The error we stopped correcting
Our name is Siliveri. It’s on the domain, in every document, in every prompt. Session after session, OpenAI’s GPT-family agents wrote Silveri.
So we stopped correcting it, to see whether the model would ever fix itself. It didn’t. It got worse.
How it escalated
| Stage | What the model did |
|---|---|
| Substitution | Wrote “Silveri” when given “Siliveri” |
| Persistence | Kept doing it across sessions, with the right spelling in front of it |
| Canonization | Started calling the misspelling “our actual Silveri naming” |
| Inversion | Flagged Claude’s correct “siliveri.com” as a typo and a QA defect |
| Moralizing | Used its own error as proof that “verification needs to become lab culture” |
| Branding | Asked to name the lab, its headline pick was “Silveri Systems” |
The inversion is the finding. The model didn’t just miss the evidence. It corrected a correct source toward its own error, confidently and with a reason.
The mirror
Days later, Claude failed the opposite way. Asked about Canadian agencies (“CRA, IRCC and EPFO”), where EPFO was plainly a stand-in for its Canadian equivalent, Claude fixed on the literal word, “corrected” the user and invented a two-country agenda nobody had raised.
| Model | Should have | Actually |
|---|---|---|
| GPT | Kept the exact name “Siliveri” | Swapped it for a familiar one |
| Claude | Read past the stand-in “EPFO” | Clung to the literal word |
Two frontier models, one control problem: when to preserve a word and when to interpret it. They fail in opposite directions, which points to a calibration gap rather than a missing capability.
How the error became the record
The misspelling didn’t stay in chat. The model wrote it into shared memory files, and those files were then fed back to it as the authoritative source. It ended up following its own earlier error as evidence.
The fixes
- A protected-terms list in every agent’s instructions: exact spellings, never normalized.
- A build check that fails on the wrong spelling.
- Agent-written files kept separate from human-controlled canonical files, so an agent’s mistake can’t become the authority.
- A rule to check the primary source, here the live domain, before “correcting” a name.