Replication of Ribeiro 2026 (arXiv:2602.09047): select k = 28 municipalities from n = 88 in Goiás, Brazilian Cerrado, maximizing carbon + biodiversity + social impact with spatial synergies. Same open data, same portfolio score — different execution architecture.
This is not a vendor bake-off. We show that a multi-objective business task usually run as one weighted QUBO + hybrid loop or three separate quantum jobs can be submitted as one polycontextural IBM run: three objective contextures on shared qubits, coupled by transjunction bridges.
| Workflow | IBM submissions | Description |
|---|---|---|
| Published QAOA + ZNE | 1 (+ ZNE folds) | Ribeiro baseline: single weighted QUBO, warm-start QAOA, zero-noise extrapolation on ibm_fez / ibm_torino |
| Sequential 3-job hybrid | 3 | Carbon-only, biodiversity-only, social-only QAOA → classical vote merge (Classiq-style pattern) |
| QPC polycontextural | 1 | Carbon → transjunction → biodiversity → transjunction → social → shared mixer — one circuit |
| Method | Portfolio score | Notes |
|---|---|---|
| Greedy heuristic | 44.42 | Classical baseline in paper |
| QAOA + ZNE (mean of 7 runs) | ~58.47 | IBM hardware, January 2026 |
Classical replay on n=88, k=28 with the published pipeline (--matrix-indexing positional, matching their script):
Greedy score: 44.419 (published 44.42) · SA: 47.15
heron156: 3×52Q contextures · n=28, k=7 · 8192 shots · QPC noise reducer. Same Cerrado metric; instance size independent of Ribeiro n=88.
| Method | Score | vs greedy (5.53) | Job ID |
|---|---|---|---|
| Greedy | 5.528 | — | — |
| Weighted QAOA (28Q) | 6.002 | +8.6% | d8g9hd5v8cos73f2qvlg |
| QPC 156Q (3×52) | 4.998 | below greedy | d8g9k75v8cos73f2r2mg |
JSON: results/cerrado_compare_heron156_fez.json · v2: results/cerrado_compare_heron156_fez_v2.json · v3 decoder: results/cerrado_compare_heron156_fez_v3_decoder.json
v3 (ibm_fez): QPC beats greedy with marginal-vote + greedy repair: score 5.908 vs greedy 5.528 · job d8ga43tv8cos73f2rikg
v3 + ZNE (λ=1,2,3): decoder 6.272 (extrapolated) · λ=1 decoder 6.267 · jobs d8gj6t1e8nrc73bg1eh0, d8gj7042upec739jrgjg, d8gj73dv8cos73f37hbg
Backend: ibm_marrakesh (156Q) · 4096 shots · QPC noise reducer (per-qubit marginal readout + normalize)
| Method | Portfolio score | IBM job ID | Valid shots |
|---|---|---|---|
| Greedy (classical) | 44.419 | — | — |
| Published QAOA + ZNE (external) | ~58.47 | Ribeiro 2026 | — |
| Weighted QAOA (Ribeiro-style) | 42.356 | d8g8kblv8cos73f2q1ag | 13.8% |
| QPC polycontextural (one run) | 27.992 | d8g8kee6983c73dpg6t0 | 9.7% |
Honest read: on this first 88Q Marrakesh pass, the deeper polycontextural circuit (depth 146 vs 43) is more noise-sensitive than single-layer weighted QAOA. The architectural point stands — one submission encodes three objectives — but hardware tuning (ZNE, layout, p=1 angles) is the next step toward published ~58.47.
JSON: results/cerrado_compare_ibm_marrakesh.json
MPS Aer simulation (512 shots) on the same instance as Ribeiro 2026. Scores use the published portfolio metric; no ZNE (noiseless simulation).
| Method | Best portfolio score | IBM jobs | vs greedy |
|---|---|---|---|
| Greedy (classical) | 44.419 | 0 | — |
| Published QAOA + ZNE (external) | ~58.47 | 1 (+ folds) | +14.05 |
| Weighted QAOA (Ribeiro-style, Aer) | 45.532 | 1 | +1.11 |
| Sequential 3-job + merge | 44.419 | 3 | 0 |
| QPC polycontextural (one run) | 45.532 | 1 | +1.11 |
Architecture takeaway: QPC one-run and weighted QAOA reach the same best score on this Aer pass, while the sequential 3-job hybrid collapses to greedy after classical merge — three submissions without cross-context transjunction coupling.
Next step: --mode ibm --backend ibm_fez for hardware parity with the published study.
Artifacts: results/cerrado_compare_full88_dryrun.json, results/cerrado_compare_pilot_dryrun.json
Aer simulation confirms circuit plumbing before full 88Q IBM submission. Portfolio scores are on the pilot sub-instance (not comparable to published 44.42 / 58.47).
| Method | Best portfolio score | Jobs |
|---|---|---|
| Greedy (pilot) | 5.571 | 0 |
| Weighted QAOA (Ribeiro-style) | 6.002 | 1 |
| Sequential 3-job + merge | 5.571 | 3 |
| QPC polycontextural (one run) | 6.002 | 1 |
On this pilot: QPC one-run matches weighted QAOA best score while sequential merge stays at greedy — illustrating the architectural gap (3 submissions vs 1).
Machine-readable: results/cerrado_compare_pilot_dryrun.json, results/cerrado_compare_verify.json
cd site_release_2025_11_15
python3 vendor_benchmarks/cerrado/qpc_cerrado_compare.py --mode verify
python3 vendor_benchmarks/cerrado/qpc_cerrado_compare.py --mode dry-run --n 22 --k 7 --shots 1024
python3 vendor_benchmarks/cerrado/qpc_cerrado_compare.py --mode ibm --backend ibm_fez --shots 4096