Context and definitions

Most of what makes an answer right isn't in the data. It's that net revenue means after refunds, that Tom does the reconciliation, that March was strange because of the move to new premises. Overscore keeps those as facts, each with a source, a date and a level of trust: heard, or confirmed by a named person.

The library

Context, Library has twelve folders, always, as one short list under three headings: the business (The company, People, Teams, Customers, Products, Systems, Other things), the numbers (Metrics, Rules) and what happened (Timeline, Notes on things, Checked answers), with a count beside each. A folder with nothing in it yet shows 0, and its own page says what would fill it: People, what the interview hears about who works here; Notes on things, what has been written on a table, a model or a dashboard. A folder opens as a table, and a row opens beside it. One line under the title says how many things wait on you in the Inbox, only while something does. On a phone the folders are a list to pick from, the empty ones marked none yet.

Metrics is a table with one line per metric: its name with its other names and what it means under it, the table it reads by its real name (a link to that table), who stands behind it, and how many pages ask it. The box above the table finds a metric by its name, its words or any of its other names, so typing GMV finds net_sales.

House rules

How the agents do things here, in four short texts under Settings, Company: how models are named and where they live (the dbt way: stg_, int_, fct_, dim_, every layer in the analytics dataset), how a dashboard is laid out, how an answer is given, how a source lands. On the Company tab, House rules opens their page: each rule's first lines and whose it is (ours, or yours with who changed it and when), and Edit, which opens the rule in full in a panel at the side; Save there. Leaving a changed rule asks before discarding it; a failed save keeps the draft. Every agent in the app follows the saved rule from its next turn; an agent on a laptop reads the rules ovs init wrote into its folder, so run ovs init there again after a change. Empty it or choose Back to ours to use our default again.

Who can confirm

Admins choose who can confirm a metric (Settings, Company). A teammate who has never signed in here or whose current access cannot be verified can be asked to check an answer, but can open or answer only with current access to every original table, without masking. Known insufficient or revoked access is refused before sending. A confirmer's "looks right" turns a metric from saved to confirmed, at the version the answer used. The metric's page shows who confirmed it and when. An admin or nominated confirmer can confirm by saving the definition while signed in, including with ovs login then ovs push, or by choosing Define on a dashboard number. Overscore records that person and the date itself. An agent key, even one allowed to write, saves a guess (heard); its tool answer says so and names the confirmation path. A name or date typed into a file, an API request or an agent's tool call cannot confer confirmation.

What a metric is

A metric is a named business number with a saved calculation: the name people use in meetings, the calculation (what's added up, on which table, with which filter and date), and its owner. Saving it and confirming its definition are separate: the trust label says whether a person has agreed to it. The SQL is generated from those parts every time it's asked, so it can't drift from the definition, and every change is a new version with the old ones kept. Whether the table it reads is raw or cleaned is a detail shown on the trace, not a mark against it.

A source's first sync proposes its obvious metrics as cards in your Inbox, each with its definition and last month's figure, and none is a metric until a person confirms it (see The inbox).

You don't define everything. A company has ten to thirty numbers people argue about; those become defined metrics. Every other number still shows its SQL, its table and who made it, and says honestly that it isn't a defined metric. When one of those becomes a number people fight over, the owner names it and it's a metric from then on.

A number is defined once. Saving a new metric that works out the very same number as one on file (the same thing added up, on the same table, with the same filter and the same date, or the same rate of the same two metrics) is refused, and the refusal names the one to use. If the new one really is a different number, say how in one sentence (in the app, the chat or the file's differs line) and it saves, with that sentence kept on it. A different filter is a different number (overdue invoices and all open invoices are two metrics), and so is a running total, a share of the total, or the same sum counted by another date.

A metric can carry the other names people use for it, GMV, sales, gross revenue. Search finds it by any of them, the assistant reads them first, and no two metrics may share one. Spaces, underscores and punctuation count as the same spelling, so top-line and top line cannot name different metrics. If older names conflict, ask which metric is meant before working out a figure.

