You're asked who made the call, and you can't say.
“Humans are still accountable for the code that ships, whether they reviewed it or not.” Something breaks, and you're asked who decided it. You can't say whether a person chose it or a model suggested it and nobody pushed back.
In a chat, “sounds good” means two different things.
In a chat, your choices and the model's suggestions mix into one long transcript. “Sounds good” can mean “I decided this” or “I didn't read it”. Once the chat is closed, nobody can tell those apart, and the model can't either.
What doesn't work.
“Co-authored-by: AI” lines. They tell you a model touched the change, not who made the call.
Keeping chat logs. The line exists, but nobody can find it, and it's mixed in with everything else.
Asking the model later. It rebuilds a plausible answer. It doesn't remember.
What actually helps.
Most of this isn't us.
When a person makes a call, write it down in their own words, with the date. Don't paraphrase it.
Keep a short list of decisions a person made, separate from the build notes.
Make “I didn't decide this, the model did” an acceptable answer, and fix the gap when you hear it.
A ruling, kept in the operator's own sentence.
We keep a person's calls as their own kind of record, quoted word for word. On 13 October 2026 the operator looked at a setup step that only worked from a terminal and ruled on it in one sentence. That sentence was recorded exactly, and the work built because of it cites the ruling.
It works in the other direction too. In July a model proposed a new database table for scheduled work. The operator saw it duplicated something that already existed, and the corrected spec says so in its first line, so anyone reading it knows a person made that call.
| What was measured | What the record says |
|---|---|
| the operator's ruling, verbatim | “Why are we doing this from CLI? This would immediately kill a novice.” |
| recorded | 13 October 2026, 22:10 UTC, as a ruling |
| the corrected spec's first line | The operator caught that the proposed table duplicated a primitive that already existed |
| recorded | 14 July 2026 |
| where it's from | Vinculum's own project record, which runs on Vinculum |
Where a shared memory fits.
Vinculum stores rulings as their own kind of entry, quoting the person, so “did a human decide this?” has an answer you can open.
A model still writes the entry. If it paraphrases the person badly, the record is wrong, so the quote has to be exact.
Common questions.
Isn't a Co-authored-by line enough?
It records that a model touched the change. It doesn't record who decided, which is the question you get asked when something breaks.
Can't I just ask the model who decided?
It doesn't remember the conversation. It reconstructs a plausible answer from what it can see now, which is not the same thing as a record.