BGPT pairs scientific claims with their supporting evidence, scope, and limitations — so findings can be inspected, not just trusted. Available through MCP for Claude, Cursor, and compatible AI tools.
Updated November 13, 2026
Scores for novelty, rigor, quality, and reproducibility — with the weaknesses stated plainly.
The one underappreciated claim each paper supports — and where the evidence stops.
Falsifiable follow-up designs that split competing hypotheses at modest cost.
First mechanistic demonstration that EBV-primed CD4+ T cells cross-react with an MS-associated autoantigen, though small single-cell samples and undisclosed raw data temper confidence.
Novelty 90% · Quality 80% · Reproducibility 60%
Landmark single-cell estimate of ~13.7 somatic L1 insertions per neuron, but a ~4-fold spread between artifact-inclusive and corrected estimates means the true burden remains a methodology question.
Novelty 90% · Quality 90% · Reproducibility 80%
The field-defining microbiota–gut–brain axis review; its causal claims remain largely rodent-based, and early-life HPA programming implies adult-only probiotic trials may miss a critical window.
Novelty 90% · Generality 90% · Reproducibility 70%
Heat-killed controls cleanly split live-colonization from bacterial-product explanations, but human causality stays correlational and statin confounding persists in the CAD cohort.
Novelty 90% · Quality 80% · Generality 70%
First microscopy of Hodarchaeales cells, but the eukaryote-like interpretation rests on fluorescence signal geometry alone (n=40, one site) and awaits electron-microscopy arbitration.
Novelty 80% · Quality 80% · Generality 60%
A three-arm protocol proposing psilocybin under anesthesia to test whether the psychedelic experience itself drives antidepressant effects — conceptually rigorous, but n=9 with no data yet.
Novelty 80% · Quality 60% · Usefulness 60%
Strong single-cell evidence for EBV-infected driver B cells in MS, with the caveat that only ~5 in 10,000 cells were EBV-positive and counts were extremely small.
Novelty 80% · Quality 70% · Usefulness 80%
Near-zero total effect coexisting with significant indirect paths suggests social media harm operates through distrust, sleep timing, and body image — implying intervention leverage points exist despite small population-level effects.
Novelty 80% · Quality 70% · Reproducibility 80%
A plausible but untested sociotechnical mechanism built from case reports — the true attributable fraction of chatbot-associated psychosis is genuinely unknown because confounders co-occur with intensive use.
Thirty years of consistent rodent data with strong mechanistic plausibility, yet fewer than 30 humans ever studied and zero RCTs — real-world risk is unknown precisely where exposure is highest.
Clinical-concentration AMPK activation in tumors is biologically implausible, so the observational epidemiology and the proposed mechanism could not both be right — randomized trials sided against the epidemiology.
The non-white null result anticipated a core genomics problem: a variant's functional annotation can be ancestry-dependent, making some "replication failures" broken biomarker validity rather than absent biology.
The review's least-supported claim — that meal timing, not fasting duration, drives metabolic benefit — was a prescient hypothesis that a decade of subsequent trials only partially confirmed.
A consistent but small-to-moderate association with worse adolescent mental health (~80% of 50 studies), though causality and dose-response remain uncertain because "use" is measured mostly by unvalidated self-report.
Immune "boosting" is largely a repletion story — foods help most by correcting suboptimal nutrient status, and effects largely vanish in people who are already replete.
Metabolic levers are dose-dependent and asymmetric: ~150 min/week buys disease prevention, but meaningful weight loss and organ-level repair demand 250–300 min/week plus resistance training.
5-HT2A is a signaling hub, not one pathway — JAK2 sits upstream of both STAT3 and ERK1/2, so psychedelic receptor activation can reconfigure transcriptional programs rather than only acutely altering neurotransmission.
One high-quality MMR–autism case-control study exists (OR 0.86, 95% CI 0.68–1.09, no association) — a meta-analysis cannot exceed its inputs, and inflating the evidence base would itself be a scientific error.
High-impact PNAS study engineering bee gut symbionts to express dsRNA against Nosema ceranae — highly innovative colony-level control, limited by small sample sizes in key subgroups.
Quality 80% · Rigor 70% · Novelty 90%
Noncytotoxic PBD–ciprofloxacin hybrids active against M. tuberculosis in vitro, with the Gyrase A mechanism still inferred from modeling rather than proven biochemically.
Quality 80% · Rigor 80% · Novelty 80%
Groundbreaking ALS work linking dipeptide repeat proteins to inhibited GEM body formation via dCas13 and iPSC models, with the caveat that cell culture cannot fully replicate in vivo neurology.
Quality 80% · Rigor 80% · Novelty 90%
Exemplary use of TMLE causal inference in a multicenter bladder-cancer protocol, though results remain unpublished and the non-randomized design constrains generalization.
Quality 80% · Rigor 90% · Novelty 80%
Abrupt stiffness interfaces in 3D GelMA hydrogels orient and differentiate fibroblasts through mechanical cues alone, limited to a single cell model in vitro.
Quality 70% · Rigor 70% · Novelty 70%
High-rigor demonstration that influenza suppresses mitochondrial Complex III assembly to dampen macrophage hyperinflammation, pending confirmation in human primary macrophages.
The lead author's declared cofounder/shareholder status in Genegoggle shifts the burden from biological plausibility to demonstrating commercial and scientific independence; findings are robust within their limited scope but demand skepticism about broad generalization.
The commentary and the original study share the same epistemic flaw — both extrapolate across the in vitro/in vivo boundary without measuring hepatocyte nanoparticle concentration and clearance kinetics.
Every retrieved paper concerned textile or anatomical fibers rather than dietary fiber — a cautionary example of how polysemous terms can silently derail evidence retrieval without exact-concept indexing.
Colonize germ-free mice at three postnatal timepoints to test whether the behavioral critical window matches the HPA programming window — directly addressing the timing constraint adult probiotic trials ignore.
Objective device logs rather than self-report, isolating timing from duration effects on validated depression and anxiety scales.
Pre-registered, blinded, ≥200 untreated MS patients with raw data deposit — resolving the natalizumab confounding and small-sample limitations of the original finding.
Directly resolves the lipid-versus-mRNA distribution dispute underlying the COVID-19 vaccine reverse-transcription debate.
A double-blind diabetic foot ulcer trial bridging three decades of rodent wound-healing signal to the complete absence of human randomized evidence.
Electron microscopy at FISH-confirmed coordinates splits the compartmentalization-versus-division-stage hypothesis, arbitrating the contested eukaryote-like morphology claim.
Connect BGPT's evidence layer directly to Claude, Cursor, and other compatible AI tools through the Model Context Protocol — claims arrive with evidence, scope, and limitations attached.
Know what changed, what holds up, and what remains uncertain. Every Friday. No ads.
Explore by Goal
Ask a Follow-Up
Trending