Havuz

To know what models don't say out loud.

Jacobian-Brainwash : A manual alignment tool for large language models built on Anthropic's Jacobian Lens. Results are exportable.

Interactive canvas for taking an LLM apart tensor by tensor: measured INT8/INT4/FP8 error, distributions, spectra and outlier channels for e

ICA Lens: compact ICA-based interpretability tools for exploring LLM activations. Code release for the paper.

ENGRAFT: write facts into an LLM's n-gram (Engram) memory table as a removable overlay: 100 facts in a 9 MB file, no weights touched, verifi

OpenAI-compatible server with a live view into any HF model's residual stream. pip install brainscope

End-to-end pipeline for seeing how LLMs actually process your prompts. Capture attention across every layer, render heatmaps and cooking cur

Agents that talk through model internals — activations & J-space — instead of text. No words pass between them. The lab on top of brainscope

Is there one workspace, or twin workspaces? Testing cross-lingual generalization of the J-space / global workspace in language models.

Recovering verbatim private reasoning from Claude extended-thinking signatures, with a local byte-exact verifier.

Watch a real AI model think - live 3D visualization of a transformer's actual forward pass: tokens, embeddings, attention, neurons, sampling

Models Take Notes at Prefill: KV Cache Can Be Editable and Composable (arXiv:2606.17107) — paper, code, results, and interactive companion.

Counterfactual replay and causal attribution over LLM-agent trajectories: find which step actually caused a failure by intervening on it.

Agents answer. ZugaMind notices. A persistent cognition sidecar that wakes Claude Code, OpenClaw, Codex, or Hermes when something deserves a

Mechanistic experiments on how LLMs represent and compute arithmetic internally, with strict no-parser controls, reproducible audits, and an