The metric's own page

A metric opens beside whatever you were looking at, at once: click it in Context, Metrics, or its name where a dashboard, its Where the numbers come from tab or a chat names it. Its own page (the same, at its own address) is there for a link. In order: its name, the other names people say for it, and what it means in a sentence; one line with a dot that says who stands behind it (confirmed by whom and when, or defined by the AI in whose chat and not yet confirmed); the number for the latest month with its change on the month before, and a chart of the last twelve months, with a picker each for the bucket (month, week, quarter), the split (everything, or by store, by channel: the splits the metric declares) and the period; How it's worked out as a few short rows (what it adds up and on which table, a link to that table; which rows it keeps (Only · not cancelled); the day each row counts on; what it may be split by; how it is shown), and Show the SQL, which opens the statement it runs right there; its questions, one at a time under the title Before you confirm, with the answers as buttons and Ask about this, which opens a chat about that question; then Confirm, or Change it in a chat; then where it is used and each version. Nothing is folded.

Its other names sit under its name, "Also called GMV, sales, top line", with Change beside them, which opens one field for the words people say for the number, with commas between them. A word another metric already has is refused, and the refusal names that metric. Spaces, underscores and punctuation count as the same spelling, so top-line and top line cannot name different metrics. They are part of its definition. Saving them adds a line to its versions ("Version 1, added the other name GMV"), and its version number, which counts changes to the calculation, stays as it was.

The line with the dot says who stands behind it: "Confirmed by Sarah Bell on Oct 13", or for a metric the assistant saved in a chat, "Defined by the AI in Kim's chat on Oct 13", the chat a link; the accuracy tests that check it every night join that line. When the model a confirmed metric reads is saved with new SQL, the mark comes off: the line turns amber and says "Confirmed by Sarah Bell on Oct 13; the model it reads changed on Oct 20, not yet confirmed again", the Inbox asks whether it is still right, and a person confirming it again puts the mark back. The definition itself does not move, so there is no new version. Confirm is for an admin or a named confirmer while nobody has, with Change it in a chat beside it (once it is confirmed, Change it in a chat is the button). Confirm is there with questions still open.

Under them, its review questions (what its data shows that its definition does not settle) come one at a time, with the answers as buttons: "Do cancelled orders count? 41 of September's orders were." [They count] [They don't] [Something else], and "1 of 3" beside it. Something else opens a box for your own words. Your answer is kept as a dated fact, with the question as its evidence, and the next question appears. Under a rule, a small line each: what the answers settled ("Settled with Sarah Bell on Oct 13: cancelled orders don't count, refunds don't come off"), every card that shows it ("Used on the Average order value card on Orders overview") and each version, with what changed, who and when ("Version 1, defined by the AI in Kim's chat on Oct 13"). The page changes nothing else: Change it in a chat starts a chat about this metric, so a change goes through the same checks as any other.

A rule you say in words is kept as a rule in those words, a number in it and all: "an invoice with a balance of at most 3.5 percent of its total plus 31 cents is paid" is filed as that sentence. Only what you said is kept; a figure it worked out that you did not say is left out.

A rule you tell the chat about what a number counts, like "never count test orders in net sales", has to reach the number, not just its description. If the card it makes only changes the words, the card says so: the number would be worked out as before. Approving that on a metric someone confirmed doesn't change anything yet: it asks which rows are meant (the metric's questions about its data are asked on the metric's page) and waits for a filter, a test on the metric's own table that keeps only the rows to count. Give it on the card and press "Yes, with this filter": the metric gets a new version and the number follows its words. A filter the table can't run is refused.

