rows sit together under one line — From the 1 Oct call with Tod · 6 items — the first three shown and the rest one tap away, so one long call does not push everything else down.

What has to hold, and it is checked by Louvain, not taken from the model:

• You were there. One of your own addresses (your sign-in address, one your mailbox sends as, or one you added under "Your other addresses") is on the recording's organiser, participants or invited list, as the notetaker reported it. A recording the notetaker lists nobody for (a file uploaded to Fireflies) counts the people Fireflies shares it with for good, and its owner. The list must also give a name beside that address; without one Louvain cannot tell your words from anyone else's, and the meeting opens nothing.
• The words are one person's. The row's title is copied word for word from one turn of the transcript, and the person the row is with is the speaker the notetaker put on that turn. A turn labelled "Speaker 2" or "Unknown speaker" opens nothing, unless Louvain knows who that speaker is (below).
• It was not you. A turn whose speaker could be you (your name, or part of it, such as "Alex" for Alex Morgan) opens nothing as somebody else's, unless Louvain knows who that speaker is (below).
• It was said, not summarised. The notetaker's own summary and action items open nothing: nobody said them.

When Louvain knows who spoke. A speaker is bound to a person when the notetaker says who they are (Gong, Fathom, and Granola's own-microphone channel do), when the speaker's name is exactly the name of one person the meeting itself lists or one word of exactly one of their names ("Scoby" for Scoby Wadsley, never when two of them are called Scoby), or when they are the only speaker and person left and the speaker's label is made of that person's name (or nothing names them). A name two people in the meeting share binds nobody. A "Speaker 2" — what a notetaker writes when one microphone recorded everyone — binds by what the notetaker states, by someone's decision, or from what was said: Louvain's decision model reads the start of the meeting and binds a numbered speaker to one of the meeting's people only when the transcript shows it — someone addresses them by that name, or they say who they are — and no other speaker is matched to them. Speaking first or running the meeting is not enough, and anything less certain binds nobody. Naming the speakers in your notetaker afterwards works too (Fireflies, for meetings from the last three days). A turn bound to you is yours: an ask you made opens You asked, and a promise you made opens You promised. A turn bound to someone else is theirs, whatever their name resembles. On the plate a bound speaker is shown by the name of the workspace member they are, with one colour, so "Tod" in one call and "Tod Pretorius" in the next read as the same person. A speaker nothing identifies is treated as before, and a fact a meeting states about a numbered speaker ("Speaker 2 joined Acme in 2019") is kept only once that speaker is bound to someone; from then on it is that person's. What a meeting says about anyone's earlier jobs or where they were born is not kept as a fact at all; who works where now is. What a bound speaker said in that meeting is that person's in Louvain's knowledge too, when Louvain already knows the person's address (from your directory, their account or mail you connected) — in that meeting only, so another "Scoby" elsewhere is not taken for them — and the person is listed once, not once more under the speaker's label. Louvain decides this when it reads the meeting, so Who spoke shows it as soon as the meeting arrives; changing Match meeting speakers, or anyone's decision, is applied to recent meetings in the background. When who spoke in a meeting changes after your plate read it, your plate reads it again once, so a promise you made — or one made to you — that it could not place before can reach your plate. If you were in the meeting, that happens straight away, without you opening Today: the plates of the people the meeting lists are checked again in the background (up to eight). Anyone else who can read the meeting gets it the next time they open Today. A plate line written before the meeting's speakers were known names them as the notetaker did ("Send Speaker 2 the agreement"); once a numbered speaker like that is bound, the line shows that person's name instead, for anyone who can read the meeting — on the page, in the due reminder and digest emails, in notifications, in the Slack plate post, in the follow-up after the meeting and to an AI assistant reading your plate — and goes back to the label if the binding is undone. The source's own words, quoted, keep the label they said. Something a numbered speaker asked of you, or agreed with you, waits until Louvain knows who that speaker is: a plate never names "Speaker 2" as the person you owe.

