cascade-heat10-var4-svcal-pool-rebalance

Heat10 is Heat9 code, mathematics, and throughput behavior unchanged, plus one conservative 1.5-percentage-point family-weight rebalance. It retains Heat9's empirical-Bayes stochastic-volatility calibration, Numba kernels, batching, prefetch behavior, requirements, and every other throughput optimization.

Exact weight changes

Final family weights:

• ar2: 0.105 -> 0.095
• chaotic: 0.010 -> 0.005
• physical_sensors: 0.085 -> 0.095
• seasonal_counts: 0.075 -> 0.080

Curriculum start weights:

• trend_seasonal_ar: 0.125 -> 0.120
• ar2: 0.130 -> 0.120
• physical_sensors: 0.095 -> 0.105
• seasonal_counts: 0.080 -> 0.085

Chaotic already has zero weight at curriculum start, so trend_seasonal_ar supplies the additional 0.5 point there. Both the final and curriculum-start distributions sum to exactly 1.0.

Integrated, OU stochastic-volatility, threshold AR, regime, spectral GP, and long-memory weights are unchanged. This preserves energy and high-frequency coverage as well as Heat9's mathematical correction.

Evidence and rationale

The all-snapshot evaluation contains 10 snapshots and 2,663 windows. Heat8's overall score was 0.34579, with MASE 1.40925 and MWSQL 0.08485. Heat8 beat Heat3 in nature (0.22394 vs 0.26586), healthcare (0.34820 vs 0.39200), and sales (0.71085 vs 0.84711). The 07-25 pool expanded from 292 to 462 windows: healthcare grew from 13 to 140 and nature from 135 to 158. Physical-sensor and seasonal-count semantics align with these growing slices.

The donor mass comes from generic AR2/trend structure and chaotic dynamics. Heat3's early curriculum placed much more mass on generic AR2/trend structure yet performed worse overall, while the final weights were identical. The chaotic final weight of 1% is exotic and weakly aligned with the expanding slices. This is indirect evidence, not causality: Heat3 also changed curriculum parameters, and its final weights were identical.

This rebalance is not a proven or guaranteed improvement. A production-faithful paired Heat9/Heat10 heat A/B with identical trainer configuration, hardware, seeds, and evaluation is required. Do not infer benefit from the retrospective slice alignment alone.

Validate

From the repository root:

cmp generators/07-26/cascade-heat9/generator.py \
  generators/07-26/cascade-heat10/generator.py
cmp generators/07-26/cascade-heat9/requirements.txt \
  generators/07-26/cascade-heat10/requirements.txt
python -m py_compile generators/07-26/cascade-heat10/generator.py
uv run ruff check generators/07-26/cascade-heat10
docker run --rm \
  --entrypoint /root/cascade/.venv/bin/cascade \
  -v "$PWD/chain.toml:/workspace/chain.toml:ro" \
  -v "$PWD/generators/07-26/cascade-heat10:/workspace/candidate:ro" \
  ghcr.io/tensorlink-ai/cascade-worker@sha256:f73c77276b80942c61e7c5570bbbafbd69c8fd15e79cffbbcf75c8855715c9a0 \
  verify /workspace/candidate --chain-toml /workspace/chain.toml

Also compare the parsed Heat9 and Heat10 JSON objects, allowing only name, description, and the eight weight paths listed above, and assert both weight sums with decimal arithmetic. Validation must not start training.