The evidence layer for scientific AI — claims, scope, and limitations
Evidence supports the wavefunction as a mathematically defined state description that predicts experimental probabilities; whether it is a physically real object remains unsettled.
Established meteorology: buoyancy depends on temperature at the same pressure, so "denser air above" alone does not make a warm layer rise.
Strong multi-layer evidence (gene knockout, LC-MS/MS, gnotobiotic mice) shows Bacteroidota enzyme BF2170 methylates bezafibrate, raising drug exposure — but validated so far in mono-colonization models only.
Entry-exclusion genes on one engineered plasmid blocked >99.9% of AMR plasmid acquisition in vitro and in mice, but small mouse groups and single-strain limits cap confidence.
Repurposing natural entry exclusion as an anti-resistance firewall is plausible and partly self-limiting, yet coverage of the full clinical plasmidome and escape evolution remain untested.
Mechanistically feasible in simplified systems (BGPT odds ~62%), but human safety and infection-control effectiveness are not supported by current evidence.
A 1,500-fold conjugation-frequency gap between IncX1 and IncU plasmids mirrors two-regime spatial models, suggesting plasmid size or spatial access as the bottleneck — testable, not yet proven.
Solid evidence that motility costs substantial cellular energy; whether selection acts on heterogeneous motility distributions rather than averages is a plausible but unproven hypothesis.
Environment-responsive motility portfolios are a defensible, testable idea, but supplied evidence does not directly measure within-population heterogeneity or its energetic benefits.
Localized resource use can suppress invaders while sparing residents, but "unused niche" preservation remains a testable inference rather than a measured outcome.
Across ~4,940 species, song complexity correlates negatively with brain size after controlling for migration — refuting the "songs signal intelligence" narrative at this scale, though correlational design limits causal claims.
Six preregistered experiments (N=1,554) show a computational model tracks human intimacy judgments at r≈0.99, but results are US-context vignettes and one relationship dimension only.
Enformer-based scans across human, archaic, and ape genomes found 37,592 candidate regulatory changes with R=0.81 luciferase validation, but most predictions remain functionally unconfirmed.
High-quality evidence (>25,000 crossovers, 3,842 genomes, Fancm validation) shows single-nucleus ATAC-seq can haplotype even protamine-packed sperm chromatin; per-cell distance recovery is lower than prior methods.
~61% of the germline genome is discarded at the 12-cell stage at precise palindromic motifs — well demonstrated, but the elimination nuclease and evolutionary cause remain open.
Adult AAV-FMR1 delivery suppressed audiogenic seizures in male Fmr1 mice — evidence for adult circuit reversibility — though cohorts are small, male-only, and Npas4 causality is untested.
Biologically plausible but unproven: evidence comes from cell and mouse studies; no randomized human trial shows increased survival or lifespan.
A transparent accounting model estimates ~41 kcal/day per kg of new muscle — an order of magnitude below common surplus advice — but the estimate has zero primary data and rests partly on infant/animal parameters.
A large cross-sectional study (N=21,851) links PM2.5 to sperm DNA damage, but ZIP-level exposure proxies and missing smoking/BMI data mean only cautious, non-causal inference is warranted; the age effect is larger.
Dual agonists outperform on HbA1c and weight per stronger trials, but the reviewed synthesis itself has methodological red flags (no PRISMA flow, no registration, conflicting search descriptions).
Durable 1918 H1N1 antibodies coexist with zero SARS-CoV-2 cross-reactivity in 33 Sicilian centenarians, ruling out antibody cross-protection — but the cohort is tiny and essentially unreplicable.
A coherent narrative review links UV photochemistry to DNA damage and immune modulation, but provides no systematic search protocol and compresses heterogeneous evidence into precise attributions.
Intrathecal antibodies to an EBV BRRF2 motif cross-reacting with human CNS proteins appear in ~8% of MS patients with 99.4% specificity — validated but causality (initiation vs amplification) is unresolved.
In propofol-anesthetized rats, local brain dynamics reverse between loss and recovery of consciousness while global network trajectories do not — demonstrated in one anesthetic, one species, no causal tests.
A testable review framework proposes loss aversion is noradrenergic while reward seeking is dopaminergic, but supporting pharmacological studies are small (n=12–48) and unpreregistered.
First causal demonstration (lab, field, model, inactivation) that bats phase-lock calls to quiet gaps, though sample sizes are modest and only one species and masker rate were tested.
A preregistered, fully reproducible hypothesis links von Economo neurons to emotion-recognition decline beyond ~340 days, but the central cognitive claim rests on one digitized observation.
Argued from convergent evidence: durable records should carry type, origin, confidence, and permitted use — persistence alone is storage, not evidence.
Leakage-controlled benchmarking (CodonBench) shows reported codon-model gains collapse under gene-held-out splits — a rigorous, reproducible audit with public code and data.
The best model depends on the metric — TxGemma-9B on ROC-AUC, LlaSMol-Mistral on top-1% enrichment — but evidence is retrospective ChEMBL data with only two fine-tuning replicates and no prospective validation.
otto-SR improves screening and extraction efficiency, supporting supervised acceleration — not autonomous reviews; generalizability and independent replication are incomplete.
A pretrained latent-diffusion model improves sparse spatial assay reconstruction (PCC 0.31 vs 0.24 baselines), with strongest evidence for clustering, not gene-specific inference; inferred genes should not be confused with measured ones.
A 1B-parameter model trained on 23M microenvironments forward-simulates tissue responses and yielded one validated combination, but only one designed perturbation was tested in one assay and code is unreleased.
An unusually transparent governance prototype converts safety properties into testable software invariants, yet biological hazard detection remains proxy-based and partly circular.
Visual narratives of targeted siRNA delivery should present tumor accumulation, receptor identity, and apoptosis as model-specific hypotheses, since only the endosomal-release step has direct siRNA evidence.
Evidence from sponges to mouse circuits supports cell-type diversification via regulatory recombination plus spatial wiring — but no single universal evolutionary mechanism is established.
Hydrothermal autocatalytic chemistry is real, but the universal "period-five" life theory rests on analogy across only two reaction networks and lacks statistical testing.
No existing source reports quantified survival fractions; periastron — not separation alone — likely governs stability, a testable claim requiring new N-body grids.
Falling atmospheric pressure predicts >4-fold higher methane ebullition above 3000 m (R²=0.38), but the pattern is dominated by the Tibetan Plateau and direct methanogenesis controls are absent.
Well-established pathway: CYP450 converts AFB1 to a reactive epoxide forming N7-guanine adducts that persist as FAPy lesions — but whether mutation spectra track persistent adducts over total adducts is untested.
A falsifiable quality-factor screen shows a carrier can separate labels yet fail on persistence, but conclusions depend on estimated coherence times and incomplete direct measurements.
Know what changed, what holds up, and what remains uncertain. Every Friday. No ads.
Explore by Goal
Ask a Follow-Up
Trending