The Evidence Layer for Scientific AI

BF2170 methylation of bezafibrate raised gut-wall permeability ~20-fold in cell assays and plasma drug exposure in mice colonized with defined bacteria; demonstrated so far only in mono-colonization models.

One engineered plasmid carrying three entry-exclusion genes blocked >99.9% of resistance-plasmid acquisition in vitro and gut colonization in co-housed mice; groups were small (n=3) and generality to complex human microbiomes untested.

Laboratory systems support feasibility of curbing plasmid spread, but human safety, efficacy, and compensatory resistance routes — biofilms, alternative plasmids, chromosomal resistance — remain untested.

A proposed contained design tracks transfer, loss, mobile-element replacement, fitness, and resistance across IncX1 vs IncU clinical plasmids; the experiment is designed, not yet performed.

Fiber loss, fewer butyrate producers, and barrier failure are each individually supported, but the full chain has never been tested with ultra-processed foods, resting on observational diet-pattern data.

Across ~50,000 participants in ABCD, HCP, and UK Biobank, median brain–behavior effects were tiny (median |r|≈0.01), so typical small neuroimaging samples produce inflated, irreproducible estimates.

Resampling and Spearman-Brown analyses indicate measurement reliability — not only participant count — governs replication needs; a prospective validation has not been run.

In propofol-anesthetized male rats, local neural activity largely reversed between anesthesia loss and recovery while global network trajectories did not; one anesthetic and one species limit generality.

RNAi knockdown of the receptor subunit chrnal-E increased sleep in Cassiopea — the first functional genetic sleep evidence outside bilaterians; the preprint has small samples and unresolved species-identity issues.

Perceived isolation predicted cognitive decline and Alzheimer's risk in older adults even though autopsy data showed no neuropathology link, pointing toward reserve effects rather than direct degeneration.

The protease gene HTRA1 is downregulated in male and upregulated in female nucleus accumbens under stress (humans and mice), with causal mouse manipulations; preprint without independent replication.

Plasticity-inducing stimuli recruit IRSp53 to assemble Arc capsids and release vesicles that suppress excitatory receptors (AMPARs) in recipient neurons — shown in cultured rodent neurons, not intact brain.

Across six preregistered vignette experiments (N=1,554), a dyadic inverse-planning model matched human judgments at r=0.987–0.991; samples were US-only and one relationship dimension was tested.

In an exploratory fMRI analysis (n=18 per group), psilocybin increased while retreat training decreased topological distance between brain-state graphs, hinting at counterbalancing rather than synergy.

Mapping model training onto drift, immigration, recombination, and isolation yields closed-form predictions with two honestly reported failed tests; key mechanisms are simulation-only and LLM validation is thin.

Governance properties become testable software invariants with preregistration and deterministic reruns, but biological hazard detection stays proxy-based — computational feasibility, not demonstrated biosecurity effectiveness.

A 4,400-prompt audit found uneven bias resilience and universal jailbreak susceptibility; a single judge and skipped adversarial testing of already-weak categories make measured vulnerability a lower bound.

Screening and extraction at review scale performed well with blinded adjudication, supporting human-supervised acceleration rather than autonomous reviews; evidence covers twelve reviews and independent replication is pending.

Acute caffeine (200–300 mg) transiently raises blood pressure ~8/6 mm Hg, while habitual 2–3-cup intake tracks modestly lower mortality in cohorts; causality of the cohort benefit remains unresolved.

In UK Biobank (n=190,275), genetically predicted omega-3 raised heart-attack risk while measured levels appeared protective, with rs964184 genotype modifying the benefit; ancestry-limited and causally unresolved.

Animal and cell studies consistently show inflammatory and oxidative-stress signaling, but measured human exposure is poorly quantified and a causal human-harm link is nearly empty.

A Navigation Guide systematic review rates prenatal pyrethroid exposure as sufficient evidence of harm at moderate quality and postnatal exposure as limited, with thyroid disruption the best-supported mechanism.

