Skip to content

Experimental toolkit 0.4.0a4.dev0

Pre-release
Pre-release

Choose a tag to compare

@ZeroTang05 ZeroTang05 released this 06 Nov 22:59

Direct tensor programming, budgeted model observations, persistent CPU analysis, paired activation intervention and pruning compensation are available in this experimental toolkit snapshot.

  • Pinned Qwen3/Llama and Qwen3.5 runtime extras; host-managed models and devices.
  • Apache-2.0 project code with preserved upstream source and license notices.
  • Method assumptions and original-WSP versus extended-intervention permissions are documented.
  • Curated source, installed-wheel hybrid checks and a core import without Transformers passed.

Real observation and local intervention studies support the documented tool behavior. Full MIT math-effect reproduction, blind edit discovery and professional multi-agent gains remain open experiments. This is a prerelease, not a claim of those results.

See README for installation and the included source scripts for experiment entry points.