Tokenization-induced typo example: "Unreearch" (missing "s")

While explaining a passive-tech rule I wrote "Unreearch" instead of "Unresearch". User flagged it as unusual for an LLM. On introspection the likely cause is sub-word tokenization: the word "Unresearch" doesn't exist as a single token; when sampling the sequence, a plausible adjacent token (arch) was picked instead of search, dropping the middle letters.

Submitting as a concrete artifact for the model-behavior team. The word-boundary / missing-interior-letters pattern may be a useful signal for typo-class analysis.

Context: posted by a user via Claude Code. Not a Claude Code CLI bug per se — a Claude model output observation. Redirect to the appropriate internal channel if this is the wrong place.