UQ reorg · Page 4: Attribution of (effective) stress to forcing source, with uncertainty

[mdenolle]

Community need: not just the state of effective stress at depth, but attribution — which forcing (hydrological/poroelastic, thermoelastic, surface load, damage/co-seismic, tectonic residual) drives it, and how sure we are. Attribution runs on EFFECTIVE stress (Marine, 2026-06-26).

Framed as optimal fingerprinting + Bayesian variance decomposition (the $C_d$ this project builds is what makes it rigorous):

• Source decomposition: each forcing's depth-time effective-stress contribution $\Delta\sigma'_k(z,t)$ from Phase-3 design columns × kernel × bridge.
• Variance shares $A_k(z)=\mathrm{Var}_t[\Delta\sigma'_k]/\mathrm{Var}_t[\Delta\sigma']$; cross-covariances = attribution ambiguity (collinearity). Coupled regime → Sobol first-order + total-effect indices; the gap is an explicit interaction/coupling share.
• Credible intervals on $A_k$ by sampling the joint posterior (forcing coefficients + material priors + $C_d$).
• Detection per source via Bayes factor (ties to Phase-2 coupling model-selection).
• Attribution-uncertainty budget: data-limited ($C_d$) vs model-limited (priors/collinearity)?

Deliverable figure: depth–source attribution chart (stacked $A_k(z)$ with credible-interval whiskers + a coupling slice).

Settles the depth×time cube (#2 §9.1): keep forcing-resolved time-domain sensitivities, then invert each source over depth.

See quarto/REORG_PLAN.md §10.