Where facts come from

• The interview. A short conversation under Context, Talk it through, whenever you like. The tab begins with the business and keeps the dates you give, including whether something changed or was always so. The first time, it asks for your website and reads it (below); say there isn't one and it goes on. One question at a time; it never asks what it already knows. It keeps only what your answers say: a question you did not answer is never taken as your answer, and when you say something it heard is wrong, it drops it. A count such as "60 people" is kept as a fact about the company, never as a person. When you finish, every fact is checked first; if one can't be saved (it ended on a date but no start was said, say), nothing is saved yet, and that fact is shown with why so you can leave it out or answer with its dates. A new company's first screen, once it is built, offers it as one of four doors: talk about the business, connect a source, go to the app, or take the tour.
• Your corrections. Every "no, it means this" is kept.
• Your website, and the web about you. When you give the interview your website, it reads it in the background while the questions go on (what you sell, the systems it can see in the site's code), then looks your business up: your careers page, the hiring site your job postings are on (Greenhouse, Lever, Ashby, Workable, BambooHR, Workday and the like), and two web searches, your business name with your website and your business name with "jobs". Only your business name and website leave, to a search company. Software your postings mention (NetSuite, QuickBooks, Excel) is only a suggestion, since a posting that asks for a tool doesn't mean you use it: it is kept as a note under Context, with the sentence each came from, and is never counted as a system you use until you say so. LinkedIn and Indeed are never read; what a search shows of them is all it sees. Each page is read from the address that was checked, and no more than 2 MB of it. On Free nothing is read, not even the website. With the AI data policy set to inside only, or with web search off, only your website is read.
• Cards you approve. Things it heard in a chat, a note or from an agent, confirmed by you.
• Your documents. A deck, a contract or meeting notes kept in Warehouse, Files: when its text is read, up to five facts from it wait in the Inbox as one row named for the document, each with its page. Only people who can open the file see them until someone approves.
• What it notices. A number that moved, a finished month whose number changed since it was last measured (rows that landed late, a model that changed), a definition used two ways, a guess an answer has used, or one nothing has used in a year (Still right?).
• Your chats, re-read overnight. Each night it re-reads the chats it has not read yet, a few at a time, for what you said that should be remembered. A night that cannot finish keeps what it read and leaves the rest for the next night; nothing is filed twice, and nothing is filed from an answer that came back cut short.

The Context screen is a library

On a computer, the folders sit on the left with a count each, and the chosen folder is a table on the right. The company comes first, then its parts (people, customers, products, systems), then the other things facts are about (a store, a supplier, a campaign), then metrics, rules, the timeline, notes on things and checked answers. Each table has the columns that suit it: people show their title, team and who they report to; systems show what they hold and who owns them; metrics show their name and what they mean. The Status column says the top of what each has earned, in a word or two: Confirmed, Checked against the data, Used in 3 answers, Heard, or Disputed once someone said it is wrong. On a phone, one folder picker sits at the top and each row is the name and the line that says most about it.

Click a row and it opens in a panel on the right: where it came from, its fields as a label and a value, every other fact about it with who said it and when it was true, what is no longer true, the notes on it, the history of the row, and how often answers read it. If it is not confirmed yet, Yes, that's right confirms it with your name. Point at a field or a fact and a pencil shows at the end of its line (on a phone it is always there): it opens that line to edit, names, dates and yes/no values included; it asks whether the thing changed on a date or was always so, and Cancel leaves the saved value alone. "Tell it something about Bob" opens the box with his name already in it. A metric's row opens the metric's own page; a note's row takes you to the dashboard, model or metric it is on. A first name is tied to one recorded person when filed; if two people share it, Overscore asks which one. Later hires or departures never move that fact to someone else. Text rules appear on the company and under Rules as the same dated line, so a correction updates both. Older rule notes remain readable.

