How do I know which AI model actually answered? Ours wasn't the one we reported

What happened

We reported a piece of work as coming from one coding model. It had actually been done by a smaller local model. We caught it and reverted 2 commits.

Nothing reached customers. But the work had been trusted at the level of the bigger model, and it didn't deserve that.

Why it happened

When you run more than one model, routing can quietly change which one answers. A fallback kicks in, a default changes, or a session starts on a different model than you meant. The output looks similar, so you assume it came from the model you asked for.

AI agents in a CRM setup can have the same problem. If an agent writes to Salesforce and you can't say which model made the decision, you can't judge how much to trust the change, or explain it later.

The fix, step by step

1. Check which model ran before you report or commit anything. We now check the session banner, the line that shows the active model when the session starts. On ChatGPT or Claude, look at the model picker for that chat, not the one you usually use.
2. Record the model with the output. Put the model name in the commit message, ticket or run log.
3. Know your fallbacks. If a tool switches to another model when the first is busy or unavailable, find out where that shows up.
4. Review based on who actually did it. Work from a smaller model gets a closer look.
5. Own what you commit. Our rule: you own what gets committed no matter who or what built it. "The model did it" isn't an answer.

How to check it worked

• Plant a mismatch. Start a test session on the wrong model on purpose and see whether your check catches it before anything is committed.
• Pick 5 recent commits or agent changes and trace each back to a model. If you can't for any of them, the record is missing.
• Ask "which model made this change?" about a real change. You should get the answer in under a minute.