The calibrated posterior
Eight parameters, fit by Bayesian MCMC against the clean-decade geodetic mass balance, 22 stakes, and 9 trusted branch snowlines. The corner plot is the joint posterior — every pairwise trade-off and every marginal, drawn from 500 thinned draws.
One card per parameter: the prior (grey, ±1σ around μ) above the posterior (ink, 5–95% around the median) on its own bounds, then what the parameter is, what the posterior says, and the literature behind its prior (linked to the Literature station). A posterior narrower than or shifted off the prior means the observations were informative; an overlapping band means the fit leaned on the prior — an honest thing to see.
The other half of calibration: with the posterior in hand, re-run the model and compare it back to the observations it was fit to. Modeled vs observed for each constraint group — points on the dashed 1:1 line fit perfectly; systematic offset is bias. The honest results (a mild geodetic residual, the winter-stake offset) are surfaced, not smoothed.
Provenance & caveats
The posterior is 500 thinned draws from a 32-walker × 5,000-step ensemble MCMC, fit to the 2010–2020 geodetic mass balance, 22 stake readings, and 9 trusted branch snowlines (clean-decade likelihood, coupled geometry). The two thickness tracks give statistically identical calibration posteriors — the bed barely affects mass balance over the fit window; it expresses itself later, in projection longevity. σ_extra is an error-model hyperparameter (m w.e.), not a throwaway nuisance — its posterior is itself a finding about stake–geodetic tension (see the panel below).