Changelog#

All changes are listed by pull request on GitHub. Everything below is part of the 0.1.0 alpha, which is not on PyPI yet.

0.1.0 (alpha)#

New methods#

  • Feature Generation: a language model writes binary rules as Python lambdas that then run on your data with no further model calls (#63), with optional cognitive-mode prompting (#68).

  • Verifiable RL: an adaptive, sequential information-gathering classifier (#72).

  • Reasoned Rule Mining: an interpretable if-then rule-mining classifier (#73).

  • Policy Induction (#32, #79).

  • Random Rule Forest (#15) and GPTree (#7), with shared LLM utilities (#2).

Random Rule Forest#

  • Elastic-net learned-weights aggregation (#74).

  • Concurrent predictions and checkpointing in predict() (#69).

  • Correlation-based answer similarity (#67).

  • cross_validate_aggregation for unbiased evaluation (#62).

  • F-beta as an aggregation metric (#61) and configurable F-beta question scoring (#50).

  • Prompt presets for domain-specific prompts (#60).

  • Cost-sensitive mode with early pruning (#58).

  • Founder-level aggregation (#55).

  • Early semantic filtering during fit (#54) and an exclusion_report for filtered questions (#53).

  • Batched question answering in training and prediction (#43, #52).

  • A flag to turn cumulative memory on or off during question generation (#42).

  • Clearer prompt instructions and logging (#16) and lower memory use (#17).

GPTree#

  • Bug fixes, lower cost and class-weighted Gini (#65).

  • Refactored prompts (#11).

Fixes#

  • Per-call log-probability retry flag instead of instance state (#66).

  • Sentence-transformer encoding and a Gemini log-probability fallback (#64).

  • xAI system instructions are now sent (#57).

  • Parameters are filled in the LLM fallback error message (#56).

  • Normalisation and error-message formatting (#33).

Tests and tooling#

  • Offline Random Rule Forest test harness with a fake LLM (#49).

  • Pyright set up for a local virtual environment (#71) and pre-commit ruff aligned with the dev dependency (#75).

  • Open-source polish (#20).

Documentation and website#

  • thinkreasonlearn.com served from this repository, with the docs under /docs (#86).

  • Docs built and served from a container (#22, #23, #25, #85).

  • README lists every method and uses current model IDs (#78), with earlier updates in #21, #28 and #30.

  • Docs theme, layout and contributing guide (#9, #12, #24, #26, #27, #31).