Experiments

From a concrete question to evidence someone else can inspect.

Four laboratory notes report local llama.cpp RPC runs. Two of them are one series, Packet loss on llama.cpp RPC: a slow tail on short replies, then the same loss with generation held at 128 tokens. The earlier pair is a 50 ms egress delay and a worker killed mid-generation.

Each experiment will specify its question, system and failure model, baseline, procedure, measurements, and limitations, with links to the relevant code and data.