What masque protects, and what it does not | Two axes of control: roles and actions | Depth controls | The conditional clone: preserving the treatment effect | What the copula carries: monotone, not non-monotone, association | The leakage audit | Multi-table sets | Geographic coordinates | Operational guidance | References

What masque is for | The one-call path | The masking plan: roles and actions | Editing the plan as code | Local and collaborate modes | Tidy, dates, and depth | More than one table | Multi-environment structure | Where to go next

Why the recipe matters | What a recipe holds | Printing is redacted | The round-trip | Fail-closed translation | Saving the recipe | Multi-table bundles

Why an operator calculus | A worked example | Pushforward through a linear sensor | Bayesian update on a noisy linear observation | Kalman parity check | Aggregation through a coarsening matrix | Missing-data conditioning | Filtering over time -- the Kalman filter as a special case | Bounding the filter -- mixture reduction | A bounded filter in one call -- gmm_filter() | Composition rules | When to not use the operator calculus | Comparison to a Gaussian-process latent | References

Why a from_kde() constructor | Recovering a known mixture | Quality of the importance-sampled fit | Bandwidth sensitivity | Visualising the compression | Scope and limits (v0.2) | When to prefer fit_proxymix(regime = "sample") instead | References

Banana | Donut | Three-mixture | Bounded support | Summary | Reference

Regression -- from lm() to a conditioned mixture | Kernel regression -- from Nadaraya--Watson to a conditioned KDE | Clustering -- from kmeans() to the mixture components | Principal components -- from prcomp() to the fitted covariance | Penalised regression -- ridge as a covariance ridge | A few more, in one line each | Summary | When to reach for the specialised tool | References

A target you cannot sample from | Fit a Gaussian-mixture proxy | Overlay the proxy on the target | Closed-form mixture operations | Where to next | Reference

Regime (i): closed-form moment matching | Regime (ii): classical EM on samples | Regime (iii): importance-sampled KLD-EM | Side-by-side overlay | A sanity check: (i) and (iii) agree at N = 1 | Reference

A real unnormalised posterior | Fit the proxy | The evidence, with an independent check | Closed-form reads off the proxy | Error bars on the proxy itself | When to use this, and when not to | References

The closed-form spine | Renyi-2 entropy | Shannon entropy, bracketed | Cauchy--Schwarz divergence | Mutual information and predictive entropy | When to use which | Deterministic annealing and phase transitions | Maximum-entropy targets | Selecting a component count with ICL | Dependency structure of a mixture | References

Imputation is conditioning | A two-cluster example | Impute, then look at the shape | Pooling an estimand | Scope

An objective is a density you can evaluate but not sample | A bimodal objective in one dimension | Four minima at once: Himmelblau | Notes on use | References

Imputation is conditioning on a gate | A value-dependent mechanism biases the ordinary fit | The slope is a sensitivity parameter | Censoring is the exact, identifiable case | Scope and references

The end-of-sample question | The statistic | A worked example | Robustness to non-Gaussian noise | Scope | Reference

Overview | Lineage and scope | Cross-tool benchmark: PESTO vs PEST vs PEST++ | Accuracy, calibration, and cost | Interpretation | Performance characteristics | A worked example: a well-posed inverse problem | The problem | Building a PEST control file | Running PESTO | Comparison with the reference smoother | Limits | Reproducibility | References

Why this study | The calibration target | The synthetic observations | The calibration | Convergence | Parameter recovery | Identifiability -- what the data can constrain | Out-of-sample validation -- and the under-dispersion caveat | The ensemble manifest | Part 2 -- calibration to real observed data | Estimates, with an independent cross-check | Reproducibility | What this study shows | References

Why a callback driver? | A synthetic recovery example | The forward-model contract | Parallel, fault-tolerant ensembles | Guard: calibrate to field-realistic uncertainty, not the mean's standard error | Driving APSIM through apsim_callback() | Failure handling | Concurrency | Multi-fidelity calibration | ODE / compartmental forward-model templates | Sequential (filter-mode) assimilation | When to prefer pesto_ies() instead | Reproducibility

The problem: finite-ensemble under-dispersion | A linear-Gaussian problem with a known posterior | The bare smoother collapses | Inflation re-expands the spread | Localisation suppresses spurious correlations | References | Session information

