How the evaluator is put together, and which part is actually responsible for the answer. The short version: retrieval gathers evidence, deterministic code decides, and the language model explains. Those three jobs are kept separate on purpose.
The verdict comes from deterministic code, not from the model. The model writes the prose around it. If the model reaches a different conclusion it is recorded as dissent and shown, but it does not change the answer. This is the difference between a tool you can audit and a tool that sounds confident.
One evaluation, start to finish.
Finds the closest known algorithms to the problem description, using keyword and vector search together.
Supplies evidence. It does not decide anything. A close match is not proof that a problem is quantum.
Applies fixed rules to the retrieved evidence and the problem text, and produces the verdict and the recommended platform.
This is the part that decides. Same input, same output, every time. It is plain code with unit tests, not a model.
Checks proven speedup, I/O cost, error-correction overhead, whether the problem is naturally quantum, and whether the crossover is reachable.
A problem has to survive all six to be treated as a quantum candidate.
Writes the human-readable assessment and cites the retrieved references.
It explains the verdict, it does not set it. When the model reaches a different conclusion, the disagreement is recorded rather than silently applied.
Prices a reference run on quantum, AI/ML and HPC using live retail rates, falling back to pinned rates when the pricing service is unreachable.
Per-unit rates across three different billing models. Context, not a verdict.
Emits a quantum kernel for quantum verdicts, or infrastructure-as-code for classical ones, plus a resource estimate.
Optional. Off by default because it roughly triples the response time.
The browser can ask questions. It cannot write to the knowledge index. Everything behind the API authenticates with a managed identity rather than a key, and there is no ingestion path reachable from the public web. That matters because the index is what the router reasons over: anyone who could add documents to it could change the verdicts.
This page names service roles, not deployed resources. Instance names, hostnames, deployment names, index names, subscription and tenant identifiers are all left out on purpose. Knowing the shape of the system should not tell you where to knock.