Hypothesis

Readers pre-activate the sound of upcoming predictable words, so the N400 to the article before a noun would scale with how predictable the noun is, as DeLong et al. (2005) reported.

What failed

In a preregistered direct replication across 9 laboratories (N = 334), the noun effect replicated but the article effect did not, in either the preregistered or the single-trial analyses.

Broken assumption

Readers routinely pre-activate the phonological form of predictable upcoming words.

Why

Exploratory Bayesian analyses suggest the article effect may be non-zero but far smaller than originally reported, too small to see without very large samples. The authors conclude the results do not support routine pre-activation of word form.

From the published paper: Nieuwland et al. eLife, 2018. ↗CC BY

Hypothesis

Adjusted indirect comparisons, increasingly used to compare treatments never tested against each other, give reliable confidence intervals and hypothesis tests.

What failed

In simulations, indirect confidence intervals suffered under-coverage and indirect hypothesis tests had low power whenever between-study heterogeneity was moderate to large; the risk of overestimation was large when one of the two direct comparisons rested on a single trial.

Broken assumption

Indirect comparisons support confidence intervals and significance tests as dependable as direct ones.

Why

The authors trace the fragility to between-study heterogeneity and to unbalanced comparisons, where one side rests on little evidence.

From the published paper: Mills et al. PLoS ONE, 2011. ↗CC BY

Hypothesis

Reminding people of a precaution they have already taken makes them judge the related risk as less likely (the protection effect), alongside the tempting-fate effect.

What failed

Three direct replication attempts of Tykocinski's protection-effect experiment did not reproduce the effect, while two replications of the Risen and Gilovich tempting-fate experiment did reproduce theirs.

Broken assumption

Recalling a precaution taken reliably lowers people's perceived risk of the related bad outcome.

Why

The abstract does not offer a reason for the non-replication.

From the published paper: Wolferen et al. Judgment and Decision Making, 2013. ↗CC BY

Hypothesis

Raising the energy density and the volume density of rations should act together on appetite, and the positive link between energy density and intake should still hold at moderate-to-high densities.

What failed

Twenty healthy men completed four two-day crossover treatments; no energy-density-by-volume-density interactions appeared, and raising volume density produced no compensatory cut in self-selected intake, while total intake was 276 kcal lower (p = 0.01) on the low-volume-density treatments.

Broken assumption

Appetite compensates for the volume of food eaten the way it compensates for its energy density.

Why

The authors note that only a small isovolumetric portion of the diet was manipulated, and read the volume-density null as an opportunity rather than a problem.

From the published paper: Hatch‐McChesney et al. Appetite, 2024. ↗CC BY

Hypothesis

The five-peptide cancer vaccine S-588410, given after curative resection of HLA-A*24:02-positive, node-positive esophageal squamous cell carcinoma, would prolong relapse-free survival.

What failed

Among 276 randomised patients, median relapse-free survival was 84.3 weeks with the vaccine vs 84.1 with placebo (P = 0.8156), and overall survival did not differ (P = 0.6533). The primary endpoint was not met.

Broken assumption

Inducing cytotoxic T-lymphocyte responses against cancer-testis antigens delays relapse after curative resection.

Why

The vaccine did engage its target — cytotoxic T-lymphocyte induction in 132 of 134 patients within 12 weeks — but that immune response did not translate into longer relapse-free survival. The authors note longer survival among subgroups with high CTL intensity, upper thoracic tumours or grade 3 injection-site reactions.

From the published paper: Makino et al. Esophagus, 2024. ↗CC BY

Hypothesis

Anomalous current-phase relations and nonreciprocal supercurrent in topological-insulator Josephson junctions signal Majorana bound states.

What failed

Reconstructing the current-phase relation as a function of magnetic field, the anomalous peak-dip structure and the nonreciprocal supercurrent both appeared when one flux quantum threaded the junction, but no evidence of Majorana bound states was found despite the extensive prior literature.

Broken assumption

Anomalous current-phase features in a topological-insulator junction are a Majorana signature.

Why

The authors show the features follow from a nonuniform supercurrent distribution across the junction, which also produces the tunable Josephson diode effect.

From the published paper: Kudriashov et al. Science Advances, 2025. ↗CC BY

Hypothesis

Decay would make velvet worms appear progressively more primitive, as experiments on chordates showed, so badly decayed lobopodian fossils would be misplaced towards the stem of the panarthropod tree.

What failed

Experimental decay of onychophorans produced no significant stemward slippage, so this particular bias does not explain interpretations of non-mineralised lobopodian fossils; decay was far from unbiased in other ways, changing body proportions significantly, affecting anatomical features at very different rates and causing correlated loss of topologically associated characters.

Broken assumption

The decay-driven stemward slippage found in chordates generalises to other animal groups.

Why

The authors reason that slippage bites where phylogenetically informative characters sit in decay-prone tissues, which is not the case for lobopodians, and that this makes claims of preserved neural tissue and complex musculature in Cambrian lobopodians hard to reconcile with their results.

From the published paper: Murdock et al. BMC Evolutionary Biology, 2014. ↗CC BY

Hypothesis

A mouse model of breast cancer driven by polyoma middle T antigen would reproduce the systemic changes in monocytes seen in human breast cancer, making it a usable model for studying them.

What failed

Tumour-bearing mice did show more monocytes in bone marrow, blood and spleen, from faster proliferation, and transcriptional changes in classical monocytes led by down-regulated type I interferon signalling and antigen presentation. But these transcriptome changes did not match those found in patients with breast cancer.

Broken assumption

The monocyte response to cancer in this mouse model mirrors the human systemic response.

Why

The abstract does not explain the mismatch; the authors conclude the mouse model may be inadequate for studying systemic effects of human cancer.

From the published paper: Robinson et al. Cancers, 2022. ↗CC BY

Hypothesis

A short-term drop in ion-induced nucleation, of the kind caused by a decrease in cosmic-ray ionisation, would measurably change atmospheric aerosol and cloud-condensation-nucleus properties — the proposed explanation for observed cosmic-ray/cloud correlations.

What failed

In a global aerosol microphysics model, a 10-day 15% reduction in the total nucleation rate (an upper limit on the ion-induced effect) lowered global mean surface cloud-condensation nuclei by only 0.15%, and against the model's own aerosol variability no statistically significant response could be detected in particles above 10 nm, CCN above 70 nm, or the Ångström exponent.

Broken assumption

Cosmic-ray-driven changes in ion-induced nucleation are large enough to alter cloud and aerosol properties.

Why

The modelled signal is far smaller than natural aerosol variability, so the authors conclude the observed cosmic-ray/cloud correlations cannot be explained by associated changes in the large-scale nucleation rate.

From the published paper: Dunne et al. Atmospheric chemistry and physics, 2012. ↗CC BY