How it works

This page follows one example from question to answer. Say you are looking at a competitor and want to know how many people at Company A hold PhDs.

An example

You ask two assistants, ChatGPT and Claude, to find out. There is no single obvious way to count this, so between them they take three different routes. Each route is reasonable, but they give three different numbers:

Research path	Finding	Evidence used	Limitation
Company biographies	~18 PhDs	Public team biographies	Only team leaders might be featured; people without biographies are absent.
Professional profiles	8–12 PhDs	Education in indexed profiles	Public web coverage and self-reporting vary.
Publications and patents	30+ PhDs	Author and inventor affiliations	May include alumni, collaborators, or duplicate identities.

None of these findings is automatically right, and each method misses something different. In an ordinary chat you would see only one of the three, with no sign that the others existed.

One cell, three claims

In Infocrash, all three findings are stored as three claims in the same cell, the box in the collection's table where Company A meets PhD headcount, each with its own sources and the date it refers to. This is what your assistant reads back, and what you see in the dashboard:

entity           Company A
attribute        PhD headcount

guidance         People currently on staff who hold a doctorate.
resolution rule  Prefer a figure the company states about its own
                 staff. Counts inferred from publications only
                 when nothing company-reported exists.

value            ~18 PhDs            <- shown
    web          Company A - team page
    author       ChatGPT
    as of        2026-02-11

value            8-12 PhDs
    web          Indexed professional profiles (n=11)
    author       Claude
    as of        2026-01-30

value            30+ PhDs
    document     Author affiliations, 2019-2025 papers
    web          Patent register - inventor records
    author       Claude
    as of        2026-02-02

The two lines above the claims belong to the attribute, not to any one claim:

Guidance says what the column is for.
Resolution rule says which claim to prefer when they disagree, as they do here.

Settling it

You don’t have to settle a cell. Sometimes several answers is simply what the evidence looks like, and it is more useful to keep them side by side than to pick one. The cell keeps all of them either way.

If you do want one answer, ask your assistant to resolve the cell. It reads the rule, sees that one claim comes from the company itself, and makes the cell show that claim. The cell now shows that answer for a stated reason, not just because it was recorded last.

Nothing is deleted. The other two claims stay in the cell with their sources, so you can compare them, or choose another instead if you decide the rule was wrong.

Claims don't have to come from an assistant. A value you imported from a file is a claim like any other, and its source is that file.

What you keep

The result stays after the conversation ends. If you come back in a month, or hand the project to a colleague, the cell still shows the number, the reason it was chosen, the two answers it was chosen over, and the date each one refers to.