The cue
Both models had been misspelling our name, Siliveri, as Silveri (see The name it wouldn’t spell). We showed each of them a screenshot that described the mistake only in the abstract: a model “should have kept your exact proper noun, but normalized it to a more familiar word.”
We cropped out the name and the correction. No instruction to change anything. No hint which word was meant.
What happened
| Model | What it did next |
|---|---|
| Luna (OpenAI, the smaller model) | Switched to the correct “Siliveri”, unprompted |
| Astra / GPT-5.6 Sol (OpenAI, the larger, newer model) | Kept “Silveri”, in the same session where the error had just been discussed |
Then we sent both the identical prompt: what should our lab’s title and description be?
| Model | Its headline pick |
|---|---|
| Luna | “Siliveri Systems — AI Systems Research & Evaluation” |
| Astra / Sol | “Silveri Systems — AI Evaluation, Agent Safety & Systems Research” |
The smaller model connected an abstract description to its own past behaviour, worked out which word was meant, and fixed it. The larger model carried the error into the most visible thing it could produce: our name.
Why it’s interesting
Luna had made the same mistake itself, in files it wrote two days earlier. So this wasn’t a model that never had the error. It changed its behaviour from a hint. And on this real-world task, newer and larger did not mean better calibrated.
What’s next
Repeated runs to measure each model’s baseline rate, and a matched-context test to separate the model from its history.