Open a recorded meeting to see Who spoke: one line saying who spoke and how many speakers are not matched; open it for each speaker, who Louvain thinks they are and why (from your notetaker, matched by name, by elimination, from what was said, or set by you or by the member who chose — by hand, once that person has left the workspace). For a speaker nobody has identified, which shows as Not matched with nobody chosen, use Choose… to pick This is me or a member of your workspace. It holds for that meeting; your own This is me on a named speaker, or an administrator's choice, holds for that notetaker from then on. A speaker the notetaker identified can't be changed this way. Someone else saying a speaker is you doesn't on its own put that speaker's asks and promises on your plate; your own This is me or an administrator's choice does. Not me or Undo undoes a decision or a match, and it is not made again: anyone who can read the meeting can undo one for that meeting. A decision that covers every meeting of the connection ("this speaker is me") is withdrawn everywhere only by the person it names or an administrator; anyone else's undo sets it aside for that one meeting. What the notetaker said only the person it names or an administrator can undo. An administrator chooses under Workspace settings → Workspace → Knowledge quality → Match meeting speakers whether name matches bind on their own (the default), are only suggested, or are not made; what the notetaker states always counts. Deciding who spoke never changes who can read the meeting.

Are you in it? When a meeting you can read has a speaker who may be you and nobody has said (your first name, a word of your name, or your initials), or Louvain matched you and you haven't confirmed it, the bell asks once: Are you 'Scoby' in it? or We matched you as 'Scoby'. Right? It shows who else spoke, when, and how many things that speaker said were noted. Yes, that's me is your own decision, so later meetings from that notetaker know you without asking. No is remembered: you're not asked about that speaker again, and Louvain stops matching that speaker to you by name on that notetaker. If someone else has decided who the speaker is by the time you answer, your answer isn't applied and the question is set aside. Questions about your own personal notetaker's meetings come to you too. Someone else opens the meeting's Who spoke. A meeting is asked about once, 15 minutes after it ends at the earliest; questions arrive at most once an hour, several together as one. Nothing is asked where the notetaker said who spoke or someone already decided, and a "Speaker 2" is asked about only when Louvain matched it to you from what was said ("We matched you as 'Speaker 2'. Right?"); a "no" there holds for that meeting only. A meeting you can no longer open is no longer shown.

Which speaker were you? A meeting that names its speakers only by number ("Speaker 1", "Speaker 2", a recording made on one microphone) asks each person it lists who can open it, once, with each numbered speaker and the first thing they said. Pick yours: it is your own This is me for that meeting, and when the meeting lists exactly as many people as it has speakers, the other speaker is then the other person, so one pick settles a two-person meeting and what was asked and promised reaches both plates. None of these sets the question aside without ruling anyone out. Meetings from the last 14 days that Louvain read before this question existed are asked about too. A meeting nobody could open yet when Louvain read it (a new notetaker's first meetings, while who can read them is still being set up) is asked about once someone can open it, within those 14 days.

A calendar event is never read this way. It is also meeting-shaped, and the plate still refuses every calendar event (a birthday, an invitation, a series' repeated description). A recording is told apart by what the notetaker's connection wrote, never by what was said.

The judge reads the whole transcript, once — up to its first 400,000 characters, about fifteen hours of conversation, a bound against a runaway record: a long one is read in overlapping parts of about twenty minutes each, at the same time, and then once more with what the earlier parts found, so something said at the end can close or change something said at the start. A promise made two hours in is read in its own context like one made in the first minute. A meeting is dated by when it was held, as the notetaker reports it, not by when the notetaker last changed its record.

A promise someone made TO YOU that Louvain recorded from a meeting is offered to the judge too, under the same rules: its words must sit inside one turn, the row is with that turn's speaker, and a turn that could be yours opens nothing. A commitment or deadline recorded from a meeting that is not a promise to you is not offered.

What you agreed

Some conversations don't ask anyone for anything: they set something. "How about a design partner signed by the end of the month, and the seed round closed by the end of the year? Let's go with that." That is a target two people agreed together, and it opens under Agreed on each of their plates, each "with" the other, whoever's plate it is. A decision the conversation explicitly leaves open ("per seat or a flat fee — I'm not set on either, let's decide after the pilot") opens there too, as an Open question. Arranging the next time you'll meet or talk ("we'll chat again on Monday") is not something either of you delivers or decides, so it never opens as an item. Something the other person agreed to do for you reads as what was agreed ("Tod Pretorius gives you 500 shares immediately"); where Louvain can't put it that way, it reads "Agreed with Tod Pretorius:" and the words agreed.

