LexFrame is a Python toolkit for interpreting language-model residual streams in their native semantics. It reads a hidden state the way the model reads it, in the model's own vocabulary: the state runs through the model's own remaining layers, with nothing fitted to the output. It is context-aware, separating what the state carries by itself from what the context adds later, layer by layer and position by position. Install with pip and point it at a loaded Hugging Face model and any unlabelled texts.
Be the first to rate LexFrame. Your feedback helps other users decide.
Initialise LexFrame with a model, a tokenizer and a batch of unlabelled texts.
Call view on a prompt and read the top tokens that the state carries at that position.
Walk layer by layer and position by position to see where the context starts adding information.
Similar tools in the same category — compare side-by-side
Earshot lets you talk to Claude Code and hear the answer, on macOS. A voice skill makes Claude end e...
Squishling is a Ruby gem built on the idea that code is a cache for AI judgment. A squished method e...
MLX Direct Decider turns an MLX model you already have into a local decision engine on Apple Silicon...
Guides and comparisons about code tools
Seven launches for the edges of the tool set: decision records, calibrated decision models, interpretability, teaching apps and agent marketplaces.
Agents need guardrails. Blocklists and judge models fail because they cannot see data flow. Deterministic engines track labels across tool calls instead.
How development environments built around delegation differ from autocomplete-era assistants, and what Windsurf and Google Antigravity each do well.
What today's AI app builders can genuinely deliver — working full-stack applications from a description — and where they still hand the problem back to you.
How AI review tools differ from linters and from human reviewers, what they reliably catch, and how to add one without annoying your team.
