RLTrainner/generator-v17

generator_v17 — boundary-aligned forecast-task ablation

This candidate forks generator_v12 and preserves its deterministic single-stream prefetch path, fixed 4096-point rows, and established stochastic families. Its decisive change is a new forecast_tasks family occupying 45% of the final corpus and 60% of the clean early curriculum.

Forecast-task compiler

Each forecast-task row is a continuous composition of:

• deterministic quota-sampled periods from 7, 15, 24, 48, 60, 96, 144, 240, and 288;
• slowly varying seasonal amplitude and phase templates;
• smooth global trends and predictable local slope corrections;
• persistent AR residuals with slowly varying, context-visible volatility;
• sparse precursor → 64-step event → recovery episodes whose boundaries are exact multiples of 32.

The 32-point alignment matches Toto2's patch size, while the 64-step episodes match Cascade's two-patch evaluation horizon. Event sign, approximate magnitude, and volatility are visible during the preceding context; this teaches conditional forecasts rather than independent surprise corruption.

Period quotas use a 70% revealed-pool-frequency and 30% uniform blend. Counts are allocated exactly inside each generated family block and then shuffled by the generator's seeded RNG.

Curriculum and corruption

The forecast-task family starts at 60% with forecast_task_hardness = 0.15, then blends to a 45% share at full difficulty. Existing families retain the remaining 55% as a regression guardrail.

Compared with v12, final independent corruption is materially reduced:

• shock-row rate: 0.40 → 0.08;
• irregular-hold row rate: 0.25 → 0.08;
• irregular hold probability: 0.01–0.10 → 0.005–0.04;
• censor rate: 0.08 → 0.04;
• regular-hold rate: 0.08 → 0.04;
• TSMixup: 0.03 → 0.01.

Positive and count observation links remain, at slightly lower rates, because they model domains rather than independent corruption.

Determinism and throughput

generation_workers remains 1. The new family consumes only the existing numpy.random.Generator stream, uses vectorized NumPy operations plus the existing SciPy AR filter, and introduces no new dependencies or unseeded state.

Verification

cd /root/cascade
cascade verify /root/generator_v17

For a direct generation benchmark, instantiate both generators with the same seed, drain the same row count, and compare points per second. This benchmark tests generation only; it does not launch a heat evaluation.