Decoupling molecular clock dating from phylogenetic tree reconstruction
Dating a population requires only two algebraic operations: anchoring the ancestral root and weighting shared ancestry across sublineages. ChronAeon executes both in continuous distance space:
Matches Bayesian rates and root dates across 42 empirical BEAST benchmarks ($\log_{10} R = 0.94$) in seconds while outpacing Delphy by two orders of magnitude.
Processes an uncurated century-long BV-BRC dump across 16 HA subtypes in minutes, quarantining 453 temporal outliers and resolving 72 host/clade clocks.
Streams 6,317,042 SARS-CoV-2 Spike genomes (2020–2026) in flat memory ($M = 1{,}887$ landmarks), locking onto emerging VOC clocks weeks before 50% dominance.
Profile rooting, foundation covariance, AutoClock, and online coresets
Weights sequences exponentially toward earlier sampling dates to form a soft nucleotide profile $\mathbf{Q} \in \Delta^{L \times 4}$, contracting saturated distances back into the linear regime.
Extracts cross-taxa attention covariance $\mathbf{K}$ from HyphAeon in one forward pass and computes exact asymmetric confidence sets via Fieller's theorem ($g < 1$).
Recursively bisects the normalized graph Laplacian along its Fiedler vector to separate co-circulating rate regimes while flagging high-leverage outliers via Sherman-Morrison LOOCV.
Locks 256 ancestral anchors + 1,792 online Approximate Linear Dependence (ALD) landmarks to stream millions of genomes in $\mathcal{O}(N \cdot M)$ time and flat memory.
Dating pandemic HIV-1 Group M ($N = 466$) without building a tree
HyphAeon cross-attention covariance ($\lambda^* = 0.943$) with a restricted cubic spline dates the HIV-1 Group M crown to 1901.8 CE [1891.2, 1912.4], matching Worobey et al.'s 1908 biopsy-calibrated estimate.
42 published BEAST datasets, LTEE ground truth, and Delphy speedups
AutoClock raises modern concordance across 42 BEAST studies from $R = 0.71$ to $R = 0.86$ ($\mathrm{MAE} = 9.59\text{ yr}$, 25/42 CI overlap) and $\log_{10} R = 0.94$ across five millennia. On the 50,000-generation E. coli LTEE, ChronAeon achieves -0.8% rate error and dates the 1988.15 founding clone to 1987.51.
Streaming 6,317,042 SARS-CoV-2 genomes (2020–2026) with 18–119d lead time
AutoClock resolves multi-wave Simpson's paradox into a conserved $1.43 \times 10^{-3}\text{ subs/site/yr}$ within-lineage rate and avoids the 25–93d detection delays of $N=3{,}000$ subsampling.
Run ChronAeon in your browser or reproduce all 53 manuscript claims
Drop a FASTA alignment and date table into primaeon.org/time for instant client-side dating, LOOCV triage, and AutoClock partitioning.
Includes all 42 BEAST benchmarks, 148,516 simulations, BV-BRC streaming scripts, and exact sequence ID manifests (10.5281/zenodo.23560893).
Simon Frost's chronaeon.timetree extension builds full binary time-trees and coalescent skylines ($\hat{N}_e(t)$) in milliseconds without MCMC.