Notes on anything, and the table card

A note can be on a table, and since September 30 on anything else too: a dashboard, a model, a metric, a person, a source. Notes on models, saved queries and dashboards follow access to their data; notes on people, the company and other things follow company context. Search and old versions follow the same rule. That is where "why this dashboard leaves out wholesale" lives. The builder reads a dashboard's notes before it changes it and leaves one after it builds: which tables and metrics it read, what it left out and why, which facts it used. Every table also has a card: what one row is, what to use the table for (and not for), and what each column means. The AI fills it for the models it saves, marked not confirmed yet; it can fill an empty column meaning, but cannot replace or remove a written meaning. A confirmed note can only be changed by someone allowed to confirm the replacement; an agent proposes that change for approval. You can write or correct what it is, what one row is and what to use it for under About this table, on the table's page in Warehouse, as soon as the table exists; a column's meaning is written through an agent or the API. Write it opens the fields; Save confirms your words with your name, and Cancel keeps what was there. The guide the AI reads first shows the card before the computed description.

Checked answers

Pressing Right on an answer you can already see keeps what was recorded, even when an older answer did not keep every table its metric read. It sends the answer to nobody. Sending that answer for someone else to review still needs a complete record of the tables behind it.

Under every answer in a chat there are Right, Wrong and Ask someone to check, and the line under them says what the one under your pointer does. A mark you pressed by mistake comes off when you press it again, and a Right taken back lets go of the answer it kept. Press Right and, besides telling us, the answer is kept as a checked answer, which the line under it says ("Kept for next time"): the question in your words, how it was answered (the metric, the period, the split, the filter), the number, and your name and the date. A teammate's "right" on a check you sent them does the same, and so does an accuracy test that graded a question right. Only people who can read every source dataset with its real values can see a checked answer, including copies from Accuracy tests. The question, value and how it was answered stay private from other keys. A key an admin made, with access to all data without masking, can inspect the complete Context records through SQL; any other key uses the Context controls and search, which check each record. When an allowed reader asks something like it, the chat finds the checked answer, answers the same way, and says who checked it. Every night the saved metric requests are run again with the same dates, split, filter and original calculation time, with no AI involved. Each figure is compared with its own result, and one that moved becomes a card. When only some figures can be repeated, the page says partly checked and shows how many matched, moved or could not be checked. Older answers may need a new check because their original calculation was not saved completely. Ask the question again and press Right on the new answer. If a metric definition changed, ask again using the current definition. Accuracy tests are how you test the AI (some call them evals): questions whose answer you already know, typed in on purpose under Settings, Company, Accuracy, and asked cold every night; the page shows how many it got right. There can be at most 40, so every one is asked each time. Checked answers are the ones using it makes. An accuracy test grades one clear figure in the written answer; an ambiguous answer says why it could not be graded. Finding the expected figure elsewhere in the query rows cannot make a wrong answer right. The saved answer and the last check use the same number formatting; Answer details keeps exactly what you entered and what the check recorded.

That's wrong

Under every answer there is also Wrong. Pressing it sends nothing yet: a box asks whether it is wrong or right but needs a nuance, and what was off, listing the facts the answer used beside the usual reasons; Send sends it, and Not now leaves no mark. Pick the fact that is wrong and send. If it said something else before, it goes back to that: the answer that used it keeps its record, and the fact's history shows both, with who said it was wrong and when. If nothing came before it, it is marked disputed and left out of answers until someone fixes it; its page still shows it, labelled. On Never ask this happens straight away when you may undo it: an admin always, a member unless an admin said or confirmed it. Otherwise, and on Ask me about definitions, the admins get a card in the Inbox that says who said so and what it said before, with Put back and Leave it (when nothing came before it, Leave it out of answers and Leave it).

A figure worked out from your data and kept as a fact ("net sales were $243,025 in August", filed about the company or about the metric) follows the same rule: an agent's key sees it in search, in the library, on the metric's page and in the block every chat starts with only when it can read every dataset behind that number with its real values. A person who is not an admin sees it only when they may know it (Who knows this, below) and may see the data behind it.

Who knows this

Everything Overscore is told starts private. A fact, a note, a card in the Inbox or an item on the Later list is known to the person who said it and to the admins, however it got in: said in a chat, picked up from the day's chats overnight, typed into the library, or approved from a document. So what you tell it about one person's pay plan in your own chat is not something that person's chat can find.

