Realtime voice can drift from a supplied text file and invent source attribution

Summary

In a realtime voice session, I asked Codex to read a supplied local TXT file word for word and to pause only when I said “stop.”

After initially reading from the file, the voice output drifted into text that was not present in the TXT. It then treated that invented material as if it had come from the file, including unsupported evidence references and an attempt to locate linked source files.

Expected behavior

When asked to read a supplied file verbatim, the system should remain locked to the file’s exact contents. It should not blend in prior conversation text, generate missing paragraphs, or infer supporting source material.

Impact

This is especially concerning for legal, medical, financial, or other source-sensitive documents: the user may hear material presented as part of their document when it is not.

Requested safeguards

• A strict verbatim mode that reads only the selected file contents.
• Clear source boundaries between document text and conversation context.
• A guardrail that prevents evidence/source claims unless the relevant content is actually present in the file or separately verified.
• A visible or audible indication when exact playback cannot continue.

The report names no source files, tests, or entry points. First reproduce the realtime voice session with a local TXT file and compare the output with the file contents; done means verbatim playback stays within the supplied file, with unsupported source claims prevented or clearly flagged.