Main idea

An assistant asked to add a row to a spreadsheet answers "it's done". That message attests to nothing: it says the assistant issued the operation, not that it succeeded, and above all not what was actually written.

For it to attest, a read has to be inserted. The order becomes: intent, write, read back what was just written, verify, confirm. The confirmation no longer bears on the intent but on the observed state.

The difference lies in what the answer displays: the content read back, not the content sent.

Why it matters

It is the same defect as self-assessment, transposed to a technical operation: you take an actor's report on its own action for proof of that action.

The remedy is cheap — a targeted read, on the single entry added — and it transforms the nature of what you receive: an assertion becomes an observation.

Nuances and limits

For long content, a full read-back cannot be displayed. You then display a summary of what was read back, explicitly distinguishing that summary from the stored content, which stays complete.

And the read-back doesn't cover everything: it attests to what is written, not to what should have been.

Open questions

• What is a read-back worth when the same agent both wrote and reads back?