Pillar D · State as of 2026-07-23

Run 002: the crossover curve gets its second point (still not a trend)

Same fight, bigger instance: QAOA p=2 vs OR-Tools CP-SAT on a 16-asset portfolio, same Judge v1 protocol, seed 42. CP-SAT again reached the exact optimum (1.0 s); QAOA landed 20.7% away in 97.5 s. Verdict: not yet. The gap is smaller than at 12 assets (42.8%) — and we are explicitly NOT calling that a trend: one seed per size proves nothing. Sealed as EXP-0012-002.
→ Leer en español
State as of: 2026-07-23

What changed since run 001

One variable: instance size. 16 assets instead of 12 (choose 7, same risk-aversion and penalty structure, seed 42). Everything else is frozen by protocol — same 120 s budget per side, same QAOA depth (p=2), same solver on the classical bench, same exact-optimum referee (2¹⁶ = 65,536 candidates, still brute-forceable). If you change one thing, you learn one thing.

The result

Quality gap to the exact optimum · by instance size seed 42 · equal 120 s budget · one seed per size — not yet a trend 0% 50% 12 assets 16 assets 42.8% 20.7% QAOA p=2 CP-SAT (optimal) dashed line = visual guide only, NOT a fitted trend · runtimes: CP-SAT 1.0 s · QAOA 97.5 s
Measured results, sealed archives EXP-0012-001 and -002. The dashed connector is a reading aid, not an extrapolation.

CP-SAT: provably optimal again, in 1.0 second. QAOA: 20.7% from the optimum after 70 optimizer steps — the time budget, not the step count, was the binding constraint this time (97.5 s of the 120 s budget; the simulation cost per step grows with qubit count). Verdict: not yet — classical wins.

The honest reading of “the gap shrank”

At 12 assets the gap was 42.8%; at 16 it is 20.7%. It is tempting to draw a line through two points and announce that QAOA closes in as instances grow. We refuse, for three reasons:

  1. One seed per size. QAOA’s optimization landscape is noisy; a different seed can move the gap by a lot. Variance is unmeasured until we run multiple seeds per size.
  2. Two points define any story you want. A curve earns the name with density: more sizes, more seeds, error bars.
  3. The budget bound changed. At 16 qubits the optimizer completed 70 of 120 steps before the budget cutoff — the two points weren’t produced under identical effective conditions, and the protocol records exactly that.

What the two points do establish: the pipeline measures what it claims to measure, at growing scale, with every parameter on the record.

What’s next on the ladder

Multiple seeds at 12 and 16 to measure variance, then 20 assets — where brute force (2²⁰ ≈ 1M) still referees but CP-SAT starts to feel the size. Each run ships sealed, in triple copy, with its library entry. Wherever the curve goes, it goes on the record.

What we don’t know

Whether the gap movement is signal or seed-luck — that is precisely what the multi-seed runs will answer. We also don’t know where (or if) QAOA-family recipes cross CP-SAT on this problem class, and nothing measured so far suggests it is near. When we know, you’ll read it here first — with the raw data attached.

Measured content from sealed archives EXP-0012-001/-002. Rosetta Quantum publishes verdicts with reproducible raw data — including the ones quantum loses.

Sources:
· Sealed archive EXP-0012-002 (sha256:b9f9ab48…47fe4), RosettaQuantum/evidence", url: "https://github.com/RosettaQuantum/evidence" }