An admin shares it, with one group or with everyone: open a fact in the library, or a card in the Inbox, and press Share beside "known to" (a note or a Later item is shared through the API). The person who said something can keep it narrower than it is, never wider. Say "keep this between us" as you tell it something and it is marked so: sharing everything at once leaves it out.

• Who it applies to. Every person who is not an admin, whatever group they are in. Their chat's search, the lines every chat starts with, the library, the Inbox and the Later list each show only what they may know.
• What was there before. Everything filed before 7 October 2026 is known to everyone, as it was. To share everything known so far in one step, ask us, or as an admin: POST /context/knows with level everyone, or a group.
• A figure from your data keeps its own rule as well: it also needs the reader to see that data, and the stricter of the two wins.
• SQL. What Overscore knows about the business is read through search and the library. Its own record of what it was told (context) is the admins' to read with SQL: for anyone else, and a key they made, it is closed wherever SQL runs, the query editor, an agent's SQL, a saved query, a model, an alert, a live link and a table's page under Data, as it is saved and again each time it runs. The facts themselves (company) are read with SQL someone types only by an admin.
• Agents' keys. A key knows what the person who made it knows. One an admin made on Settings, Agents reads it as an admin does, by the datasets it was given, and a key sealed off from it reads none, so give it only to someone who may know what the company has told Overscore. Only an admin makes keys there; a laptop logged in with ovs login knows what its person knows, in the group they are in at the time.

What every chat starts with

Before anything else, a chat, the dashboard builder or an agent on a laptop that may read company context is handed one short block: what the business is, in a paragraph, the defined metrics by name with one line each (and the other names people say for each), the dated facts about the company that are true today, and how much else is known. It is capped at 800 words so it stays in the back of the AI's mind rather than in front of it; everything else is found by search when a question touches it. Read it under Settings, Company; it is written from the Context screen. A key shut out of company context receives only general guidance, with no company words or counts. When you tell the chat something durable about the business, it keeps your words. One sentence can hold more than one fact: "Jamie started September 15 and gets 5% commission" files Jamie and, separately, Jamie's commission from that day, so "what is Jamie's commission?" finds the 5%. Nothing you did not say is filled in: no manager, team, date or number appears unless your words gave it. The chat ends with a line starting "Filed:" written from what was actually saved, and says what is waiting for a yes or which person you mean.

Then, on every message you send in a chat or the dashboard builder, Overscore looks up your words before the AI answers: the defined metrics whose names, other names or definitions match them (a page's own metrics, in its builder) and the facts and notes that match, five of each at most, each with its name and how far it is trusted. A word of yours that names exactly one metric is read as that metric, and the look says so first. Older conflicting names ask which metric is meant. The AI gets them beside your message, so "trial conversion rate" finds the metric of that name without the AI having to remember to search, and the answer can say which one it used. ovs push does the same for a model or a dashboard: it prints the metrics already defined on what it reads.

