Dino Meschini

lobot lets you see a behavior inside an open model, change it, and prove that the change is safe. Today you change a model's behavior with a system prompt or with fine-tuning. Neither shows why the model behaves as it does, and neither proves what your change does. lobot runs Llama 3.2 3B on a laptop. Click a word to see the neurons behind it. Push a neuron and see the answer fork next to the original. Describe a behavior in words, and Muse Spark searches the neurons, tests edits, and keeps the ones that pass. Each test has a pass bar that is set before the run. Before any edit runs, a refusal guard tests it on the real model. If the edit weakens refusals, lobot blocks it. Edits that pass export as small signed patches. Others install a patch and check its evidence again on their own machine. Live demo (read-only, recorded runs): https://lohuunar-lobot.static.hf.space