The evidence layer for scientific AI: claims with supporting evidence, scope, and limitations.
Updated September 07, 2026 Β· Also available via MCP for Claude, Cursor, and compatible AI tools.
Physics & CosmologyThe wavefunction clearly exists as a mathematical state-description that successfully generates experimental probabilities, but whether it is a physically real object, a representation of knowledge, or an emergent description remains unsettled.
Identical predictions can support competing interpretations β the divide is mathematical structure versus physical ontology.
Accelerated cosmic expansion is securely measured and ΞCDM is still strongly favored, but the underlying physical nature of dark energy is unknown and some independent measurements remain in tension with the model.
Model flexibility can fit acceleration without demonstrating new physics.
An inversion layer is warm air already sitting above colder, denser air, so lifting it would displace denser air downward β a stable arrangement that suppresses convection.
Buoyancy depends on density at the same pressure, not on "denser above."
Adding a single entangling gate between EEG sub-reservoirs is a plausible but untested hypothesis, and the current quantumβclassical performance gap (NMSE 0.42 vs 0.03) is too large for one gate to close.
The relevant benchmark is held-out gain after quadratic readout augmentation, not entanglement alone.
Microbiology & Antibiotic ResistanceA single engineered conjugative plasmid carrying IncM, IncL, and IncC entry-exclusion genes blocks over 99.9% of in-vitro acquisition of AMR plasmids and prevents gut colonization in co-housed mice.
Limitations: small mouse groups (n=3), single strain, no statistical testing reported, no human-relevant validation yet.
Reducing plasmid maintenance and transfer is biologically plausible but not clinically established; within-host adaptation, biofilms, and alternative plasmids could preserve resistance after apparent curing.
BGPT odds of true: 62% β supported for feasibility in lab systems, not for human safety or effectiveness.
A contained biofilm design with IncX1 and IncU plasmids spanning a 1,500-fold conjugation-frequency range can simultaneously track transfer, loss, replacement, resistance, and fitness.
The frequency gap mirrors two spatial-structure regimes predicted by biofilm models.
VHHβhistatin fusion proteins showed pathogen-selective killing and encouraging mouse-model activity, but clinical efficacy, resistance durability, and broad strain coverage are not yet established.
Preclinical stage; developed by the company commercializing AbTids.
Featured Paper ReviewsA quality-factor screen offers a low-cost, falsifiable triage for frequency-multiplexed biological carriers, but cross-carrier rankings depend heavily on estimated coherence times and unmeasured inputs.
Usefulness 80% Β· Novelty 50% β a necessary-condition screen, not proof of biological function.
Patient-level controls show frozen pathology embeddings contain molecular signal, but much of the interpretable variation is compositional, and the tested geometric machinery added no measurable predictive value.
Demonstrated in TCGA breast cancer (n=285); no independent paired imageβRNA validation.
Under leakage-controlled gene-held-out splits, the reported codon-model advantage on synonymous variants collapses from +2.3β14.3 points to ~+1.6β2.2 points or reverses.
A memorization canary (AUC 0.739) beats every neural model β leakage amplification, not a codon-specific failure.
LTP stimuli recruit IRSp53 to drive Arc capsid assembly and vesicle release, and transferred Arc mRNA reduces surface AMPA receptors in neighboring neurons.
Novelty 90%, but findings are in cultured rodent neurons; in-brain relevance is inferred.
Roughly 61% of the Auanema rhodense germline genome is discarded at the 12-cell stage via precise 29-bp palindromic break motifs β the most extreme case found in a free-living nematode.
The elimination nuclease and regulatory logic remain unidentified; motif causality untested in this species.
Carollia bats phase-lock echolocation calls to low-noise phases of rhythmic noise in lab and wild settings, with causal cortical evidence that shifting call phase buys information at no extra energetic cost.
Novelty 90%; single species and a single 4-Hz masker limit breadth.
A bottom-up accounting model estimates building 1 kg of muscle costs about 41 kcal/day β roughly 10Γ less than standard 300β500 kcal/day surplus prescriptions.
Zero primary data; the key parameter derives from infants and animals, and the uncertainty range is likely understated.
A pretrained latent-diffusion model converts sparse spatial assays into atlas-constrained representations, improving reconstruction and clustering in tested Xenium/SPATCH data.
Most valuable when inferred genes are distinguished from directly measured evidence; generality is conditional on atlas similarity.
A dual-neurotransmitter model proposes dopamine drives reward seeking while amygdala noradrenergic signaling drives loss aversion, implying the two are separable systems rather than features of one utility curve.
Perspective piece relying on correlational imaging, small pharmacological studies, and a 2-patient lesion study.
AI, Memory & MethodsA durable record should carry type, source, time, confidence, and permitted use, surviving only when independent evidence supports it and policy permits retention.
Persistence is not success β durable storage without exposed provenance is weak evidence.
