Executive Summary
This page previously reported that QCOS Autopilot reached fidelity 0.90 on a GHZ-6 circuit in a median of 4 evaluations, against 133 for SPSA (33.25x), over 3 seeds on the QCOS CPU state-vector simulator. Sales material stated the same comparison as 25–47x. Internal review on 25 August 2026 found that the setup favoured the optimiser. The arithmetic was right and the records verify, but the comparison did not measure what it claimed, so we withdrew the result rather than restate it.
The Challenge
GHZ state preparation is a common benchmark for near-term quantum computers. On real devices, high fidelity means overcoming obstacles inherent to NISQ (Noisy Intermediate-Scale Quantum) hardware. The withdrawn benchmark ran on a simulator, so these effects were never part of it:
- •Gate errors — imperfect gate operations
- •Decoherence — T1/T2 noise causing information loss over time
- •Readout errors — measurement inaccuracies
- •Crosstalk — interference between neighboring qubits
Variational quantum algorithms call the circuit many times before they converge, making the number of evaluations a critical metric for practical quantum computing.
QCOS Autopilot Architecture
The QCOS Autopilot is a closed-loop control system that runs on the backend you name. The withdrawn run used a CPU state-vector simulator:
What Review Found
The same GHZ-6 circuit and simulator under three setups (median evaluations to fidelity 0.90):
| Setup | Optimiser | Result | Backend |
|---|---|---|---|
| As published: start near the answer, fixed parameter order | QCOS Autopilot vs SPSA | 4 vs 133 (withdrawn) | CPU state-vector simulator |
| Same setup | A published optimiser of the same kind | 4 (identical) | CPU state-vector simulator |
| Random start in [−π, π], random parameter order | QCOS Autopilot vs the same published optimiser | Tie | CPU state-vector simulator |
Why We Withdrew It
A Head Start
Every run began close to the answer, and the parameter that sets the GHZ state was the first one tuned, in a fixed order. An optimiser that tunes one parameter at a time could hardly miss in that setup.
A Known Method
A published optimiser of the same kind, implemented from its paper, reached the same 4 evaluations. The figure measured the class of method, not anything specific to QCOS Autopilot.
Intact Records, Flawed Design
Both evidence records from 4 July 2026 still verify, and they disagree: one shows 33.25x, the other 0.31x. A sealed record shows a result is unchanged; it cannot show that a comparison was fair. Review caught that.
What Replaces It
The replacement benchmark is designed so that its setup cannot favour any optimiser:
- Random starts: every run starts from parameters drawn at random in [−π, π]
- Random order: the parameter that sets the answer lands at a random position in each run
- Failures counted: a run that misses the target counts against the method at the full evaluation budget
No result from it is published yet, and until one is, we cite no optimiser benchmark figure. See benchmarks for what we publish and how each result is recorded.
Document Information
Document Type
Withdrawal Notice
Version
2.0
Classification
Public