Research & Experimental Record
The hypervisor was built from hardware experiments, not assumptions.
Prometheus grew out of an experimental program examining how physical placement, routing cost, hardware conditions, and measured output interact on real quantum processors. The public research record contains five experimental programs, randomized instances, multi-QPU studies, IBM execution evidence, aggregate results, and explicit limitations.
5
Experimental programs in the public research record
The repository separates individual experiments so hypotheses can be tested under different workloads, scales, topologies, and physical processors.
A scientific development loop
Prometheus research does not assume that lower routing cost always means better hardware output. The experiments test that relationship directly, look for repeatability across randomized instances, examine scaling boundaries, and compare behavior across IBM Heron processors. The public record also retains mixed and negative outcomes rather than only favorable examples.
View the public research record →
01 · DISTRIBUTION
Distributional concentration
Tests whether physical routing cost and observed output-distribution behavior remain coupled on real hardware. One QFT-8 experiment on ibm_kingston showed a more concentrated output despite substantially greater physical routing cost.
QFT-8 · 100,000 shots · ibm_kingston
02 · ROBUSTNESS
Randomized instance robustness
Matched randomized QAOA and CrossEnt instances test whether observed behavior survives beyond manually selected circuits.
40 matched instances · N=6–9 · ibm_marrakesh
03 · SCALING
Micro-gradient scaling
Sweeps workload size and algorithm family to identify both favorable topological regimes and routing-penalty regimes rather than assuming that extra routing is always beneficial.
QFT / QAOA · scaling boundaries
04 · TOPOLOGY
Topological mapping
A 13-qubit topology-stress benchmark examined whether a substantially different physical realization could avoid the ground-state collapse observed in a depth-optimized baseline.
13 qubits · 80,064 shots · ibm_kingston
05 · MULTI-QPU
Three-QPU mechanistic sweep
Related workloads were examined across Fez, Kingston, and Marrakesh to test whether global device statistics are enough to describe the quality of a particular physical execution.
3 IBM Heron processors · workload-dependent study
Randomized validation
The experiments were pushed beyond one hand-picked circuit.
In the randomized-instance study, the same logical problems were independently compiled with SABRE O3, TKET, and Prometheus and executed on the same IBM Heron processor. Prometheus recorded higher measured Hellinger fidelity than SABRE O3 in 30 of 40 matched instances. The result is explicitly described in the research record as evidence that the observed behavior is not confined to a small manually selected set, not as universal superiority.
30 / 40Matched instances above SABRE
N = 6–9Randomized workload sizes
Why the research matters
Routing cost is only part of the physical story.
The research repeatedly tests a narrower proposition: minimizing routing cost alone may be insufficient to predict hardware execution quality.
- Higher routing cost can sometimes coincide with better measured fidelity.
- The sign of the effect can change with workload size and topology.
- Global device statistics can miss useful local structure in the physical graph.
- The same execution philosophy can behave differently on different QPUs.
Public evidence standard
The research repository is designed as an experimental record rather than a public release of the proprietary Prometheus implementation. It preserves enough evidence to inspect the experiments while keeping the execution-selection mechanism protected.
IBM result payloadsReturned execution results are retained where supplied in the research archive.
ProvenanceSafe metadata and SHA-256 evidence indexing support post-publication verification.
Mixed outcomesFavorable, mixed, and negative results remain part of the public experimental record.
Protected implementationProprietary formulation, weighting, heuristics, and physical trajectory artifacts are intentionally withheld.
The public research record explicitly does not claim universal superiority, guaranteed fidelity improvement, or that a single physical variable completely determines execution quality.