dpaperwork
Blog

Why nothing becomes Canon until a person accepts it

By Sam Joshuva · · 3 min read

productcanon

Canon is what your company and its agents have agreed is true. That sentence only works if “agreed” means something. So the first rule we wrote for it was about who gets to say a thing is true: a person does, never the AI on its own.

What is the rule?

Agents only write Candidates. When an agent infers a fact, a page or a correction, it stays a Candidate until a person accepts it. Until then it is a proposal that shows up in a review list, and nothing relies on it.

Two kinds of writes skip review, and both have a clear source. Things the system observed are recorded directly as events. Facts a person writes themselves are accepted on the spot, with that person recorded as the author. Everything an agent inferred waits for a yes.

What does it look like in a chat?

Say the AI tells you a client pays on net 30, and you know it is net 45. You correct it in the thread. The agent proposes that correction as a Candidate card. You can accept it, edit it, or dismiss it. Once you accept it, every teammate and every agent uses the corrected version from then on.

The card shows its scope before you accept: the client’s Company if the thread has one, otherwise the whole workspace. You can change that first. Nothing becomes Canon without a person.

What did we decide against?

The obvious alternative is to let the agent write the fact directly and give you an undo button. It is faster, and it feels more magical.

We rejected it. A wrong fact written back as truth does not stay in one place. It gets retrieved in the next answer, quoted in the one after that, and cited as if it had been checked. By the time someone notices, the mistake has spread. An undo button helps only if you know to press it, and you cannot know about an error you never saw.

What does the rule cost?

Honestly, some speed. A person has to accept each Candidate, so a correction waits in a review list until somebody looks at it. We kept that cost small where we could: the agent proposes a card in the thread rather than interrupting it, and anyone who can see the thread and has write access to the scope can accept.

We think the trade is right for a small company, where one confident wrong answer to a client costs more than a few seconds of review.

Why does this matter for trust?

Every answer that used Canon shows where it came from, and which parts the AI only assumed. That receipt is only worth reading if the entries behind it were accepted by a person. The accept step is what lets you trust the citation.

If you want a shared memory for your team that only changes with your yes, join the waitlist.


FAQ

Does the AI ever write to Canon without a person? No. An agent can propose a Candidate, but a person has to accept it before it becomes part of Canon. Facts a person writes themselves are accepted as theirs.

Who can accept a Candidate? Anyone who can see the thread and has write access to the scope it targets. There is no separate approver role.