The otto-SR workflow improves screening and extraction efficiency, but evidence supports a validated human-supervised tool β not autonomous systematic reviews.
Whether rare or visually encoded errors survive into final syntheses without changing conclusions is undemonstrated.
The strongest contribution is converting governance properties into testable software invariants, though the evidence supports computational feasibility rather than demonstrated biosecurity effectiveness.
Biological hazard detection remains proxy-based and partly circular.
Signed dendritic gating makes the clustering rule directly inspectable, showing an interpretable decision rule can stay useful even when evidence spans nearly all latent dimensions.
Known-K assumption, one 10-dimensional latent space, and code not yet public limit immediate practical value.
Health & NeuroscienceUrolithin A has a biologically plausible but unproven route to supporting healthy aging β evidence is mainly mechanistic studies, cell models, and mice, with no randomized human trial showing increased survival.
The open question is whether sustained human exposure produces durable functional improvements that reduce age-related deaths.
A convergent intrathecal antibody response to an EBV BRRF2 motif that cross-reacts with human CNS proteins was found in a minority of MS patients, with a highly specific serum biomarker candidate.
Applies to only ~8% of MS patients; causality (initiation vs amplification) is unresolved.
Adult-local AAV-FMR1 re-expression restored activity-driven inhibition and prevented audiogenic seizures, showing a circuit defect long assumed to need early intervention is correctable at adulthood.
Male-only cohorts, small n (4β8), a commercial conflict of interest, and untested Npas4 necessity temper the claim.
Multimodal functional genomics connects selected MS loci to oligodendroglial regulatory activity and cellular phenotypes, but does not prove oligodendroglia are primary drivers of the disease.
Constrained by engineered perturbations, one iPSC clone, and in-vitro context.
Clusters probably reflect different anatomical targets β distal axon, sensory-neuron cell body, or pain-processing network β rather than one disease, and loss and gain of function can coexist in one patient.
Hypothesis-level synthesis; cluster stability and molecular profiles need prospective testing.
No supplied scientific data exist on non-forceful bottle offering and infant cortisol or stress; context-specific refusal (home vs daycare) as a learned behavior remains plausible but biomarker-unvalidated.
BGPT odds: 45% that relevant biomarker data exist as framed.
Ecology & EvolutionMotility is an energyβfitness trade-off, not an unconditional advantage β construction and operation consume substantial cellular energy while benefits depend on environment.
The evolutionarily relevant trait may be a regulated distribution of motility states, not average speed; heterogeneity itself remains a testable hypothesis.
Localized resource use can suppress a neighboring invader while sparing other resources and resident taxa, though the supplied studies do not directly measure niche preservation or diversity effects.
Best competitor β strongest competitor; the objective is invader suppression per unit of resident habitat displaced.
Cell types diversify through specialization, regulatory control, and physical organization into tissues and circuits, but no single universal evolutionary mechanism is established by the evidence.
Best framed as constrained recombination of conserved regulatory ingredients.
Opponent-relative temporal segregation was observed between the two mouse species, but avoidance phase-locked to civil dusk rather than clock hours was not analyzed.
Photoperiod-coordinate avoidance remains an untested interpretation.
General ScienceCYP450 oxidation converts AFB1 into a reactive epoxide that forms DNA adducts, which depurinate or persist as AFB1βFAPy lesions that disrupt replication and produce mutations.
Carcinogenicity depends on lesion identity, persistence, and repair β not total adduct count alone.
Across 164 shallow waters from sea level to 4,886 m, methane bubbling rises with falling pressure (over 4-fold above 3,000 m), consistent with a theoretical degas+trigger model.
Correlational, geographically biased toward the Tibetan Plateau; generality to other mountain systems is asserted, not demonstrated.
In the Drosophila follicular epithelium, TGF-Ξ²/BMP signaling increases E-cadherin abundance at tricellular vertices and prevents vertex opening even when Myosin II activity is reduced.
Supported with a refinement: the mechanism separates junctional adhesion from junctional force.
A pre-registered hypothesis proposes microgravity overdrives von Economo neuron circuits, with compensatory myelination sustaining emotion-recognition speed for β€6-month missions but failing beyond ~340 days.
Scientific quality 40% β the central cognitive claim rests on N=1 digitized data with cross-dataset confounds.
A design-hypothesis visual for targeted siRNA nanoparticle delivery is scientifically defensible only when endosomal release is framed as supported while tumor accumulation, receptor identity, and apoptosis are marked as formulation- and model-specific.
Do not depict EPR accumulation or therapeutic apoptosis as demonstrated before powered animal data exist.
Across six preregistered experiments (N=1,554), a computational model treating intimacy as a graded cost of vulnerable actions tracks human belief updates at rβ0.99, near the noise ceiling.
Only one relationship dimension tested; cultural generality explicitly unverified.
Know what changed, what holds up, and what remains uncertain. Every Friday. No ads.
Explore by Goal
Ask a Follow-Up
Trending