Reliability & Interpretation · v5
Calibrated Probabilities Are Not Yet Good Decisions
Reliability diagrams, proper scores, and decision costs answer different questions; a useful probability estimate must survive all three.
Read the review →AUTURNAL / RESEARCH LIBRARY / 2026.10AI meets discovery.
AI meets discovery.
Follow the evidence.
170 Chinese and English research reviews connecting primary studies, mathematical models and inspectable evidence.
170guides526primary sources510illustrations and charts
THE RESEARCH COLLECTION
Explore the research
Reliability & Interpretation · v5
Generalization Under Shift Requires Identifiable Assumptions
Unlabeled target inputs can reveal some shifts, but no adaptation algorithm can recover information that its assumptions leave unidentified.
Read the review →
Reliability & Interpretation · v5
Preference Optimization Does Not Identify Human Utility
Pairwise feedback identifies a statistical preference model under assumptions; it does not by itself define a universal or uniquely recoverable utility.
Read the review →
Reliability & Interpretation · v5
What an Interpretable Feature Actually Establishes
Readable activations become scientific explanations only when their basis, causal role, and unresolved alternatives are made explicit.
Read the review →
Genetic engineering · v5
The Precision Ledger of Prime Editing: Evidence from Sequence Rewriting to Whole-Genome Outcomes
Successful rewriting, product purity, cellular state, and long-term stability must be measured separately; rebuild the credibility of prime editing using complete denominators.
Read the review →
Genetic engineering · v5
Programmable Large-Fragment Integration: Positioning, Integrity, and Genomic Costs of RNA-Guided Transposition
Integration rate is not a single endpoint; position, orientation, cargo integrity, and multiple copies must form joint evidence linked to the same molecule.
Read the review →
Biophysics · v5
The Real Limits of Reading Individual Proteins: Nanopore Rereading, Conformational Noise, and Proteoform Identification
Single-residue sensitivity, multi-pass classification, and de novo sequencing of arbitrary native proteins are three different scientific objectives.
Read the review →
Evolution and Human History · v5
How Ancient Genomes Reconstruct Migration: Ancestry Models, Temporal Sampling, and the Limits of Historical Interpretation
Moving from fragments in bones to population history requires navigating preservation bias, reference population selection, and multiple equivalent histories.
Read the review →
Genomics · v5
Whose Genome Defines the Coordinates: Pangenomes, Variant Discovery, and the Measurability of Reference Bias
Replacing a single path with a multihaplotype graph both increases visible sequence and changes the definitions of alignment, truth, and error.
Read the review →
Microbial ecology · v5
The Logic of Discovering Microbial Dark Matter: From Metagenomic Catalogues to Verifiable Functions
New sequences, candidate new species, and new functions are distinct achievements; assembly, binning, and annotation errors must be propagated along the evidence chain.
Read the review →
Earth systems · v5
The Depth Ledger of the Ocean Carbon Pump: Particle Transport, Fragmentation, and Actual Sequestration Time
Surface production, export from the euphotic zone, and long-term isolation must be measured separately to connect ocean processes to reliable assessments of carbon storage.
Read the review →
Remote sensing and ecology · v5
From Laser Waveforms to Forest Carbon: Sampling Design, Scaling, and Conservation of Uncertainty
Finer pixels do not mean more accurate carbon stocks; footprint measurements, plot calibration, area weights, and change detection each require their own evidence.
Read the review →No reviews match these filters.