The Polycentria Dual-Helix on a problem the SHD-CCP geometry is genuinely suited to:
learned-heuristic routing over the 720-node golden Clifford-torus manifold. The geometric
oracle proposes routes; an exact certifier scores them. Because the objective
is a function of the geometry, the oracle decisively beats random search — the opposite of
the factoring case (ε ≤ 0).
Live race · oracle vs random
oracle best value—
random best value—
this race Δ—
Measured Oracle Bias ε
run the benchmark →
ε = mean best value(oracle) − mean best value(random), at equal evaluation budget.
Honest notes
Travel cost between packets is quadrance (squared distance) on the golden
Clifford torus (exact radii r1²=(φ+2)/5, r2²=(3−φ)/5). The certifier’s route
value is exact — the oracle is never trusted blindly. The full Python reference measures
ε ≈ +12.4 (many σ) on the exact ℝ⁴ manifold; a compressed-memory prior was measured too and
is marginal (within 2σ) — the geometry is the real win, not stored history.