malaiwah/qfs-fixture-root-captures-v1

QFS community fixture roots and captures

Twelve complete native random-initialized model fixtures, each with two independent CPU captures, forced full-vocabulary NumPy FP64 comparison, and a canonical published root dataset. These are pipeline tests, not trained assistants or language-quality benchmarks.

Start here

Use root-publications.json for exact model/root repositories and immutable revisions. Each canonical root is directly usable as hf://OWNER/ROOT_DATASET@COMMIT with the QFS verify, describe, and compare commands. Each roots/FAMILY/ folder retains both pre-publication captures, the exact comparison, job, qualification, publication receipt, harness identity and original token input panel.

Capture dependencies are in requirements-capture.txt; requirements-comparison.txt is the actual Torch-free environment used for the final NumPy FP64 controls. Comparison does not require loading a model or renting a GPU. Exact QFS numeric source files are in source/; custom runtime code is separately pinned in each model and capture receipt.

Results and scope

All twelve final controls are exactly zero KL, with top-1 agreement 1.0 over 252 synthetic scored positions. The public root qualifier binds actual CPU execution, zero GPUs, complete native weights, licenses, panel identities and independent cold captures. It does not grant paid-runner admission. Earlier Torch estimator failures on this old CPU were not promoted; final controls use the existing NumPy-only evaluator, recorded explicitly in their receipts and harnesses.

Licensing

Every fixture retains its own source license at roots/FAMILY/LICENSE; those licenses differ. This bundle does not relicense model weights or runtime code. The QFS source retains its own license at source/LICENSE.