cascade-king

A synthetic time-series generator for the cascade subnet (Bittensor netuid 91).

Miners submit code, not weights: the subnet owner trains a fixed Toto2-4M forecaster from random initialisation on whatever this repo streams, and scores it on a private, rotating pool of real-world series. The corpus is the only source of learned signal, so this repo is the entire submission.

generator.py         `class Generator(DataGenerator)` — the entrypoint
config.json          length band, block size, family mixing weights
requirements.txt     hash-locked, allowlisted deps (numpy, scipy)
kingprior/           the prior families
  spectral.py        stationary Gaussian draws by circulant embedding
  gp_fft.py          the kernel bank and KernelSynth-style composition
  sarima.py          SARIMA simulation in the frequency domain
  nonlinear.py       GARCH, regime switching, intermittency, chaos, ODEs
  structural.py      trend, seasonality, change-points, observation effects
  augment.py         the augmentation cascade
scripts/             throughput benchmarks, public-receipt reader
REPORT.md            the full situation report — read this first
HANDOFF.md           setup, the iteration loop, hardware, submitting
research/INTEL.md    the measured competitive picture in detail

Two things at once

The trainer stops at whichever comes first: a fixed token budget or a fixed wall clock. Realised training compute is therefore min(token_budget, throughput × wall_clock) — a corpus that arrives slowly is a corpus the model never sees. So the generator has to be both broad and fast.

Breadth comes from a two-stage mixture: a carrier drawn from one of eleven families, then structure and observation effects applied through random per-row gates. Speed comes from never generating a series on its own — every family is vectorised across a block, and every stationary Gaussian family (composed GP kernels, ARMA, long memory, quasi-periodic) is drawn with two batched FFTs rather than an O(L³) Cholesky.

Use

cascade verify .                                   # every check the trainer runs
python scripts/bench_generator.py . 2048           # points/s
python scripts/fetch_receipts.py --per-source      # the live public scoreboard
cascade score . --pool-dir ~/pool_v1 --device cuda # train + score (needs a GPU)

generate() is a pure function of (seed, n_series) — no global RNG, no hash(), no wall clock, no torch — so the corpus is byte-identical across runs and processes, which is what the trainer's determinism check and the round audit both require.