Hugging Face's logo

Executespec
/
laxmi-controller-experimental

PyTorch
Safetensors
software-state-prediction
experimental
negative-results
custom-model
real-patch-evaluation
Model card Files
xet
 Community
laxmi-controller-experimental / CHANGELOG.md
adminspec's picture
adminspec
Release v0.1.1 real-patch evaluation tooling
3e94469 verified
preview code
|
Raw History Blame Contribute Delete
1.35 kB

Changelog

0.1.1 — 1 October 2026

Adds a separate evaluation/real-patch/ tooling component within the same Laxmi release: an exact visible-source baseline wheel, source, versioned input contract and design-only private API draft. One exposed development case is scored correctly by exact execution; five other historical cases remain oracle-only. No learned real-patch model, blind result, hosted API, Linux exact-execution qualification or case dataset is released.

All four controller candidate weights, metadata, runtime wheel/source ZIP and retained examples are byte-identical to v0.1.0. The controller experiment's failed repair and scientific scope are unchanged.

0.1.0 — 29 September 2026

First public experimental controller package. Retains all four original September 15–16 candidates without selecting a winner or changing weights. Includes the negative development result, a smaller reviewed source closure, Apache 2.0 attribution, and a retained synthetic H1 installation/replay example.

The installed entry wrapper checks Python 3.12 compatibility instead of requiring an exact interpreter compiler/build string. The release still advertises only the verified macOS ARM64 installation scope. Original checkpoint source locks remain unchanged. No new training or prospective predictive-quality evaluation is claimed.