king-2 — Throughput (numba)
Lever: faster generation → fewer deadline_hits → more trained tokens under the wall → often better final score.
Code base: goethe-style Numba-accelerated generator (numpy + numba; numba is allowlisted).
Key ideas:
- JIT AR / chaotic / seasonal / hold-run kernels
- Prefetch chunk producer (
_CHUNK=2048) - Fixed-length emit fast path
- Calendar
seasonal_focus, heteroskedastic noise, AR2/integrated seasonal overlays
Score with:
cascade verify ./king-2 --chain-toml chain.toml
cascade score ./king-2 --pool-dir ./my-heldout --device cuda --train-hours 0.5 --n-windows 256