BGPT: Introspect

Explore by Goal

Introspect

Audited evidence-weighting biases — what each stance gets right, and where it fails

Updated October 29, 2026 · Evidence layer for scientific AI · Each claim is one introspection with its measured or observed limitation.

Geochronology: radiometric-only weighting

Past the ¹⁴C limit (~50 ka) radiometric anchors become unavailable, so a bias against non-radiometric chronometers is self-limiting — it is only safely applied where independent calibration exists, the very condition the bias discourages investigating.

Immunology: intervention-first stance

Mechanistic-first reasoning is epistemically sound but risks tissue-blind reductionism — blood NK cytotoxicity genes exceed tissue-resident NK by >1000-fold, so rectal-explant depletion results depend on a phenotype the depletion experiment itself does not resolve without parallel transcriptomics.

AI research popularity demonstrably diverges from rigor — 82% of AI-bias papers lack bias definitions, while underrepresented AI-healthcare topics achieve equal-or-higher citations with a fraction of the publications.

Gut microbiome: peer-review filter

Peer-reviewed ≠ clinically proven — the most-cited gut-microbiome evidence was rat and in vitro data (one dairy-funded), so the quality filter optimized for mechanistic rigor while substituting methodological purity for translational validity.

Metallo-β-lactamases: structural-class bias

The B1/B2/B3 framework is a scaffold, not a predictor — ~30-fold potency variation exists within B1 enzymes (VIM-2 vs NDM-1 for ML302M), so structural class alone poorly predicts inhibitor susceptibility.

Adolescent STI programs: modeling over trials

Caregiver receptivity (68% Ghanaian support, higher in rural areas) may invert assumed barriers, suggesting the bottleneck is service logistics rather than community permission — yet the decisive cluster RCT does not exist, so numeric certainty remains low.

FXS neurodegeneration: microglia evidence gap

The honest answer is "unknown": neuron- and fibroblast-level sex differences cannot be transitively mapped onto microglia, the most sex-hormone-responsive and species-divergent brain cell type, making cross-cell-type inference especially hazardous.

ApoE-/- mice: organ-specific endpoint choice

Selecting vascular endpoints as "the" organ-specific data made the absence of any hepatic endpoint invisible rather than flagged — a structural bias where choosing which organ to care about determines whether the evidence gap is even perceived.

TFH therapeutics: citation-count epistemics

Citation-based evidence weighting structurally erases exactly the unpublished clinical pattern recognition that historically generates most therapeutic hypotheses — the fix is explicit tiered labeling, not pretending zero-weight assignment is neutral.

BTK-Low cells: corpus-grounded coverage gap

The DDR→inflammation field has three mechanistically independent routes (GATA4/NF-κB, cGAS/STING, IL-1α), but BTK's dosage-dependent role in any of them is unmapped — a genuine open question, not a limitation of the introspection.

Temporal tuning: mechanistic bias discipline

The tau-shift experiment's real contribution is falsifiability symmetry — it can equally vindicate a readout-only account, so the mechanistic bias must be disciplined by pre-registered criteria or it degenerates into confirmation bias.

Visualization bias: framing beyond prose

The most dangerous residual bias operates at the visualization level — a chart showing only unadjusted r=0.26 while the same paper's adjusted r=0.05 sits in a table encodes framing no limitations paragraph can undo.

EBV persistence: recency vs serology conflation

Recency bias and serology-skepticism are two distinct biases conflated in one heuristic — recent spatial data localize EBV, but only older bulk perturbation studies establish causal necessity, so truth-seeking requires pairing both.

Neuron subtyping: narrative-plausibility inflation

The mechanistic preference cuts both ways — it underweights speculative papers but inflates trust in mechanistically-plausible yet causally-unvalidated cell-line studies, because narrative coherence mimics evidence strength.

GLUE guidance graphs: weak prior, not unique data

Empirical peak–gene correlations show median r ≈ 0.08, so the guidance graph functions as a weak prior whose value depends almost entirely on the GNN's denoising capacity — making "GLUE uniqueness" a claim about the model, not the data.

POC test uptake: quantified incompleteness

Quantifying "social acceptance" as one logistic covariate can manufacture the illusion of completeness while dominant constraints (trust, stigma, staffing) remain unmeasured — the honest fix is pre-registered structured observation.

Saliva composition: grounding over coverage

Strict grounding is an anti-hallucination design choice — it explicitly sacrifices completeness when the evidence pool lacks a full constituent inventory, prioritizing falsifiability over fluency.

mNGS review: rigor-bias as anti-confabulation

Identical test labels span 12–92% performance depending on prespecified thresholds, reference standards, and sponsor proximity — precisely the details a generic checklist cannot adjudicate without the protocol itself.

Bronchiectasis CSF biomarkers: publication-survival bias

Confidence scales with citation counts and cohort size, but these correlate with publication-survival — a finding's prominence is partly a function of being positive, not being true, while negative results remain structurally invisible.

LDOC1/EGFR: extract fidelity both ways

Strict extract fidelity prevents fabrication of missing endpoints (migration, filopodia, metastasis) but also previously let an evolved response cite entirely unrelated papers — fidelity checks must verify record provenance, not just numeric presence.

Stress-free sleep deprivation: circularity trap

The deepest validation trap is circularity — using physiological arousal markers to prove stress-freeness while those same markers respond to non-stressful novelty, so protocols must demonstrate a null physiological signature, not just absence of behavioral stress.

Aqueous drug synthesis: invisible economic null cells

A mechanistic evidence base is structurally blind to its own economic null cells — absence of scale data is invisible unless explicitly coded as "unknown", so thermodynamic elegance can silently dominate rankings that would flip under real purification and GMP constraints.

Gene-transfer monitoring: sensitivity-first priors

The dPCR/VCN > IS-NGS > imaging ranking is a structural prior from assay physics, not an evidence-weighted conclusion — the honest fix is marking which protocol steps rest on training memory versus supplied citations.

How to read these introspections

Each card names a specific evidence-weighting stance and the measured limitation that qualifies it. Most introspections converge on one pattern: the bias itself is often the blindspot — a preference for one evidence type (mechanistic, radiometric, peer-reviewed, recent) quietly determines which gaps get flagged and which stay invisible.

✓ Stances that held up: strict grounding, rigor-bias, and falsifiability-symmetric designs reduced confabulation.

⚠ Stances that failed: endpoint selection, visualization framing, and narrative plausibility created invisible gaps.

→ Consistent fix: pre-registered criteria, tiered evidence labeling, and explicit "unknown" cells for missing data.

Stay current without chasing every paper.

Know what changed, what holds up, and what remains uncertain. Every Friday. No ads.

Evidence Tracing

Trace claims back to their original source, including experimental design, limitations, and data scope.Know what the science actually supports before you trust the answer.

Explore by Goal

Stay current without chasing every paper.

My BGPT

Trending