Evidence Standards and Falsifiability

Falsifiability in quantum gravity is layered. An observation can exclude a coefficient, a model instance, a mechanism, or—only when the prediction follows from its core assumptions—a broader programme. The distinction prevents a null time-delay result from becoming “quantum gravity is ruled out,” while also preventing a programme from evading every adverse result by changing its bridge to observables after the data are known.

Required background. Equivalence Principles and Gravitational Measurement supplies the operational definition of gravitational observables and apparatus response.

Helpful background. Flat-Spacetime Quantum Field Theory supplies EFT control and radiative consistency; Einstein Equations supplies the classical null model; Quantum Information supplies channels and witnesses; Quantum Gravity supplies programme scope; and Gravitational Radiation and Waveforms supplies likelihood and population methods.

Results and limitations

This discussion reflects results available through 2026-08-20. The rows compare claims with different logical strength; parallel placement does not imply equal empirical support.

Claim family and evidence date	Degrees of freedom and observable	Evidence or controlled result	Adverse evidence or decisive failure test	Scope, limitation, and unresolved boundary
Low-energy quantum GR, 2023 source status	Quantized metric and matter fluctuations; nonanalytic long-distance amplitudes	EFT fixes selected infrared terms order by order for energies and curvature below the cutoff Donoghue 2023, §§2–6	A precise gauge-invariant amplitude disagreeing with the fixed nonanalytic coefficient in its regime would falsify that low-energy quantization framework	Accessible experiments are far from many clean effects; agreement would not determine the ultraviolet completion
Modified propagation, 2009–2026 status	Photon or graviton dispersion coefficients; arrival time and polarization	GRB 090510 gives a source-assumption-dependent null bound on named dispersion terms Fermi LAT and GBM Collaborations 2009, pp. 331–333 and table 1	A redshift-scaling delay that disappears under intrinsic-lag marginalization is adverse; a null result excludes only the declared coefficient and energy dependence	Emission histories, operator mixing, helicity dependence, and deformed versus broken Lorentz symmetry
Gravity-mediated entanglement, 2025–2026 theory status	Two quantum masses plus a gravitational or composite mediator; entanglement witness	A commutative mediator with alternating local completely positive channels cannot entangle initially independent systems Ludescher et al. 2026, §III, corollary 1 and lemma 2	Aziz and Howl exhibit virtual-matter contributions they argue permit entanglement with classical gravity Aziz and Howl 2025, eqs. (8)–(10), pp. 815–817; scaling and channel decomposition must discriminate	No accepted gravity-mediated-entanglement dataset; locality, composition, and the identity of the mediator remain contested
Primordial features, 2026 data status	Initial state, background, and perturbations; correlated CMB spectra and bispectra	Planck PR4/NPIPE CamSpec searches use explicit oscillatory templates and a global trials treatment Raffaelli, Ballardini, and Barbieri 2026, §§3–6	Local $\Delta\chi^2$ gains reach roughly $-15$, but the largest PR4 global significance is $1.7\sigma$ and there is no Bayesian support	Template choice, reheating, foregrounds, cosmic variance, and the bridge from a microscopic state
Compact-binary strong-field tests, 2026 data status	Binary and remnant geometry; residuals, polarizations, PN phase, and QNMs	GWTC-5.0 applies seven tests to a selected cumulative sample of 168 events and finds overall GR consistency LIGO–Virgo–KAGRA 2026, §§2–8	Waveform-family, glitch, calibration, mode, and population changes must not remove a deviation; current residuals and parameterized tests show no significant excess	A generic deviation is shared by classical modified gravity and exotic compact objects and is not a unique quantum-gravity signature

The conservative position requires a controlled derivation, a regime-valid bridge, and a unique cross-channel pattern before using “quantum-gravity evidence.” A permissive position counts exclusion of every local commutative gravity mediator as evidence that gravity has nonclassical degrees of freedom, even though no ultraviolet theory is selected. A pragmatic intermediate position treats such a result as a major low-energy discovery but reserves programme-level support for predictions that distinguish microscopic alternatives. The 2022 phenomenology review likewise emphasizes the map from framework to operator and from operator to observable rather than universal Planck-scale slogans Addazi et al. 2022, §§2 and 5.

At the cutoff, these positions remain genuinely different. The disagreement is not cured by saying all approaches are equally plausible: controlled EFT predictions have a firmer derivational status than unconstrained noise ansätze, while present observations have not selected a microscopic programme.

The four-link prediction chain

A testable claim should expose four maps:

$$\begin{gathered} \text{programme}\longrightarrow\text{controlled model} \longrightarrow\text{observable template}\\ \longrightarrow\text{predicted data}\longrightarrow\text{data likelihood}. \end{gathered}$$

The first map declares fields, states, symmetries, approximation, regulator or contour, and correspondence limit. The second derives a normalized template and theory error. The third applies propagation, source population, apparatus calibration, and selection. The fourth specifies data version, likelihood, priors, nuisance parameters, trials, and decision rule.

A failure at the last link may be an instrumental problem. A robust exclusion of one template falsifies that model instance. A mechanism is excluded only when all admissible states and nuisance choices preserve the failed relation. A programme is threatened only when the mechanism follows from its defining assumptions in a controlled regime and no alternative internal realization changes the prediction. Each claimed falsifier should name the exact level in advance.

