Superseded and Unsupported Research Claims

A research direction can become unproductive because its central question has been answered, its motivating inference fails, or a claimed benefit disappears under a fair comparison. These are different diagnoses. This page examines specific claims and the evidence needed to reopen them; it does not maintain a blacklist of people, interpretations, or disciplines.

Helpful background. The scientific standards distinguish established results, approximations, interpretations, and open questions. A claim’s status must attach to its precise wording and assumptions.

Closed question, unsupported inference, or limited benchmark?

An original conjecture may be closed while its subject remains active. MIP*=RE yields a negative resolution of Connes’ embedding problem and separates appropriate closures of tensor-product correlations from commuting-operator correlations. A proposal to resolve the original conjecture should acknowledge this result. Constructive examples, restricted classes, quantitative bounds, and operational consequences remain possible research questions; see operator-algebraic frontiers.

A signal need not establish the proposed function. In photosynthesis, oscillations in a spectrum are insufficient evidence that long-lived electronic coherence enhances natural light harvesting. Cao and colleagues’ 2020 review emphasizes alternative vibrational interpretations and relevant timescales. The claim requiring revision is the inference from oscillations to that functional mechanism, not the existence of molecular quantum mechanics or the value of studying energy transfer.

A negative benchmark does not close an application class. Rønnow and colleagues found no evidence of speedup for the complete benchmark dataset they analyzed on a particular annealing device. That is a result about that comparison. It neither establishes universal quantum advantage nor rules out advantage for every future device, problem distribution, or quantum algorithm.

A theorem rules out its assumptions jointly. Bell inequalities constrain local hidden-variable descriptions under specified conditions. Their violation does not, by itself, select an interpretation or authorize controllable faster-than-light communication. A proposal framed as using ordinary entanglement alone to signal must identify which assumption of the established no-signalling argument it changes and give an operationally testable prediction.

Decide what evidence could change the assessment

For a closed mathematical question, identify a genuinely different theorem: changed hypotheses, a restricted model, or a quantitative refinement. Renaming the original question does not create novelty.

For an unsupported physical inference, write a competing model that reproduces the existing signal. Then identify an intervention or observable on which the explanations disagree. A better fit obtained by adding unconstrained parameters is weak evidence unless predictive performance and complexity are handled fairly.

For a technological benefit, fix the task and total resource boundary before collecting performance data. Include strong contemporary baselines and preserve unsuccessful instances. If a claimed gain survives only a favorable subset chosen afterward, the subset defines a hypothesis to test on new data rather than a completed demonstration.

Absence of evidence alone is not a falsification. Equally, declaring a proposal “not ruled out” supplies neither a mechanism nor a reason to prioritize it. A good topic-selection decision considers tractability and discriminatory power as well as logical possibility.

A quantitative check on an advantage claim

An algorithm that sometimes returns a valid answer must be compared at a specified success probability. If independent trials succeed with probability $p$, the number needed for success probability at least $1-\delta$ is

$$r=\left\lceil\frac{\ln\delta}{\ln(1-p)}\right\rceil, \qquad T_{\rm solution}=T_{\rm setup}+rT_{\rm trial}.$$

For $p=0.01$ and $\delta=0.01$, $r=459$. A hypothetical one-millisecond quantum trial therefore needs at least $0.459$ seconds of trial time at that confidence, before setup and verification. It can lose to a conventional 0.1-second computation even though its individual trial is 100 times faster. Conversely, reusable setup can be amortized across a declared batch. These are original illustrative numbers, not a reanalysis of the Rønnow dataset.

The assumptions are part of the test. Correlated trials, a drifting success probability, an uncertain estimate of $p$, or an unverifiable output require different accounting. Selecting only instances with large observed $p$ after running the benchmark changes the problem distribution. Test that new distribution on withheld instances before claiming a benefit. This explains why a failed broad claim may leave a narrower, useful research question.

For an experimental null result, make a similarly narrow statement: which model parameters would have produced a detectable effect with the actual noise and exposure? A confidence bound on those parameters does not exclude an entire family with unconstrained nuisance parameters. The opportunity may be a better discriminator rather than a more sensitive version of the same inconclusive measurement.

A graduate project: reconstruct one disputed inference

Choose one accessible claim with a primary paper, its strongest technical criticism, and enough methods or data to reproduce a central calculation. Begin by writing the narrowest statement both sides agree was actually tested. Reproduce that result before judging the broader interpretation.

The baseline is the published analysis with its original assumptions. On a workstation, change one contested assumption at a time: an alternative noise model, a corrected resource count, or a different statistical stopping rule. Keep agreement on all uncontested inputs. The deliverable is a reproducible comparison showing which conclusions survive and which depend on the disputed choice.

Use a staged six-week study. First reconstruct one figure and its units from public inputs. Then reproduce the strongest alternative using the same inputs and the same allowed tuning budget. Finally predict a withheld observable, instance family, or experimental condition on which the models disagree. Maintain a claim ledger with the statement, assumptions, evidence type, disagreement, and observation that would resolve it. If no feasible new discriminator can be found, the outcome is an identifiability result and a justified decision to stop, not proof that either author is wrong.

The failure criterion is that essential data are unavailable, the critique attacks a claim the original paper did not make, or both analyses fit the accessible observations equally well. Report those limits; none warrants an accusation of misconduct.

A research extension designs a new discriminator or proves a restricted theorem that resolves the ambiguity. It should produce new evidence, rather than another broad opinion about whether a field is promising.

Evidence horizon and entry reading

Reviewed 11 October 2026. The cases here use the 2014 Bell and speedup reviews/studies, the 2020 biology review, and the MIP*=RE manuscript revised in 2022. They are a bounded evidence guide, not a claim that all newer work has been assessed. Follow each linked frontier page for the live research question that remains.

Read each source with its conclusion and comparison domain together. A new counterexample to the stated theorem, an independently reproduced discriminating experiment, or a matched benchmark with a stronger baseline is an update trigger. A change in vocabulary or publicity alone is not.

References

• Brunner, N., Cavalcanti, D., Pironio, S., Scarani, V., and Wehner, S. “Bell nonlocality.” Reviews of Modern Physics 86, 419–478 (2014). Open review.
• Cao, J., et al. “Quantum biology revisited.” Science Advances 6, eaaz4888 (2020). Open review.
• Ji, Z., Natarajan, A., Vidick, T., Wright, J., and Yuen, H. “MIP*=RE.” Preprint (2020; revised 2022). Proof and versions.
• Rønnow, T. F., et al. “Defining and detecting quantum speedup.” Science 345, 420–424 (2014). Open manuscript.