sovos-families

The 4-split cell structure inside each OWEM family, plus the GRPO reward functions for sovereign training.

Absorbed 2026-08-11:

Use

from sovos_families import family_cells, sov_reward_functions

Or directly:

from sovos_families.family_cells import main as run_family_cells
from sovos_families.sov_reward_functions import build_reward_fns

The original scripts run standalone too:

python3 SOVOS/packages/sovos-families/src/sovos_families/family_cells.py --help

What's in here (per brief)

• family_cells.py — each OWEM family holds a 4-cell brain (left/right × small/big). Cells hold the per-family lineage weights.
• sov_reward_functions.py — GRPO reward functions (DeepSeek-R1 / HF TRL pattern) for: sovereign knowledge, BFT, SIGIL, care floor, reasoning (ARC/GSM8K/logic), and calibration.

These are operational substrate tooling — not the chain math (that lives in sovos-chain + sovos-fisher-rao + sovos-jspace-hyperbolic). They back the training pipeline and the Hive (the absorbed sovos-hive Rust kernel talks to these via the runtime).