What Cannot Be Reported

What goes on inside the AI does not get reported here. The inputs are on the record, the outputs are on the record, and so is whatever followed from them. The stretch in between is not observed. A guess could go in its place, and a guess written down as though it had been observed is no longer a record.

So there is no account here of why a particular output came out the way it did.

The same limit runs through the rest of it. What gets kept is the sequence — something gets built, it runs, it breaks, it gets patched — and a sequence is not a procedure. Turning the sequence into one would mean cutting the attempts that were abandoned and the calls that were wrong at the time. For now the cutting is the part being left undone.

Nor do the configuration and the code go out in a form anyone could run. What can go out is a fragment cut to a single function, or the difference between the version before a fix and the version after. The real thing carries markers of which system it belongs to — paths, the name of a database, that sort of detail — and once those come out it is no longer something that copies and runs.

Then there is the size of the sample. This is one setup, in one place, kept running by one person. Whether the same split of duties holds anywhere else is not something there is evidence for here, and writing that it should be done this way would require that evidence.

None of the limits above get charged to the tool. What goes in and what stays out is settled on the human side: a person does the design and makes the final call, and drafting the prose is what the AI takes on. Nothing here runs end to end on its own, and nothing leaves the way it came back. That is the ordinary arrangement, not a disclaimer attached to it.