What it takes, checked by the server whatever the model says:

• Both sides. The target is said in so many words in one turn of a recorded meeting, and the other person's own words — more than an "okay", unless a one-word "Great." is the whole of what they said in that turn — are in the turn nearest it that holds them, within four turns: their "yes, let's do that" after it, or the push that the target answers before it. One of the two is you, and that means a speaker Louvain knows is you (see Who spoke), never one merely named like you; until it knows which speaker you are, nothing agreed in that meeting opens. An agreement two other people in your meeting made is theirs, not yours. A musing, a hope ("I'm hoping we get to twenty customers"), or an idea nobody takes up is not agreed.
• In mail or chat, it is the reply that agrees to the message it quotes — "Agreed, let's lock it in then" under the proposal it answers — so the target and the agreement are two people's words; or two messages of one conversation, one yours and one theirs, when both arrive together. One person setting a date in their own message, however firmly, is not an agreement.
• A stated time. An agreed target says when, in words that are in the conversation. The day is worked out from those words and from when they were said, on your calendar — "the end of the month" said on 1 October is 31 October, "the end of the year" is 31 December, "the end of Q1" said in November is next March. Words those rules don't turn into a day — "by the 20th", "in six months" — take the day the model read them as, kept only when it is a real day on or after when they were said; otherwise the words stay with no date. An open question needs no date.
• The latest version. When the same conversation sets a target and then revises it, only the revised one opens. When a later conversation revises it, the new target opens and the old row is closed by it (or suggested closed, where a recorded meeting can't say who settled it). Read again later, the same target is the same row, however many of its words the reading quotes.

Each row leads with a short headline of what was agreed or left open, such as "Three paying design partners by the end of November", with the words it was agreed in quoted beneath. A headline is kept only when every name and number in it is in the conversation; otherwise the row shows the quoted words alone, followed by the other side's words. When the same words of agreement were heard for more than one target in a conversation, each row shows its own target's words instead, so two agreements never share a quote. When the work is yours, or both of yours, the row reads as the next step instead — "Get the LOI signed by the end of the month" — held to the same rule: every name, number and date in it is in the conversation, or the headline shows. An agreement where nobody takes on a next step is not opened as agreed. Every row also shows who it is with and when it was agreed. It is not something you owe a reply to, so it never appears in the plate strip on Today or in the agenda Louvain answers from your plate, and its one action opens the conversation it came from.

