Known Limitations

Area	Limitation
Standard deviation	Divides by n not n-1. For typical MCMC chains (n ≥ 500) the difference is < 0.1%, but it is technically a biased estimator. Affects skewness and kurtosis computation too.
HDI multimodal	The sliding-window algorithm returns the shortest contiguous interval. For strongly bimodal posteriors this returns the widest single mode. There is no multi-interval HDI.
No caching	sequenceStats(), variableStats(), and summary() recompute on every call. For large posteriors (> 50k draws × 100 variables), calling summary() in a tight loop is expensive.
densityPlot no spec	densityPlot() and getDensityPlotData() exist but there is no densityPlotSpec(). All other plots have a spec function.
Missing diagnostics	No BFMI (critical for HMC/NUTS), no Heidelberger-Welch, no Raftery-Lewis, no nested R-hat, no multivariate PSRF, no R*, no ESS/sec, no posterior correlation matrix, no DIC/LOO-CV.
Export formats	Only generic JSON export. No Turing CSV, Stan CSV, or ArviZ JSON re-export.
ArviZ recursion depth	The ArviZ JSON parser uses recursion for multi-dimensional parameter flattening. Extremely high-dimensional parameters (> 1000 dimensions) risk stack overflow.
NetCDF	Cannot read NetCDF files. netcdfjs only reads NetCDF v3 (classic format), but ArviZ writes NetCDF v4 (HDF5-based). They are not compatible without conversion.
Autocorrelation inconsistency	The ACF shown in autocorrelationPlot uses a simple lag formula; the autocorrelation used internally for ESS uses the FFT-based power spectrum. Both are correct but not identical for finite-length chains.