| # zenfro_v2 — Grammar of Time on the v1 weather backbone |
| |
| Cascade miner generator. Evolves `zenfro_v1` (weather-heavy aurora blend) with |
| **Grammar of Time (GoT)** and forecast-friendly seasonality upgrades. |
|
|
| ## Why this prior |
|
|
| Cascade holds Toto2 fixed and trains it from scratch on your corpus. Eval is |
| forecast skill on held-out windows: **4096-step context → next 64 steps** |
| (CRPS + MASE). Winning synthetic priors (Chronos-2, TempoPFN, CauKer) compose |
| temporal primitives rather than sampling one ARIMA family. |
|
|
| | Layer | What | |
| |---|---| |
| | Stems (v1) | weather (~34%), trend/seasonal, AR2, integrated, RFF-GP, … | |
| | GoT (~22%) | compose ⊕, splice ‖, nested ⋉, causal ▷ | |
| | Seasonality | coupled calendar pairs (24↔168, 7↔52, …) + amp/phase drift | |
| | Clean mode | low-noise seasonal rows for sharp periodic reconstruction | |
|
|
| ## Grammar of Time |
|
|
| ``` |
| Series → Production | Stem |
| Production → Compose | Splice | Nested | Causal |
| Compose → Stem ⊕ Stem [⊕ Stem] # additive / multiplicative phrase |
| Splice → Stem ‖ Stem # clause = change of generating law |
| Nested → Envelope ⋉ Carrier # slow modulates fast |
| Causal → Driver ▷ Response # lagged temporal chain |
| ``` |
|
|
| GoT stems are cheap (seasonal+AR, RFF, integrated, AR2, diurnal-weatherish) so |
| compositions stay wall-safe under the token budget. |
|
|
| ## Layout |
|
|
| ``` |
| zenfro_v2/ |
| generator.py # Generator(DataGenerator) |
| config.json # active mixture + GoT knobs |
| requirements.txt # numpy + numba |
| README.md |
| ``` |
|
|
| ## Quick checks |
|
|
| ```bash |
| cascade verify ./mywork/zenfro_v2 |
| cascade score ./mywork/zenfro_v2 --pool-dir <held-out> |
| ``` |
|
|
| ## Config knobs |
|
|
| | key | role | |
| |---|---| |
| | `family_weights` | mixture over stems + GoT | |
| | `got_depth` | stems stacked in compose (2–4) | |
| | `got_mul_frac` | multiplicative vs additive compose | |
| | `got_splice_cuts` | clause boundaries in splice | |
| | `got_nest_ratio` | slow/fast scale span | |
| | `got_causal_lag_frac` | max lag as fraction of L | |
| | `sa_clean_*` | clean low-noise seasonal mode | |
| | `tr_*` / `gr_*` | trend / growth excursion | |
|
|
| ## Contract |
|
|
| - Deterministic at a fixed seed |
| - NumPy + numba only, code-only (no shipped weights) |
| - Full-context series (`min_length = max_length = 4096`) |
|
|
