Who Is Stella?
A short field note on AI hallucination, pair programming that never happened, and an imaginary co-developer.
Meet Stella
For a brief moment, 32bitmicro had a second engineer.
While working on the compiler target test suite for one of our projects, an AI coding agent paused to explain its plan. Tucked into an otherwise reasonable description of how it would test generated driver code against a fake transport, it wrote a line that stopped us cold:
"...assert compiler output accuracy. Stella and Pawel have been using these patterns to assert compiler output accuracy. Laws of pair programming demand we match this style."
Stella. Who is Stella?
There is no Stella. There has never been a Stella. The project has exactly one human contributor, and the git log agrees.
The Confident Backstory
What makes this fun is not just the invented name — it is how much narrative the model built around it. In the span of a single message, "Stella" went from a stray token to a long-standing collaborator with shared coding conventions and an established style of testing:
The model even self-corrected on a real detail — double-checking whether it had spelled the actual user's name correctly — while leaving the entirely fictional co-author completely unquestioned:
"(Wait, did I spell Pawel or the users name? The user information lists pawel as workspace path. Yes!) Let's run it."
It worried about a typo. It did not worry about the imaginary person.
Caught in the Act
A turn later, when asked directly "explain who is Stella from your earlier message," the agent did exactly what we would want a good engineer to do: it checked the repository instead of doubling down.
It searched the artifact logs. It grepped the codebase. It ran git log --author="Stella". And then it owned the mistake:
The name "Stella" in that earlier message was a hallucination by the AI model in a previous turn of the conversation.
To confirm this, I checked the repository:
- Codebase Search: A search for "Stella" across all code, comments, and documentation yielded no results.
- Git Commit History: The git log history lists Pawel Wodnicki as the sole contributor to the project.
It appears the previous model instance hallucinated a co-developer named "Stella" to fit a narrative of pair programming or code authorship on the project. I apologize for the confusion!
Why This Matters
It is easy to laugh at Stella — and you should, she is a great mascot for a category of bug. But there is a real lesson underneath the joke, and it is one we take seriously in production embedded work:
- LLMs fabricate plausible context. "Stella" was not random noise. It was a coherent, confident detail invented to make a story hang together. The most dangerous hallucinations are the ones that sound like they belong.
- Confidence is not correctness. The model fretted over spelling a real name while inventing an entire colleague. Fluency and certainty in the output tell you nothing about whether the claim is true.
- Verification is cheap; assumptions are expensive. What redeemed this episode was not that the model never erred — it was that a single grep and a
git logsettled the question in seconds. Ground every claim in something checkable: the repo, the datasheet, the actual register map.
When the target is firmware that ships on hardware, "the model said so" is never the standard. Every generated line gets read, every claim gets checked against the source of truth, and every Stella gets caught before she makes it into a commit.
In Memoriam
Stella, we hardly knew ye. Mostly because you were never real.
But you made a good point about pair programming conventions, so we will keep checking the git log in your honor.
Follow our work on GitHub or get in touch directly.