enigma-spectre/multilingual_j_lens_lite

J-Lens Lite is a research script for peeking inside an open-weight AI language model to see how it handles the same idea expressed in different languages. It asks a simple question: when you ask a model the same thing in English, French, or Japanese, does the model's internal processing look similar across languages for that concept, or does each language get handled in its own separate way. The tool works by feeding the model equivalent prompts in several languages, then measuring how strongly each internal layer influences the model's chosen answer token, using a technique based on gradients, a standard way of measuring influence inside a neural network. It records these measurements layer by layer and produces charts and a written report comparing the patterns across languages, along with raw data files for further analysis. To run it, you set up a Python virtual environment and install the listed dependencies, then run a probe script pointing at a small instruction-tuned language model and a configuration file describing the languages and target concept to test, such as the capital of France. A small web interface built with Streamlit lets you browse the results afterward. Bundled example configurations already cover English, French, Spanish, Italian, German, Mandarin Chinese, and Japanese. The README is careful to state what this tool is not. It is explicitly not a full research-grade reproduction of the underlying academic technique it is inspired by, not a way to detect consciousness in a model, and not proof that a model is explicitly thinking in words. It notes that tokenization differences between languages can distort the comparisons, and that its gradient measurements are local and specific to the exact target being tested rather than a universal measure. This is a niche research tool for people already familiar with how language models work internally.