Introduction | Key features | Installation | Quick start: creating a model scenario | Core computational kernel | Performance benchmarking | Computing phi (the objective function) | Visualisation | Adaptive SVD backends | Adaptive-backend ensemble solution | Surrogate-accelerated IES | Adaptive ensemble sizing | Next steps | References

Overview | Regime of applicability | The test problem | Step 1: generate the prior ensemble | Step 2: train the GP surrogate | Step 3: evaluate surrogate predictions | Step 4: surrogate-assisted IES update | Comparison: standard vs surrogate | When the surrogate pays off | Effect of nonlinearity | Random Fourier features for large ensembles | Scaling comparison | When does the surrogate help? | References

Why a manifest? | Constructing a manifest from an IES run | Writing, reading, and verifying | Inspection CSVs (optional) | Tamper-detection | Consuming a manifest | Reproducibility

Calibration: ksd_test() | Concordance: concordance_test() | The validation pattern

What this vignette demonstrates | Stage 1 --- load the toy cohort | Stage 2 --- build a small illustrative feature matrix | Stage 3 --- stratified train / test split | Stage 4 --- fit the three models | Stage 5 --- predict on the held-out fold | Stage 6 --- evaluation (the manuscript headline) | Calibration headline | Decision-curve net benefit at clinically relevant thresholds | Discrimination (supporting) | How this maps to the manuscript

Status | Data source | Unit of analysis | Inclusion criteria | Exclusion criteria | Outcome | Prediction window | Variable contract | Splits and resampling | Invariants (executable contract) | Reproducibility | Out of scope | Reporting | Pedagogy (D5 default --- balanced) | Changelog

The Problem | How It Works | Example: Linear Causal Effect | Example: No Causal Effect (Confounding Only) | Example: Non-Linear Causal Effect | Diagnostic: Null Distribution | Using the Formula Interface | When to Use bd-HSIC | References

The cross-package contract | Four backends, one problem | bd-HSIC under each backend | When to reach for proxymix | References

Beyond Mean Effects | Key Concepts | Example: Mean Shift (Detectable by All Methods) | Example: Variance Effect Only (Invisible to Mean-Based Tests) | DR-DETT: Effect on the Treated | Comparing the Tests | Using the Formula Interface | When to Use Which Test | References

The ag-scenario question | One calibration, two scenarios | Run the DR-DATE scenario test | Null case — same outputs | What gets validated | Output sub-selection | When parameter posteriors genuinely differ | Where the cross-package plumbing lives | Reproducibility

What is kernR? | Kernel Basics | Computing Kernel Matrices | Two-Sample Testing with MMD | Independence Testing with HSIC | Visualising Results | Next Steps

Why Hierarchy Matters | Example: Agriculture Trial | Decomposition: Within vs Between | Weight Methods | Example: No Treatment Effect | When to Use Hierarchical Tests

Why a hierarchical version of bd-HSIC? | A multi-site stub | Naive vs hierarchical permutation | Stratified per-cluster contributions | Type-I behaviour under a true null with cluster effects | Choosing the scheme | Notes on practice | Reference

What HSIC-Sensitivity catches that Sobol misses | A sensitivity scan on a stubbed APSIM-like simulator | The scan | Visual | Side-by-side with sensitivity::sobol | Tabular form | Total-order indices | A pure-interaction example | Side-by-side plot | A near-additive example for contrast | Pair-bootstrap CI for total-order indices | Notes on total-order indices | Notes on practice | References

Three flavours of kernel downscaling | (1) kernel_downscale(): coarse climate → paddock yield | When to reach for kernel_downscale() | (2) dist_regression(): distribution → scalar | Bag-level vector targets | Outer kernel choice | Variable bag sizes | When to reach for dist_regression() | (3) aggregate_downscale(): aggregate-likelihood inversion | When to reach for aggregate_downscale() | Comparison | References

The question | The stub ensemble | Case A: well-calibrated posterior | Case B: miscalibrated posterior (variance too narrow) | Case C: miscalibrated posterior (mean-shifted) | Comparing surprise across scenarios | Visual diagnostic | Notes on practice | Cross-package handoff | References

Why screen before calibrating? | A stubbed APSIM archetype | Latin-hypercube design | Simulate and screen | Visual diagnostic | Handing off to PESTO | Tabular form | Notes on practice | References

The scaling wall | A tiny correctness check | Scaling demonstration | When to use which | Notes on practice | References