v0.2.1

[0.2.1] 2026-11-08

Bug fixes

• Report the released version and keep version strings in step on release (#7)

Documentation

• Describe the supply chain accurately (#4)
• Complete citation metadata and bump it on release (#6)

Maintenance

• Latest toolchain and CI versions (#5)

v0.2.0

[0.2.0] 2026-11-02

Performance

• Numpy autodiff engine with the same hand-written gradients (#3)

Documentation

• Standard README structure and language framing (#12)
• Horizontal layered architecture diagram
• Compact horizontal architecture diagram
• Remove authorship statements from docs and scripts
• Recorded demos of more commands, with their vhs tapes (#1)
• Recorded demos of the HTTP service and more commands (#2)

Build and CI

• Run CodeQL on main and on a weekly schedule only

Maintenance

• Add release automation, AGENTS.md, llms.txt and contributor docs
• Classify squash merges in the changelog by the commits they contain

v0.1.0

First release of a decoder-only transformer and reverse-mode autodiff engine in pure Python.

Highlights

• Forward pass matches PyTorch to within 1e-9 (opt-in validation group)
• Character-level training on public-domain prose: 3.64 bits per character versus a 4.30 unigram baseline
• Checkpointing with exact save and restore
• A pre-registered research study with 8 seeds: copying is carried by first-layer heads looking back a fixed distance, not an induction circuit; one of three hypotheses is not supported, two are mixed. See docs/RESEARCH.md

Install: pip install the attached wheel.