RETAINED RESEARCH / JUNE 18SPINFORGE / A GARMO LABS COMPANY
Compare the methods.
Keep the experiments distinct.
The supplied account compares GPU PyTorch statevector QAOA simulation with many-restart classical annealing on small random 3-regular Max-Cut graphs. It is not quantum-hardware execution.
NAMED RECORDS IN THE SUPPLIED ACCOUNTQuality, thresholds
Quality, thresholds
and application tests.
| Record | Experiment | Reported result |
|---|---|---|
| qaoa_vs_spinforge.json | Small-graph quality, N=8–16 | Classical ratio 1.0 on every disclosed graph. Simulated QAOA: p1 0.76–0.82, p2 0.82–0.90, p3 0.85–0.92. |
| qaoa_tts.json | Separate threshold study, N=10–18 | Classical: 1–3 restarts. QAOA depth needed to reach 90% recorded as >4 for every graph shown. Restart counts and depth are not wall-clock time. |
| moe_routing_qubo.json | 512–1,024 tokens; 8–16 experts | Greedy dropped 271–941 tokens per run; QUBO dropped zero with more even expert load on the described cases. |
| scheduling_qubo.json | 50–100 jobs; LPT comparison | Approximately a tie, within about 1% in either direction. |
| sigma_ln2_independent_verify.json | Overlap check, N=200–1,000 | Reported cut fractions 0.70–0.77 and a widening overlap gap. |
| leg1b_alpha_refit.png | Exponent refit; chart only | No numeric result is reproduced without the underlying fit data. |
INTERPRETATIONDo not flatten the studies
Do not flatten the studies
into one headline.
The quality experiment includes a depth-three ratio as high as 0.92, while the separate threshold study reports >4. Different records are not combined into a blanket “always below 90%” statement. The release account does not establish a general quantum advantage/disadvantage or a wall-clock time-to-solution comparison.
Source: owner-supplied experiment summary. The underlying JSON and chart files were not supplied for replay in this publication update. No new simulations were run.