aurora-mix-v5 — cascade challenger generator

A seventeen-family synthetic time-series prior redesigned around TempoPFN's full pipeline (arXiv 2510.25502). Native families use numpy/scipy; the genuine ForecastPFN component uses the vendored TempoPFN subset and its allowlisted runtime dependencies.

v5 — research-backed SDE coverage and dormant-family activation:

1. Regime-switching OU SDE. TempoPFN's ablation reports that removing its SDE prior worsens aggregate CRPS by 26%, the largest generator-level drop. sde_regime_ou follows its published parameter ranges: two persistent mean-reversion regimes, CIR-like stochastic volatility, time-varying mean/diffusion, calendar seasonality, and optional fractional innovations.
2. Previously dormant families now train. web_traffic and seasonal_level were implemented in the downloaded artifact but absent from config.json, so they emitted zero rows. They now provide positive count traffic and drift-neutral seasonal-level coverage.
3. Broad composition. ForecastPFN remains the largest component (30%), with GP/kernel (11%) and SDE (9%) as the other anchors. The remaining 50% is spread across fourteen complementary families, avoiding a narrow benchmark-shaped corpus.
4. Stratified interleaved emission remains enabled. Every consumed stream prefix approximates the configured mixture.

Study the live king first (cascade fetch)

Every committed generator is public and content-addressed. Pull the reigning king and diff your prior against the actual code that is winning right now:

cascade fetch king --network test --chain-toml chain.testnet.toml   # → ./fetched-king-uidN
python local_validator/validate.py --repo my_generator --king ./fetched-king-uidN

--king accepts any generator dir, so the compare stage scores your feature-space coverage against the live king, not just the shipped base_generator. You win by improving on the visible best — a byte-identical copy is dropped before it trains, so you must genuinely beat it.

Strategy vs the king (the TempoPFN redesign)

The king vendors ten TempoPFN families but drops the audio-inspired group (pyo runs an audio server and seeds via hash() — non-deterministic, not on the allowlist) and leaves TempoPFN's augmentation layer mostly unused. v2 attacks exactly that gap:

1. Keeps the highest-signal families (per the TempoPFN ablation), with its own implementations: compositional GP/kernel priors (KernelSynth-style, coarse-grid Cholesky + cubic upsampling) and rich trend × multi-seasonality × structured-noise composition.

2. Reimplements the king's missing TempoPFN audio priors deterministically (pure numpy/scipy):

   family	TempoPFN prior it restores	what it teaches
   rhythm	Stochastic Rhythms	quasi-periodic event trains: tempo drift, accent bars, swing, dropouts, pluck/percussive/bump kernels
   fractal_multi	Multi-Scale Fractals	piecewise-slope spectra, Weierstrass sums, multiplicative cascades (multifractal volatility)
   net_diffusion	Network Topology	forced diffusion on random directed graphs — propagation, echoes, superposition (CauKer flavour, no networkx)

   Financial Volatility (TempoPFN's fourth audio prior) is already covered by regime_garch.

3. Ports TempoPFN's augmentation layer (the vendored king has the params but doesn't use most of it): per-series time-warp, damping envelopes, and spike injection in post-processing, plus batch-level splice transitions (two series crossfaded at a changepoint — TempoPFN's transition_ratio) and the TSMixup-style mixup family.

4. Keeps the seven regime classes the king does not emit (from v1):

   family	real-world shape it teaches	literature
   chaotic (Lorenz/Rössler/Duffing/Mackey-Glass/logistic/Hénon)	nonlinear dynamics, traffic/weather transfer	DynaMix (arXiv 2505.13192)
   fgn (fractional Gaussian noise / fBm)	long-range dependence	classic Hurst
   regime_garch (Markov-switching AR + GARCH)	volatility clustering, regime shifts	econ/finance
   intermittent (zero-inflated counts)	retail demand, sparse events	Croston
   calendar (daily profile × weekly factors × holidays)	energy/traffic load curves	—
   growth (logistic/Gompertz/lifecycle)	adoption curves, saturation	Bass
   bursts_anomaly (subcritical Hawkes + anomaly-injected AR)	bursty events, robustness to outliers/level shifts	—

5. Scale/offset/quantisation diversity in post-processing: log-uniform scales over five decades, offsets, integer quantisation (count-like series), softplus positivity — the from-scratch model must learn scale-robustness from the corpus alone.

6. Length mixture biased long (55 % of crops in [768, 2048]) — more signal per series while keeping short-series coverage.

Determinism

The corpus is a pure function of (seed, n_series). Every RNG is a np.random.default_rng(SeedSequence([seed, tag, index])); there is no global RNG, no torch, no wall-clock, no hash(). Verified by cascade verify (two full draws, digest-compared) and by a cross-process digest smoke (~650k points/s on a CPU box — well above the trainer's 185k tokens/s reference rate, so stream_cpu never starves the GPU).

Verify / test / score

# from the cascade repo root
python local_validator/validate.py --repo my_generator          # CPU gauntlet (valid + regime coverage vs king)
python -m pytest my_generator/tests -q                          # contract tests
cascade verify ./my_generator                                   # the real pre-deploy check

# the REAL signal (GPU, upstream `cascade score`): does it actually beat the king?
bash local_validator/vps_score.sh                               # on a GPU VPS — A/B train+score vs the fetched king

cascade score trains the fixed Toto2-4M on your corpus at the heat budget and scores it (lower geomean = better), so you learn whether you clear the 0.02 win margin before spending a hotkey. See local_validator/README.md for the GPU VPS setup + a sufficient Vast.ai spec.

Deploy (when ready)

cascade deploy ./my_generator --hub-repo <namespace/aurora-mix> \
    --wallet-name <coldkey> --wallet-hotkey <hotkey>
# optional outage fallback (Hub is always tried first; --hf-repo alone is refused):
#   --hf-repo <hf-namespace/aurora-mix>   # needs HF_TOKEN; commits repo@hf:<sha>

Remember: one hotkey = one submission, for life. Deploy only after the local gauntlet and a GPU cascade score A/B that clears the 0.02 margin vs the current king (bash local_validator/vps_score.sh).