7.15.0 - the meaning layer
styxx.meaning_diff - the meaning-regression instrument. Did two models MEAN the same thing? Given two models concept representations, get an agreement score, a HEALTHY/DRIFTED/BROKEN verdict, the concepts that diverge most (named and ranked), and a reliability flag that says when not to trust the comparison.
from styxx.meaning_diff import meaning_diff
r = meaning_diff(model_a_reps, model_b_reps, words=concepts)
r["agreement"] # Qwen-1.5B vs Qwen-3B = 0.93 (HEALTHY); vs a shuffle = 0.02 (BROKEN)
r["verdict"]; r["divergent_concepts"]The use case nobody has a clean tool for: model-migration / quantization / distillation / fine-tune regression QA - did the new model keep the meaning of the old one, and which concepts did it lose? Norm-equalized by default (the convention validated tonight; the unweighted template average understated agreement 2-60x). Pure stdlib + numpy, installs in the core wheel, no torch. Gates D1-D5 ALL-PASS, 11 tests.
This release ships alongside the night's certified research: the norm-domination apparatus fix, radical translation between minds (GAVAGAI), structure-only novel-concept transmission (TELEPATHY), the species/scale/matcher boundaries, and B23-F (claude-fable-5: the first frontier conduct certificate, FRONTIER-RESISTANT). Every claim pre-registered before its data and OATH-certified.
Full details: CHANGELOG.md.