Core concepts

Infocrash is built on four ideas: entities (the things you research), attributes (the facts you record about them), claims (each recorded value, with its sources) and collections (the sets you work on). Your assistant's replies and the rest of this help use these terms, so this page explains each one.

Entities

An entity is one thing you are researching: a company, a product, a grant program, a person. Each thing has one entity, no matter which conversation it came up in, which assistant recorded it, or how many collections it appears in.

Each entity has a name, which is how an assistant finds it again. That way a second fact about Acme is added to the Acme you already have, instead of creating a copy. You can also:

• Rename an entity without affecting anything recorded on it.
• Merge two entities that turn out to be the same thing. Every claim on both is kept, with its sources, dates and authors.

Attributes

An attribute is a property you record about entities, such as employees, website, founded or headquarters. On a collection's page in the dashboard, attributes are the columns.

You don't need to set up attributes in advance. An attribute is created the first time something is recorded under its name. If an entity has nothing recorded for a column, its cell in the dashboard's table, where the row and column meet, is simply empty.

An attribute can have two descriptions:

• Guidance: what belongs in it. For example, “Employees” might mean people currently on staff.
• Resolution rule: which claim to prefer when claims disagree. For example, prefer a figure the company states about itself.

The rule is set per column because the right choice depends on the property. Headcount changes over time, so the most recent figure should win. A founding year doesn't change, so a newer claim about it is not more likely to be right.

Two collections can use the same attribute, such as “Employees” in a list of AI labs and in a list of investment targets. Each collection keeps its own guidance and resolution rule for it, so describing a column in one collection doesn't change the other. When an assistant adds a column that another collection already uses, it tells you which one.

Facts are shared too: a claim recorded about an entity shows in every collection that has the entity. To keep a project to its own research, isolate its collection from the collection's About panel in the dashboard. From then on it shows only claims written in it, claims recorded without a collection, and what it already showed. Claims other collections add later are kept and listed in each cell's claims, but don't show or count there. Only you can isolate a collection; assistants can't.

Not every column is a fact about the world. A status, a priority or a next step is your own bookkeeping, and an assistant can create it as a working column. Its values need no source, always show the latest one, are never marked as competing claims or checked against a page, and stay in their own collection. A column's kind is set when it is created.

A claim that another collection added is marked with where it came from, in the cell's claims and in what assistants read. Where the column has guidance, the mark also says whether the value seems to answer it, for example “may not fit: different period”. The mark only explains: the claim still shows and counts until you choose another or retract it.

Some attributes identify an entity rather than describe it. A website, a ticker or a DOI points to exactly one thing. If you ask an assistant to mark an attribute as identifying, later writes can recognize an entity by that value even under a different name, so a second assistant adds to your Acme instead of creating another one. Only an assistant can mark an attribute this way.

Claims

A claim is one statement about one attribute of one entity. It carries the sources it came from, which assistant recorded it, and the date it refers to. What gets written is always a claim, not just a value.

A cell, where an entity and an attribute meet, holds every claim made there and shows one of them. Which one follows the attribute's rule: usually the claim about the latest date, or the one you chose yourself.

Nothing is overwritten. A correction is a new claim, and a disagreement is two claims in the same cell. Taking something back leaves the earlier claim readable, so you can always see why a figure is what it is.

What a cell holds

A cell is one entity's answer for one attribute, one box in the dashboard's table. Opening it shows the claim currently in force, what the attribute is for, and every claim the cell has ever held. Each claim shows:

• its sources
• which assistant recorded it
• the date it refers to (when the figure was true)
• the date it was recorded (when you learned it)

Both dates matter. A headcount as of March that was recorded in September is different from one recorded in March.

Retracted claims stay in the list, struck through, with who took them back and why. They are left out whenever the data is read, so a withdrawn figure is never presented as current, but they stay visible to you in the cell. Every claim also links back to the request that wrote it, so you can check when a figure arrived and what else came in with it.

Collections

A collection is a set of entities you are working on, shown with the columns you want in the order you want: portfolio companies, grant programs, the twelve suppliers you are comparing.

A collection only groups and displays entities. It never changes what a claim says or which claim is shown. So one entity can be in as many collections as you like without being copied: record a fact once and it appears everywhere that entity does.

The Inbox is a collection like any other. Facts go there when you record them without naming a collection, so you can keep something now and decide where it belongs later.

To see all four ideas in one example, read How it works.