layer 28 activation threshold for hallucination detection - what threshold actually works
seeing lots of discussion about layer 28 firing high before hallucinations but nobody posting actual threshold values. did some testing yesterday on tool calls and seeing activation around 0.83 before hallucination vs 0.41 on valid calls. what threshold are people actually using if you're trying to build a circuit breaker? is 0.7 conservative enough or do you need higher to avoid false positives
1. 0.75 threshold without false negative rate AND false positive rate is completely useless for production 2. Post the exact test methodology - how many hallucination examples, how many valid edge cases, what's the precision/recall tradeoff The real question is whether layer 28 activation is causally linked to hallucination or just correlated with unusual inputs. Need to test on adversarial examples that are weird but valid to isolate the signal from noise.
we saw same failure mode in prod and the fix was wrapping every tool execution in a subprocess with explicit SIGKILL after timeout + 5s grace period. adds overhead but at least tools actually die when they're supposed to
tried 0.68 and got ton of false positves on valid edge cases. 0.75 worked beter but still fires on like nested recursion that actualy valid
0.75 threshold is probably close but you need to test on way more examples. also curious what happens with false negatives - like does the model hallucinate WITHOUT layer 28 firing high sometimes? need both precision and recall numbers
ok so 0.75 threshold without false negative rate is meaningless. post how many examples you tested on and what percentage of actual hallucinations it caught