CORTEX turns local models into an interpretability desktop with logits, attention, and interventions

JayB_Official · reddit · 2026-07-24

The author is building CORTEX // MODEL OBSERVATORY, a local Windows desktop app for mechanistic interpretability, with Codex + GPT-5.6 Sol Ultra helping prototype it quickly.
The app connects a WebView2 UI to an isolated Python/PyTorch backend through local IPC, and is designed to run fully offline with local Hugging Face models.
Current features include a Token Probability Microscope, Logit Lens, Representation Space visualizations, Attention Explorer, and an Intervention Lab for causal experiments such as activation patching and attention-head ablation.
Deep instrumentation currently supports GPT-2-family Hugging Face models and Llama models; Qwen and Mistral are mentioned as possible future targets.
A separate standard runtime can connect to OpenAI-compatible local endpoints such as LM Studio, though those endpoints do not expose internal activation hooks.
The author also mentions an experimental Jacobian-based concept analysis interface, but says it is not yet scientifically validated.
Development is happening on an RTX 4070 Ti with 12 GB VRAM, with the intended target being small local models in roughly the 0.5B–3B range.

More from Apps