There’s a strong temptation, in any answering system, to always produce something. Silence feels like failure. But a fabricated answer doesn’t just mislead the person who asked — it gets repeated, cited, and eventually treated as fact.
So we drew a bar. Below a confidence threshold, backed by too little evidence, Memra declines. It tells you the corpus can’t support an answer rather than inventing one.
The honest failure mode
The refusal isn’t a shrug — it’s specific. It names what it looked for and why the evidence fell short, so you know whether the gap is in the question or in the record.
- Not enough sources — the topic isn’t covered in the connected corpus yet.
- Conflicting sources — the evidence disagrees, and picking a winner would be guessing.
- Stale sources — the only matches are too old to trust as current state.
A system that knows the shape of its own ignorance is one you can build on. One that always answers is one you eventually stop believing.