In 21,851 US fertility-clinic samples, higher modeled PM2.5 associated with more sperm DNA fragmentation, but cross-sectional design, unmeasured smoking/BMI, and a larger age effect (+14.4 points at ≥50) cap inference.

A NEJM clinical review consolidates that 6.5–12.9% of mothers are affected and antenatal depression is the strongest predictor; the headline prevention statistic rests on one small (n=78) observational study.

The review links UV photochemistry to carcinogenesis but stresses repair, apoptosis, and clonal competition as deciding which damaged cells persist; it used no systematic search or risk-of-bias methods.

Across 164 waters up to 4,886 m, methane ebullition rose with falling pressure, matching a physical degassing model, but the evidence is correlational and dominated by the Tibetan Plateau.

In one 20-year throughfall-exclusion plot pair, long-drought trees held transpiration steadier through the 2023 El Niño while carrying ~26% smaller crowns — resilience apparently purchased with structural sacrifice.

Auanema rhodense eliminates ~61% of germline DNA at the 12-cell stage at precise 29-bp palindromic motifs — the extreme of known programmed DNA elimination, with the cutting nuclease still unidentified.

snATAC-seq of 3,842 haploid mouse sperm nuclei yielded >25,000 meiotic crossovers and confirmed elevated recombination in Fancm knockouts, with lower per-cell recovery than prior methods.

Enformer applied to 34 hominin and ape genomes predicted 37,592 lineage-specific regulatory DNA changes; 7 of 9 tested elements were experimentally active and one causal variant resolved, though most predictions remain unvalidated.

Equilibrium modeling explains why cytosine and thymine are absent from meteorites while A, G, and U persist, though the model overproduces by 3–4 orders of magnitude and omits kinetics.

Solid-state nanopores sustained a liquid pathway at −5 to −20 °C, stretching DNA translocation ~400-fold; the liquid state is inferred electrically, not yet visualized.

A 4-amino-acid degron plus RNA-allosteric folding control produced reporter fluorescence only when target RNAs were present — enabling real-time RNA imaging in whole plants — but only 2 of 12 switch designs worked.

Carollia bats phase-locked calls to quiet dips of 4-Hz noise in lab and field, and frontal-auditory-cortex inactivation disrupted timing; modest sample sizes and a post hoc model temper confidence.

Bottom-up accounting estimates building 1 kg of muscle costs ~41 kcal/day — roughly tenfold below common surplus prescriptions; the estimate is literature-derived with no primary data.

Theory suggests regulated mixtures of motile and immotile cells cut population energy costs while preserving exploration, but direct measurements of within-population heterogeneity are lacking.

Neural cellular automata evolved to build a body gradually lose regenerative goals, and resetting misexpressed cells rejuvenates virtual tissue — proof-of-concept in silico, with no empirical validation.

No simulation grid yet quantifies habitable-zone planet survival across binary separations and eccentricities; a proposed periastron-based N-body grid would replace 1974-era qualitative claims with testable corrections.

A hydrothermal-network paper claims period-five oscillations mark autopoiesis and that mortality follows mathematically from self-production, but only two networks were tested without statistical tests — rated weak.

Colonize mice with matched 20-strain communities differing only in BF2170 status, then measure full plasma drug exposure (AUC).

Dose probiotic-treated mice with escalating IncM plasmid inocula (10³–10⁹ CFU) to find the threshold where exclusion fails.

Test replication of preregistered brain–behavior associations at n = 500 to 4,000 in ABCD follow-up data, verifying the predicted replication curve.

12-month mouse exposure to real indoor-dust particle mixtures at measured human-relevant concentrations, testing whether lab-derived toxicity thresholds hold at real doses.

Test the omega-3 treatment effect within rs964184 genotype strata using stored STRENGTH and VITAL samples — a cheap, decisive test of the interaction.

Hourly time-series sampling of the hydrothermal reaction network to detect or rule out period-five oscillations — the first empirical test of the claim.