zenfro_v3 — Grammar of Time

Cascade miner generator. Built on the full-context spectral stack, with a new prior: Grammar of Time (GoT).

Idea

Cascade holds the Toto2 model fixed and scores the data prior. Chronos-2, TempoPFN, ForecastPFN, and CauKer all win by composing temporal primitives — not by sampling a single ARIMA family. already covers many primitives as mutually exclusive draws. GoT allocates mass to productions that combine those stems:

Production	Syntax	What the model learns
got_compose	Stem ⊕ Stem [⊕ Stem]	Additive / multiplicative phrase structure
got_splice	Stem ‖ Stem	Clause boundaries = change of generating law
got_nested	Envelope ⋉ Carrier	Multi-scale nesting (slow modulates fast)
got_causal	Driver ▷ Response	Lagged temporal causal chain

Single-stem families from (AR2, integrated, OU-SV, spectral GP, …) are retained so the corpus still covers the dynamics-heavy mix that screened well locally.

Layout

zenfro_v3/
  generator.py      # Generator class (cascade.interface.DataGenerator)
  config.json       # active config (copy a variant here to A/B)
  requirements.txt  # numpy + scipy, hash-locked
  configs/          # ten A/B variants
    config_01.json … config_10.json
  README.md

How to A/B the ten configs

# pick a variant as the active config
cp configs/config_03.json config.json

# contract checks (determinism, layout, deps)
cascade verify ./mywork/zenfro_v3

# local heat score against a held-out pool (needs .[train])
cascade score ./mywork/zenfro_v3 --pool-dir <held-out>

Config knobs

key	role
family_weights	mixture over stems + GoT productions
got_depth	stems stacked in got_compose (2–4)
got_mul_frac	fraction of compose rows that multiply instead of add
got_splice_cuts	clause boundaries in got_splice
got_nest_ratio	slow/fast scale span for got_nested
got_causal_lag_frac	max lag as fraction of L for got_causal
artifact_scale	multiplies measurement-artifact rates
tr_* / gr_* / sa_clean_*	trend / growth / clean-season knobs

Contract

• Deterministic under a fixed seed
• NumPy/SciPy only, code-only (no shipped weights)
• Full-context series (min_length = max_length = 4096)
• Prefetch producer thread overlaps generation with training