koine

The synthesised fourth opinion for Bayesian triangulation.

Description

koine synthesises one coherent, independent predictive opinion for a Bayesian model from a curated battery of known and underutilised inference methods -- or abstains when no diversity gain exists. It is designed to serve as a fourth, synthesised witness alongside an independent panel of inference engines.

The load-bearing idea: averaging reduces diversity and manufactures false confidence. koine therefore synthesises on an axis orthogonal to the engine axis -- predictive-space leave-one-out stacking (PSIS-gated) over model-structure and method diversity, never averaging the engines. Its disagreement carries a genuinely new signal ("sensitive to model structure / inference family"), so it stays a real fourth witness rather than a weighted echo of correlated ones.

koine is programmatic infrastructure: it exposes no user-facing modelling verbs. A caller hands it an engine-neutral request and receives a synthesised predictive opinion or a structured abstention.

Related packages

• loo computes approximate leave-one-out cross-validation for Bayesian models with Pareto smoothed importance sampling, and gives stacking weights for averaging their predictive distributions.
• BAS does Bayesian variable selection and model averaging for linear and generalised linear models.

koine uses loo's leave-one-out estimates and stacks in the same way, but over models it fits itself: a set of inference methods, a heavy-tailed variant and a range of shrinkage priors. The result is one more opinion to set beside your own fitted model, not a replacement for it, and koine returns an abstention when the set adds nothing new.

Installation

koine is distributed through the GitHub repository max578/koine:

# install.packages("pak")
# pak::pak("max578/koine")
install.packages("koine", repos = c('https://max578.r-universe.dev', 'https://cloud.r-project.org'))

or from a local clone:

pak::local_install(".")

Usage

library(koine)

# an engine-neutral request: formula, data, family, seed
d <- data.frame(y = rnorm(60), x = rnorm(60))
pass <- koine_pass(y ~ x, d, gaussian(), seed = 1L)

# synthesise the fourth opinion from the battery (or abstain)
res <- synthesise(pass)
res

synthesise() returns either a koine_result -- carrying the stacked predictive opinion, the PSIS gate report, the diagnostic layer (weight concentration, the over-dispersed linear-pool envelope, and the prior-conditional evidence axis), and an independence report -- or a first-class koine_abstention. Stating less beats a wrong fourth opinion.

The battery spans three diversity axes: inference-family witnesses (DRAM, NIMBLE-Gibbs, Polya-Gamma Gibbs, DREAM_ZS), a heavy-tail structure variant (robust-t), and a shrinkage spectrum (ridge / lasso / horseshoe).

koine also handles random intercept structure -- a single random intercept (1 | g) and the breeder-standard multi-environment-trial (MET) structure of two crossed or nested intercepts (1 | variety) + (1 | trial) -- on an axis orthogonal to the sealed engines: NIMBLE-Gibbs samples the random effects while the marginal lens integrates them out (exactly for the Gaussian LMM and the Gaussian MET via a dense Cholesky of the marginal covariance, by adaptive Gauss-Hermite quadrature for a single-intercept GLMM). The Gaussian, binomial, and Poisson single-intercept GLMMs are enrolled; a non-Gaussian MET (a dense crossed-GLMM marginal with no exact quadrature) abstains rather than approximate. The MET lift lets koine corroborate a production MET model directly -- on the variety-plus-trial structure that model fits -- rather than on a single-random-effect reduction that discards the trial structure. A spectrum of further lenses gives the random intercepts heavy (Student-t) tails across a grid of degrees of freedom, so where the Gaussian-RE lenses corroborate they earn stacking weight when group effects are outlying -- koine's RE opinion is sensitive to random-effect structure, not a weighted echo of one assumption. All are stacked on the same per-observation integrated predictive, which keeps the leave-one-out gate reliable; richer random-effects structure (a random slope, more than two groupings) is refused (feature_unsupported) rather than fitted as the wrong model.

Validation

koine-authored lenses must clear a correctness gate before they may corroborate: simulation-based calibration and an exact grid-based reference-posterior check (inst/validation/). An unvalidated lens abstains rather than vouch for itself.

Dependencies

• Imports (always required): digest, loo, posterior, RNetCDF, stats, utils.
• Suggests (optional back-ends, gated at runtime): BayesianTools, BayesLogit, bayesreg, bridgesampling, nimble, ncdf4 -- plus knitr / rmarkdown / testthat / spelling for the developer toolchain.

Contributing

See CONTRIBUTING.md and the Code of Conduct. In short: branch off main, keep R CMD check --as-cran clean, add a NEWS.md bullet for user-facing changes, and reference the issue a pull request closes with Fixes #<n>.

Licence

MIT. See LICENSE / LICENSE.md.

Citation

See CITATION.cff. Please also cite the methods koine drives -- each lens's source paper is recorded in its documentation and provenance.