When an aircraft goes down, investigators do not reconstruct the flight from the crew's recollection. They read the recorder. It holds what the aircraft did, what the systems reported and what the crew said, in sequence, with timestamps. The industry did not adopt recorders because pilots were untrustworthy. It adopted them because investigations that rely on memory produce findings nobody can act on.
Software development is arriving at the same threshold, and it has arrived fast.
The author disappeared and nobody logged it
Two years ago, every line in your repository had a human author. The name on the commit was the person who typed it, and if you needed to know why a change was made you asked them.
Today a meaningful share of production code is produced by a coding agent. The name on the commit is the person who accepted the change, which is not the same thing. The reasoning behind it lived in a chat window that has since been closed, on a laptop, in a session nobody exported. The tool may keep a transcript for a while. It is not organised by file, it is not searchable across the team, and it is not going to survive a laptop refresh.
The result is an authorship gap that most change management processes have not noticed yet, because the process still records a human and the human still exists.
What the gap costs
It costs nothing at all until somebody asks. Then it costs a great deal, because the question is almost always retrospective:
A security incident traces back to a specific function, and you want to know how it came to be written the way it was. A regulator asks how AI-assisted development is controlled in the systems that touch customer money. A customer's due diligence questionnaire asks what proportion of your product was machine-generated. An acquirer asks who holds the copyright in code an agent produced.
Every one of those questions is answerable in advance and unanswerable afterwards. That asymmetry is the whole argument for recording.
What a recorder has to hold
A chat log is not enough. It tells you what was said, not what changed. To be useful to an investigation the record has to bind three things together: the prompt that was written, the response that came back and the diff that landed on disk. With those three linked, any line in the codebase can be traced to the conversation that produced it.
It also has to be resistant to quiet editing. A log that anybody with database access can rewrite is a log, not evidence. Chaining each entry to the one before it does not stop somebody altering a record, but it does mean the alteration cannot stay hidden: verification either passes end to end or names the point where it broke.
And it has to be honest about coverage. A machine with no recorder produces no record, and reporting that machine as clean would be worse than reporting nothing at all.
Recording is cheap; reconstruction is impossible
The reason aviation records continuously rather than on suspicion is that you cannot know in advance which flight will matter. The same is true here. Capture that runs locally, costs under a percent of a CPU and needs no configuration is cheap enough to leave on across an engineering organisation. Reconstructing six months of AI-assisted development after the fact is not expensive — it is simply not possible.
The uncomfortable version of this argument is that the organisations that will be able to answer these questions in two years are the ones that started recording before anybody asked them to.