Case studies · Bug report
Our notes invented decisions. Here is how we found it and stopped it
A phone-store memo came back with “Decided: then let's ship on Friday”. Nobody said it. The words came from our own prompt.
What we saw
Testing the installed build, the notes for a phone-store memo listed DECIDED: “Then let's ship on Friday” and OWED: “Park fast charger shortage until the installer lands”. A lawyer's call got “we'll go with the second option” and “can you own X?”. None of it was said.
Why it happened
Our summary prompt explained what a decision sounds like by quoting examples: “then let's ship on Friday”, “park it until the installer lands”. Small local models (Llama 3.2 3B and 1B) copied the examples into real notes. The summary paragraphs above them were accurate, which made the invented items more convincing, not less.
The fix, in two layers
- The prompt now describes a decision and a commitment instead of quoting sentences a model can copy.
- A guard in code runs on every path to saved notes: a decision or action item must have most of its content words in the transcript, or it is removed, and an emptied section says “None recorded.” Notes translated from another script skip the word match and only lose the known leaked phrases, so a real translated decision is kept.
Before and after, same recordings
| Run | Invented items | Real items kept |
|---|---|---|
| Before · store memo, 3B | 2 | yes |
| Before · law call, 1B | 4 | yes |
| After · both recordings, 3B and 1B | 0 | yes |
What remains is ordinary imprecision: the 1B model attributed an email to the wrong person. VIBE records were never affected; every value they hold must already be on a transcript line.
Sources
- Llama 3.2 on Ollama · Ollama