This structure also limits positive claims. A statistically secure residual establishes a data-level discrepancy; a calibrated cross-observable relation may establish a mechanism; agreement with a microscopic calculation supports a programme only after viable conventional and competing quantum explanations are compared.

A solvable robustness check

Suppose a mechanism predicts two independently measured, dimensionless observables

$$y_1=A, \qquad y_2=qA,$$

with the shared sign $q=+1$ fixed by its symmetry. Measurements give $y_1=1.0$, $y_2=-1.0$, each with Gaussian standard deviation $\sigma=0.1$. The best fit is

$$\widehat A =\frac{y_1+qy_2}{1+q^2}=0,$$

and

$$\chi^2_{\min} =\frac{(y_1-\widehat A)^2+(y_2-q\widehat A)^2}{\sigma^2} =200.$$

The shared-sign mechanism fails catastrophically even though two unrelated amplitudes could fit both observations exactly. This is why cross-channel relations are more discriminating than flexible one-channel residuals.

Now change a named assumption: an alternative initial-state boundary phase changes $q$ from $+1$ to $-1$. Then $\widehat A=1$ and $\chi^2_{\min}=0$. If both phases are admissible within the programme, the data falsify only the first boundary prescription. If independent consistency fixes $q=+1$—for example a symmetry or unitarity theorem makes it state-independent—the same data falsify the mechanism. The conclusion must follow the robustness analysis rather than the desired label.

Preregistration and adverse tests

Before unblinding a signal region, freeze the observable definition, theory template, validity domain, data and calibration versions, quality cuts, foreground or environmental model, priors, population selection, trial family, and stopping rule. Publish mock-data recovery and null-injection performance. A calculation should survive independent algebra, limiting cases, and variations of gauge, regulator, truncation, state, matching, or boundary condition that the theory says are unphysical.

Background completeness is mechanism-specific. A photon lag requires source-emission models and energy response; a tabletop witness requires electromagnetic, Casimir, mechanical, and control channels; a CMB feature requires foreground, mask, beam, transfer, and look-elsewhere tests; a black-hole residual requires glitches, calibration, waveform families, higher modes, and population selection. “No known background” is not a statistical model.

The most useful adverse result is one that attacks the mechanism’s distinctive correlation. Reverse a predicted sign, change baseline or mass scaling, move to a second redshift, measure the conjugate polarization, or use an independent detector technology. A larger significance in one channel with failure in a locked companion channel can count against the theory.

Updating claims without moving the target

Every mutable conclusion should carry a dated cutoff, exact source locator, and explicit update condition. A new dataset can tighten a coefficient without changing a programme’s status; a theorem can instead change which observation would be decisive. Withdrawn results, calibration revisions, and changed likelihoods must propagate to derivative claims.

The conclusions could change if an independently reproduced cross-channel anomaly survives preregistered background and calibration variations, if a theorem changes the locality or composition hypotheses of mediator tests, if a controlled microscopic calculation fixes a previously flexible observable relation, or if evidence after 2026-08-20 changes any row’s best result, adverse check, or unresolved boundary.

Exercise

In the two-observable model above, keep $q=+1$ but let both measurements have standard deviation $\sigma=1$. Find $\widehat A$ and $\chi^2_{\min}$. What changed: the mechanism or the evidential strength?

Solution

The best-fit amplitude is still

$$\widehat A=\frac{1+(-1)}{2}=0.$$

The residuals are $+1$ and $-1$, so

$$\chi^2_{\min}=1^2+(-1)^2=2.$$

The predicted shared-sign relation and its best fit did not change. Only measurement precision changed: $\chi^2=2$ for two measurements and one fitted parameter is not a compelling rejection. Calling the mechanism falsified would therefore be premature. More precise data or an independent channel is needed, while changing $q$ would be a theory change rather than an uncertainty update.

References

• Addazi, Andrea, et al. (2022), “Quantum Gravity Phenomenology at the Dawn of the Multi-Messenger Era—A Review,” Progress in Particle and Nuclear Physics 125, 103948, arXiv:2111.05659.
• Aziz, Joseph, and Richard Howl (2025), “Classical Theories of Gravity Produce Entanglement,” Nature 646, 813–817, doi:10.1038/s41586-025-09595-7.
• Donoghue, John F. (2023), “Quantum General Relativity and Effective Field Theory,” arXiv:2211.09902v2.
• Fermi LAT and GBM Collaborations (2009), “A Limit on the Variation of the Speed of Light Arising from Quantum Gravity Effects,” Nature 462, 331–334, doi:10.1038/nature08574.
• LIGO Scientific Collaboration, Virgo Collaboration, and KAGRA Collaboration (2026), “GWTC-5.0: Tests of General Relativity,” preprint, arXiv:2607.19293.
• Ludescher, Stefan L., Leon D. Loveridge, Thomas D. Galley, and Markus P. Müller (2026), “Gravity-Mediated Entanglement via Infinite-Dimensional Systems,” Journal of Physics A: Mathematical and Theoretical 59, 215302, doi:10.1088/1751-8121/ae65e7.
• Raffaelli, Antonio, Mario Ballardini, and Nicola Barbieri (2026), “Signals from the Early Universe: A Comprehensive Search for Primordial Features in Planck CMB Datasets,” preprint, arXiv:2605.30236.