# Simulator-to-Hardware Transfer Gaps in Robust QAOA Parameter Selection

## Draft claim

Robustness-aware QAOA parameter selection reduces simulated noise sensitivity, but real-QPU pilots on `ibm_marrakesh` and `ibm_kingston` show that the simulated robust selector does not automatically become the best real-device selector. The paper should therefore be framed as a simulator-to-hardware transfer-gap study, not as a hardware superiority claim for RA-QAOA.

## Current evidence

| source | metric | value | conclusion |
| --- | --- | ---: | --- |
| 150-graph simulated benchmark | p=2 sensitivity delta | -0.412928 | robust selector lowers simulated sensitivity |
| 150-graph simulated benchmark | p=2 noisy E[C] delta | -0.143569 | robust selector pays an expected-cut cost |
| hardware-like proxy | p=2 sensitivity delta | -0.946489 | proxy signal also favors lower sensitivity |
| `ibm_marrakesh` QPU pilot | `noisy_expected` mean E[C] | 3.666016 | best real-device mean E[C] in 3 runs |
| `ibm_marrakesh` QPU pilot | `robust` mean E[C] | 3.656901 | close, but not mean winner |
| `ibm_marrakesh` QPU pilot | `noisy_expected` mean success | 0.847656 | best real-device mean success in 3 runs |
| `ibm_marrakesh` QPU pilot | `robust` mean success | 0.843913 | close, but not mean winner |
| two-backend QPU pilot | `noisy_expected` mean E[C] | 3.645264 | best mean E[C] across 4 QPU jobs |
| two-backend QPU pilot | `robust` mean E[C] | 3.597412 | lower than noisy and optimizer across 4 QPU jobs |
| `ibm_kingston` QPU pilot | `noisy_expected` E[C] | 3.583008 | best selector in the second-backend check |
| QPU metadata table | job calibration/layout | recorded | job-time calibration timestamp and recovered layout are now auditable |
| readout mitigation check | rank change | none | independent per-qubit readout mitigation does not change selector ranking |

## Interpretation

The simulated result remains meaningful because RA-QAOA deliberately optimizes a stability-aware score rather than raw expected cut. The QPU result changes the manuscript claim: real hardware can reorder selector performance because routing, calibration, queue-time drift, and sampling effects are not captured by the current simulator and hardware-like proxy. The second-backend check strengthens this interpretation because `noisy_expected` also leads on `ibm_kingston`. The metadata table improves reproducibility, and the simple readout-mitigation check shows that final independent readout error alone does not explain the selector ranking.

## Paper-ready conclusion

The strongest defensible conclusion is:

> RA-QAOA exposes a stability-performance trade-off in simulated QAOA parameter selection, but small `ibm_marrakesh` and `ibm_kingston` pilots show that simulated robustness does not directly transfer into real-device dominance. Robust QAOA selection should therefore be evaluated with separate simulator, proxy, and QPU axes.

## Required next evidence before a full submission

1. Repeat `ibm_kingston` to reduce single-job artifact risk.
2. Evaluate correlated/error mitigation or a calibration-aware selector as a separate follow-up.
3. Keep the result as a two-backend pilot/case study unless more hardware data is collected.