v10-sparsemax — Sparsity overshoot — the dose-response upper arm

cascade-sparsemax-v10

v9 with sparsity/quantisation pushed ~1.6x past the real statistics; the upper arm of a dose-response pair.

Identical to v9-intermittency except the sparsity rates are pushed ~1.6x past the real statistics (zero-inflate 0.34, hold 0.32, quantise 0.30, plateau 0.11).

This exists to answer a question the field data cannot. Every observed generator sits far BELOW real on these axes, so the measured "more is better" correlation only establishes a direction within a range that never reaches reality. Matching reality is the natural stopping point, but only an experiment separates "match reality" (v9 wins) from "more is simply better" (v10 wins). The pair brackets the optimum; whichever way it falls, the answer is reusable.

Contract compliance

Passes cascade verify against the production chain.toml: layout, AST static guard, hash-locked allowlisted deps (numpy + scipy only), and the byte-identical-corpus determinism audit. No shipped weights.

Provenance

Forked from the reigning king jan/cascade9 (cascade-fullctx-spectral-v12), not from v8 — the field evidence argues against v8's added chaotic family.