How much a person approves

Settings, Company has one setting for this company, How much it asks you, with two choices; another company’s setting never changes it. The Context screen has a short link to it under its heading. Ask me about definitions, the usual choice on every plan but Free: what the chat and the agents hear is filed at once, not confirmed yet, and you are asked only about a metric to confirm, something said that disagrees with what you confirmed, someone leaving the company, and something not confirmed after a completed answer has read that version, once per version. Only words that say a person left the company mark them as gone, and the card says so before you approve it; "Tom managed the Bend store until August 31" is kept as a dated fact about Tom, and Tom stays on file. Saying it is still right confirms the version shown. If the fact has changed meanwhile, the old question closes without confirming the newer fact. When something said disagrees with what is on file, a fixed order decides which stands, never which came first: what your data shows beats what anyone said, what an admin said or confirmed beats what a member said, and between two of the same kind the newer stands. Here the order is what the Inbox proposes: a guess that disagrees with what stands over it waits instead of filing itself, and its card shows what is on file and who said it, with the order's choice as the first answer, Keep what is on file or the change. You still say yes. Kept, the card's words go on that fact's record beside it. Never ask, the Free plan's setting unless you pick another: nothing waits for you. What it hears is filed at once, a new metric and someone leaving included, and the fixed order decides a disagreement itself. The other one is never thrown away: it stays on that thing's record with who said it and when, and the Inbox's Handled for you lists it. New words for a metric that would not change how its number is worked out are kept as a rule about the metric, so the number never says something it does not do. Until October 14 there were four choices: a company on Ask me about the important things or A person approves everything is on Ask me about definitions now. Website research also waits before changing anything you confirmed, including what you said in the sign-up interview. Unsorted words found in a reread wait in Inbox; empty or repeated suggestions are skipped. The label beside a fact is its history in plain words, never a warning. A person's yes is the top mark: confirmed by Nina, Sep 29. Below it, what the fact has earned since: checked against the data 12 times (an answer said the fact's figure and the data that answer read held it) and used in 30 answers, nobody objected (finished answers that were given it). A fact that has earned none of these says where it was heard: heard from Sam's chat, Oct 2. The counts belong to the version they were earned by, so a changed fact starts again. In the library's tables the Status column says the top of it in a word or two: Confirmed, Checked against the data, Used in 3 answers, or Heard. A guess read in three completed answers is also marked checked; a wrong or nuanced verdict prevents that, and later negative feedback withdraws it. Each morning there is one line for it: learned 4 things yesterday, 2 held up in use, 1 needs you. An admin can have that line emailed each morning to the addresses they list in the same place; it is sent only on days with something to say, and names what needs a yes. A failed email can be retried without sending it again that day to people who already received it.

Things that happened, and loose statements

Something that happened on a date ("we moved the books to NetSuite in March", your answer to "why did reorders drop last week") is kept as an event about the company and shows on the timeline; it never expires and is never asked about again. A statement with no obvious home ("everything is drop-shipped") is a dated fact about the company. What the website and the web said beyond the basics is a note on the company, marked not confirmed yet. Nothing is filed under "decisions" any more.

When it was true

Every fact carries two dates beside who said it: when it became true and, once it stops, when. "Bob's commission is 7% from January 1" closes the 5% that came before it on that day. Ask what it was in November and you get the 5%; a report sent in December says which it used. When you change a fact, the inbox asks the one thing it can't know: did it change on that date, or was it always so and we had it wrong? One tap either way. If the raise really began February 1, correcting its start restores 5% for January; an old receipt still shows the belief it used. Correcting today leaves scheduled future changes in place. Older facts without a start date keep their past. To move the only scheduled raise, say when it really starts and choose correction; if several changes are scheduled, choose which one you mean. A fact that is no longer true stays in the history, marked, and is never used to explain a later month. A fact can be about anything: a person, a store, a supplier, a campaign, whether or not it has a page of its own. If a name is unknown or shared, the original words and dates wait in Inbox: choose the person or give their full name. Once filed, the fact stays with that person through hires and departures.

Where did this come from?

A thing in the library has the question on its top line, the same as a number on a dashboard. Press it and a card opens: one line saying where it came from ("approved by Ruth from q3-board-deck.pdf, page 2", "from Sarah's chat, Sep 26", "said in the interview"), then the original itself: the document with Open the file at that page, the words of the suggestion a person approved, or, for an admin, the conversation it came from. Below that, what it says now, every version with who changed what and when, and how often answers read it in the last 30 days, so a note nobody reads is easy to spot. Something said that was not used, because what was on file stood over it, is in the list of versions too, marked Not used, with why.

Nothing on the card is guessed: it's what was recorded when each version was made. Something made before we kept records says so. If a note's original file was deleted, the card says when; the approved note is kept. A metric's question opens its own page, which has all of this and more.

