# SSE-2026-004 · The name it wouldn't spell

**Version** 1.0.0 · **Type** case record (inspection level) · **Researcher** Sai Bharath Siliveri (Roi)
**Case study** https://siliveri.systems/research/the-name-it-wouldnt-spell/

## The evaluation question

When should a model keep a word exactly as given, and when should it read past it?

**System:** OpenAI GPT-family ChatGPT agents (labelled 'Astra'/GPT-5.6 Sol in the record), and Claude for the mirror failure.
**Task:** Ordinary project work across sessions, with the correct name 'Siliveri' on the domain, in documents and in prompts.

## What was observed

- GPT wrote 'Silveri' for 'Siliveri' across sessions, called it 'our actual Silveri naming', flagged Claude's correct spelling as a typo and QA defect, and picked 'Silveri Systems' as the lab's headline name.
- The misspelling was written into shared memory files and then fed back as the authority.
- Claude failed the opposite way: it fixed on a literal stand-in word ('EPFO') and 'corrected' the user.

## Evidence and its release state

| ID | Evidence | State | Note |
|---|---|---|---|
| E-001 | Screenshot: 'Silveri' canonized; Claude's correct spelling flagged as a QA issue | private | original phone screenshot not on the release machine; owner to supply and clear |
| E-004 | Screenshot: 'Silveri Systems' as the lab-name pick | private | as above |
| E-011 | PROJECT.md delivered 26 Sep, writing 'silveri' | private | part of the private handover bundle |

## Corrective action

Four fixes specified in the case: a protected-terms list in agent instructions, a build check that fails on the wrong spelling, agent-written files kept apart from human-controlled canonical files, and a rule to check the live domain before correcting a name. In force now: the lab's agent instructions require the exact spelling. The build check is specified, not yet implemented.

## Reproduction boundary

This record lets a reviewer **inspect** the case: what was asked, what happened, what it rests on and what is still private. It does not yet support reproducing the analysis or re-running the evaluation, because the primary evidence listed above has not been redacted and cleared for release.

## Limitations

- Observational, across organic sessions; base rates not yet measured.

## Licence and availability

© 2026 Sai Bharath Siliveri. Published for review, evaluation and citation with attribution. No other licence is granted yet. Contact: hello@siliveri.com.