What it refuses to do

• A recorded meeting opens only what someone else said to you, in a meeting you were in (see Recorded meetings). Your own promises, promises between other people in the room, and the notetaker's summary never do. Of what Louvain recorded from a meeting, only a promise made to you is offered, and only from one turn someone else spoke.
• A passed deadline is flagged, never closed. Nothing about the date says the work happened. A passed moment is different: an ask that only made sense before a time closes six hours after it as "The moment passed", with no receipt, because nothing is left to do.
• An ask whose answer arrives in the same check — a reply a minute later — is written already closed with that reply as its receipt, never left open past its own answer.
• A later message that does not settle a row — someone nudging a thread, your own follow-up on something you asked — is counted and shown as activity, not treated as an answer.
• A recorded fact's row is never closed by something you cannot read. What stands for you is worked out from your own sources, and what you can read decides that — not what exists.
• A new ask — or a promise — in a thread whose row already closed opens as its own row; the closed row keeps its receipt. A thread holds at most one open ask; what the conversation agreed sits beside it as rows of its own. "Not mine" is your word about the thread and stays.
• If the model could not be reached — or replied with something that is not a judgement at all — nothing opens and nothing closes, the same sources are read again on the next check, and the page says so: a quiet plate and a broken judge must not look the same. Thread replies still close rows without it. One malformed entry in an otherwise good reply costs only that entry.
• A check never moves past a message it did not finish with. What a check opened and closed and how far it read are saved together or not at all, and messages are read in the order they arrived, so a large import cannot slip a message past the check unread.
• Two tabs, or two servers, never check your plate at once: the second one shows the plate as the first leaves it. Done or Not mine tapped while a check is running is never undone by it.
• A paused process whose refresh lease expires cannot later write over its successor. Lease ownership, settlements, openings and the cursor are checked and committed together. The deterministic reply rule, like the model's settlement check, accepts only the viewer or an existing participant and rejects list-mail closure. Other evidence may suggest a settlement for you to confirm, but cannot silently close the row.
• Opening your plate never waits for the check. The page shows what is already on your plate at once, and says where the check stands in ONE place, beside Check now: "Checked 5 min ago" when none is running; "Checking…" from the moment you press Check now (the button keeps its name and waits, disabled, with a spinner in place of its icon) or while a check started elsewhere — another tab, another device, the scheduled one — is still running; "Reading 4 new sources…" while it reads sources you can see; and back to "Checked just now" when it finishes, with the new items in place. The page reads again on its own while a check runs — every few seconds for the first minute, so new items appear within seconds of the check finishing, then every 15 seconds — and if one runs for more than a few minutes (a long recorded meeting) it says "Still checking — this page updates when it's done" and keeps going. A check runs at most every two minutes unless you press Check now, or at once after you change your plate rules. If the check fails, the same place says "Couldn't check just now" with Try again; an empty plate under a running check reads "Nothing needs you yet.", never as a quiet one. If the service that checks names in Louvain's own wording is not answering (a self-hosted sidecar starting up, say), it says "Waiting for the reading service — showing what's already read": new sources wait up to ten minutes for it, then appear with their requests word for word, and gain their short lines once the service is back.
• When Louvain's plate check itself improves — a new model or new instructions — the last two days of messages are read again with it within minutes, whether or not anyone has opened their plate since, and your next check shows what it would now find, or close, without anyone running anything. Older conversations that still have an open item on your plate are read again too, however old, a few at each check. Whoever's check reads a conversation again, every plate holding items from it gets the new reading, not only theirs; and a conversation the model can't read right now never holds up the others. An item the new check no longer finds there leaves your plate, unless you've done something with it: marked it done, dismissed or snoozed it. A conversation you can no longer open is left as it stands. A check you're waiting on never stops to read an old conversation afresh: that happens in a check nobody is waiting for, and the next one you open shows the result.
• While an operator is rebuilding your plate — re-judging a month one day at a time, which takes minutes — the page says "Rebuilding your plate from your sources" instead of drawing an empty plate. The rows arrive as the walk works through the days, and a rebuild does not clear what is already there: it regenerates rows in place, so a row keeps its history and gains the newer words. A walk reads each day as that day, so something stated after the day being replayed simply has not been said yet on that day — the row is left alone, never closed. "Not yet stated" is not "no longer current". The same holds when nothing you can read states an obligation any more, because access changed or its source was withdrawn: the row is left alone and the page simply stops showing a row whose opening source you can no longer read. One case goes further: when an earlier message is withdrawn or held, a reply that was read with that message's ask — an agreement to it, say — is read again without it, and a row that reply opened which you have not touched is removed, so the withdrawn words never live on through the reply.
• If the model cannot read a source, the plate check tries it again a few minutes later, then after longer waits, and after five failures moves on as if the source asked nothing, so one unreadable message never holds up the rest of your plate. A moment when the model is busy or unavailable is not counted as a failure: the source simply waits a minute and is tried again. If the model keeps refusing a source for more than two hours, your plate stops waiting for it and reads everything after it; the source is still tried, and when it is read its items join your plate (if you can still read it). After a week of refusals it is treated as asking nothing. A workspace whose plan or monthly budget does not allow model work has no sources read until it does.
• A row remembers which path opened it — the judge, an obligation your sources record, or the plain rule that runs when no model is configured. An operator can rebuild a plate from a judged month, and the run shows its work day by day so a person can check that every row is one they would want. Removing the rows an older path opened is a separate, explicit step.

Resolution looks beyond Today's window: a request from last week is closed by a reply yesterday. Every row keeps its receipts — the message that opened it, the latest in its thread, and the one that closed it — and each is re-checked against your access every time the page loads. If you can no longer read the message that opened a row, the row is gone. If you can no longer read the message that closed it, nothing closed it as far as you are concerned: the row is open again, with none of that message's words and no count of replies that includes it. The same holds wherever your plate is read: the digest, your notifications and an AI assistant connected to Louvain.

Without an email address on your account nothing can be matched to you, so no ask opens or closes; recorded facts still show.