king-1 — Forecast quality (process realism)
Lever: better synthetic prior → lower geomean on held-out pool.
Code base: t_smo_v5-style generator (NumPy/SciPy).
Key ideas:
weekly_demand(~8%) for retail/load calendar structure- ARIMA-style correlated increments in
_integrated - RBF/RQ spectral GP on 2L circulant embedding (no wrap artifact)
- Prefix-causal measurement artifacts
- Forecastable pulse timing
Score with:
cascade verify ./king-1 --chain-toml chain.toml
cascade score ./king-1 --pool-dir ./my-heldout --device cuda --train-hours 0.5 --n-windows 256