How it works: Two prediction systems work together:
(1) An
n-gram frequency predictor (cyan) tracks what you play after specific move sequences (window sizes 1-5).
(2) The
full SONA engine (purple, 227 KB WASM) records each round as a trajectory, trains
two-tier LoRA weights (micro + base) with EWC++ memory preservation, and transforms context
embeddings via
applyLora() to predict your next move.
The
combined predictor (white dashed) uses n-gram when available, falls back to SONA, then random.
Tip: Try repeating a pattern (e.g. rock-rock-paper-rock-rock-paper) and watch
the AI catch on. Then switch strategies and see the accuracy dip before it re-adapts.
Credits: SONA engine by
ruvnet/RuVector —
SONA crate.