Model occasionally fuses adjacent words into confident gibberish mid-sentence, with no self-correction

Summary

In an otherwise coherent sentence, the model fused several adjacent words into scrambled gibberish and delivered it at full confidence, with no self-correction.

Observed output

So we deliberately did not stuff lt withjyoursmectets, your history, and your whole life.

Intended output

So we deliberately did not stuff it with your secrets, your history, and your whole life.

The span "it with your secrets" collapsed into "lt withjyoursmectets" — adjacent tokens fused, letters transposed — mid-paragraph, surrounded by clean text.

Why it's worth fixing

• Silent failure: the broken span is produced fluently, with no hesitation or self-correction.
• Highest cost in trust-sensitive moments (this occurred while the assistant was explaining a security design to a user).
• Intermittent and not reliably reproducible, which is why the exact failing string is captured above.

Environment

• Claude Code (CLI), Windows
• Model: Claude Opus 4.8

Suggested direction

A lightweight output self-consistency / gibberish check could catch these spans before they reach the user.

Bonus context (with love)

The garble happened mid-sentence while the assistant was explaining that it was building the user's own bot a self-proofreading "sanity check" — a safeguard the assistant did not itself have. The user caught it in two seconds. Filed with appreciation. 🫠