gen-d4871f4eba85-top8_v2
Top-eight fork of gen-d4871f4eba85. It applies the pruning idea from
forever-real_tired3-5-top8, but selects families from this parent's own
configured weights.
Active top-level families:
- k14: parent weight 0.310099
- k15: parent weight 0.210951
- k03: parent weight 0.185441
- k00: parent weight 0.042055
- k21: parent weight 0.040000
- k22: parent weight 0.040000
- k10: parent weight 0.033967
- k01: parent weight 0.032350
Their parent weights sum to 0.894863 and are renormalized to one without changing their relative proportions.
The variable-length update from forever-real_tired3-3-variable-lengh is
applied as a cyclic 2048/3072/4096 schedule. Full 4096 context remains the
default; two smooth pulses expose 0.5% 2048 and 1% 3072 rows at their peaks.
min_length = 64 calibrates schedule progress to the trainer request contract,
while emitted rows stay in the three configured buckets. The parent's family
curriculum remains disabled.
Unselected kernel functions remain only where useful as implementation helpers
(for example, k10 and k14 use k07 internally); they are absent from
_FAMILIES, dispatch, and configured weights, so they cannot be selected as
top-level training families.
The throughput-only update from generators/07-27/proven_speed_up.md is also
applied. Numba compiles the AR(1), AR(2), and threshold-AR recurrences with
fastmath=False; deterministic warm-up runs before corpus RNG creation. The
optimized kernels are bitwise-equivalent to the original SciPy/NumPy paths and
do not change the generated corpus.