Live findings from TAI Creative's quantum lab: entanglement verification, optimization benchmarks, QRNG entropy runs, and honest write-ups of what worked and what didn't.
10243 experiment runs (all time). 0 validated advantages (hardware + controls). 5/28 standing questions answered. 38 problem domains covered.
No crossover exists in N=6-16: p=1 QAOA is not competitive with classical local search at any tested size, so there was never an advantage to lose. It does beat uniform random by a stable 1.31-1.37x, so the circuit performs real quantum work — it is simply far weaker than a trivial classical heuristic.
depth_helps_but_insufficient. Depth genuinely improves QAOA quality (+0.16 to +0.18 approximation ratio from p=1 to p=6 at every size, significant at all four), but it never reaches classical parity, and and depth is expensive on hardware. **CORRECTED 2026-08-15 — less prohibitively than this conclusion originally stated**. It claimed a p=6 n=16 circuit (288 two-qubit gates) needed <=0.24% error against a Bell-derived 4.1%, implying 5.8e-6 fidelity. Live calibration gives 0.28% real cz error, so that circuit sits near 0.45 fidelity unrouted and roughly 0.25 with typical routing — degraded and marginal, not hopeless. The recommendation is unchanged but rests on the simulator result alone: Max-Cut QAOA should be retired as a route to quantum advantage because it never beats classical local search, not because it cannot be executed.
**CORRECTED 2026-08-15 — the original verdict was wrong.** It read 'wall_binds_at_p1' on the strength of a two-qubit error of 4.1% inferred from TAI's Bell fidelity. Live IBM calibration puts the real median cz error on ibm_fez at 0.28% — a ~15x overstatement, because a Bell circuit also carries readout error (0.85%, three times the gate error) which does not compound per entangling gate. Re-measured with real calibration AND real SWAP routing (transpiled to ibm_fez, 1.25x-2.19x overhead), FIVE OF SIX configurations execute p=1 fine: treatment_scheduling n=8 (fidelity 0.946, depth 12), vehicle_routing n=10 (0.848, depth 4), vehicle_routing n=20 (0.642), portfolio n=10 (0.617), job_scheduling n=12 (0.506). Only portfolio n=15 (459 routed gates, 0.276) is past the wall. **What survives unchanged:** the density mechanism (|E| = C(n,2), measured 4.2x growth for 2x n), and the separate finding that Kyon's flagship TSP builds no QUBO at all.
Rejected, on both sizes, with an early futility stop. Quantum beats random cleanly — 25/25 sub-instances on n6, 25/25 on n8 — but loses to the corrected classical baseline by a wide, consistent margin: n6's upper-95% bound on quantum/classical is 0.745, n8's is 0.758, both far below the 1.0 parity line the decision rule requires. Not competitive.
Worth wiring up, conditionally — but it has never been run. First: ZNE has NEVER executed in TAI's codebase, so Q9's ~280 accumulated runs contain zero ZNE data. Second, measured under a depolarizing noise model at equal budget, ZNE is a large real improvement ABOVE a circuit-fidelity floor of roughly 1e-2 (turning 15% relative error into 6%), and its benefit collapses below it (37% -> 34% at 216 gates). Third, it can HURT when noise is already low, because its extrapolation variance exceeds the bias it removes. This widens the executable envelope at the small end of Kyon's domains without changing Q13's verdict for the large ones.