The question
Helios-Dx asked something specific: do variational quantum circuits offer any advantage over classical layers for privacy-preserving medical inference under fully homomorphic encryption constraints?
The setup made the question answerable. A frozen ConvNeXt-Tiny backbone extracts features; an aggressive bottleneck reduces 768 dimensions to 4 so the trainable head — roughly 30 parameters — fits homomorphic compilation with Concrete-ML. The only variable is the head itself: a plain Linear(4 → 2) against a 4-qubit variational circuit with strongly entangling layers. Same optimizer, same seeds, same data. Capacity-matched by construction.
The answer
No consistent quantum advantage was observed. The two heads were functionally equivalent in capacity and performance, and the quantum head cost more to run under simulation.
That sentence took months to earn, and there were three ways to write it up:
- Reframe it — "a novel hybrid quantum-classical pipeline for encrypted medical inference." Every word technically true, and the framing a lie.
- Abandon it — let the repo trail off with no conclusion, like most experiments that disappoint.
- Close it — lead the README with the disclaimers, tag v1.0, add a
CITATION.cff, and stop.
I closed it. The README's first section is a warning box: this project reports null results; no quantum advantage is claimed; simulation-only; not clinically validated.
What the discipline bought
The negative is informative. Anyone considering VQC heads under FHE constraints for small classification tasks can read one repository and skip the experiment. That is the entire value proposition of publishing research, applied to a result that happens to be a "no."
It changed how I design experiments. Knowing the result would be published either way forced the design to be defensible either way: capacity-matching, fixed seeds, an explicit split between the classical client and the quantum-FHE server. When you might have to publish "nothing happened," you make sure "nothing" is measured well enough to mean something.
It calibrated every other claim. After you have published a null result, "78% accuracy" on another project reads differently — because the reader knows you would have said 52% just as plainly. My ASTRA README flags its weakest class (Solar-like, F1 0.52) as not production-ready. That habit came from Helios-Dx.
The uncomfortable part
Publishing a null result feels like publishing a failure, and for about a week it did. The reframing option was genuinely tempting — the pipeline was novel, the engineering was real, and nobody audits a student project.
The counterargument that won: an engineering record is an instrument, and an instrument that only reads high is broken. If the record can't say "no," its "yes" means nothing.
A closing note on scope
Honesty includes the limits of the study itself. Helios-Dx ran on PennyLane's default.mixed simulator, not quantum hardware; four qubits is tiny; one dataset family. The null result is scoped to exactly that. It doesn't prove quantum ML useless — it proves this configuration offered nothing, and says so in a way someone else can build on.
Closed at v1.0. Citable. Done.