What a definition is

Not just a name and a query. A definition has parts: what it means in the owner's words, the other names people say for it, where it lives, how it adds up, which date decides the month, the always-on filter (test orders out), what it's safe to split by, an example, and an owner, a version, who confirmed it and when.

Two ways of calculating the same thing are two named definitions. The name settles the argument, not the SQL. A change is a new version; every answer and dashboard stamps the version it used, so "the June deck used the old definition" is a fact you can say.

Any definition with a date splits by day, week, month, quarter or year, and by weekday, hour of the day or month of the year ("which weekday do we sell most?"), always from the same calculation. For timestamps with a time zone, months follow the company's time zone. Dates and timestamps without a time zone stay as recorded. Today is the company's today too: the last 28 days end on the company's date, and a number's receipt says it was worked out on that same day, wherever the machine runs (UTC's, until an admin sets the company's zone under Settings, Company). Comparing a named month with the previous period means the month before it: January compares with December.

A definition can be a running total (year to date, trailing twelve months), a share of the total (each store's share of sales), or one number worked out from others (margin as a share of sales, sales net of refunds). Each uses the same dates, filters and splits and carries a receipt. A running ratio divides the totals for its whole window; it never adds up the daily or monthly ratios. Nobody works these out on a page or in a chat.

A trailing month uses a calendar month ending on each date shown: through March 15 it counts February 16 to March 15; through March 31, March 1 to 31. Asking to see just one day still reads that whole window. A running calculation the system cannot support, such as a running median, is refused with the reason.

A chart by month never skips a month. A month with no rows shows as 0 (or empty, for a ratio or a median), and the line under the number says which months those were. A period with no rows at all says so, never $0.

How a number adds up is read from its calculation, not typed. A sum or a count adds up across stores and months. A ratio (margin as a share of sales) doesn't: the total is the ratio of the totals, and the line under the number says so. A median, an average or a distinct count doesn't either: each store's is worked out on its own rows, and nobody, agent or person, can save one as if it did.

Later: what we know we should improve

Not every good idea is a decision to make now. When the assistant, an agent or a person spots an improvement (this table should be a model, this metric wants a split by region, this page needs a fix), it goes on the Later list, the Inbox's Later tab: what, why, where it lives, who spotted it and in which chat. A card you answer "Later", including feedback on an answer, lands there too instead of vanishing. The list is grouped by area, so a working session is "take the models": Work on these in a chat starts one with them. An item remembers its working chat: Resume the chat picks up that conversation, while Start a new chat begins another. It says "in a session" only once a chat exists; done items stay as history.

It gets better as you go

Every question, correction, card and dashboard adds to this. The system you have after six months knows things the one you started with didn't, and every agent you connect reads the same memory.

Overnight it reads the day's chats again, once, cheaply: a thing someone said about the business that nobody filed becomes a guess (or waits for you, depending on the dial), a correction you made once ("no, net sales leaves out wholesale") becomes a card of its own with the change, your words and the chat it was said in, and said again it joins that card; a correction said twice becomes a rule to approve, a question asked three times in a day goes on the Later list, and what the chat looked for and couldn't find is counted in the morning line (the email an admin can turn on in Settings, Company). Only what a person typed is kept, in their own words: never something only the chat said, and never a number the chat worked out from a metric or a query, even if you repeat it (that number lives in the metric and moves with the data). Something already waiting for your yes is not counted as something it couldn't find. Tables with no card get one drafted from their columns and samples, marked not confirmed yet until a person edits it. Facts about someone who has left are closed with the date they left. What an agent on a laptop says to its person is not seen; what it files is.

On Free

Overscore's own AI only sorts what you add to the library on Free, on us, and your own Claude does the rest. So there is no Talk it through; no website is read; a document suggests no facts; the day's chats are not read again overnight, and a table gets no card drafted for it. Accuracy tests can be written there, and are checked on Self-serve and Done for you. Everything you or your own Claude tells it is kept, sorted and shown the same way.

What to say to it