Wrong answers from AI: what Delentia does and does not do
We do not claim to prevent hallucination. The figures that used to be on this page were not measured and have been withdrawn. This page says what the system does, what was measured, and what has not been shown.
What the system does
- Keeps wrong answers from causing damage: every tool call passes a safety gate, and risky ones wait for a human signature. This is about what the agent does, not about whether its words are true.
- Checks the final answer against the evidence the episode gathered: a number or address that appears nowhere in the goal, the conversation or any tool result, or a claim of an action no tool performed, makes the episode not count as a success.
- Offers a jury of independent models for questions where a second opinion is worth the cost (see the SignedAI page).
What was measured
- The answer check is weak. On real answers from a small local model it let 5 of 16 bad answers through (14 of 16 before tuning) and rejected 2 of 8 good ones. It only decides what the system learns from, never what you are shown.
- A later version was tried on a second set and was not adopted, because it did not beat the criteria written before the test.
What has not been shown
- That Delentia reduces wrong answers compared with the same model alone. This has not been measured, so no rate is claimed.
- That a jury of models is more accurate than one model. No real multi-vendor vote has been run.
- Anything with a capable model. Only scripted models and small local models have been used.