---
title: "Technical note: validation record"
format: html
vignette: >
  %\VignetteIndexEntry{Technical note: validation}
  %\VignetteEngine{quarto::html}
  %\VignetteEncoding{UTF-8}
---

```{r}
#| label: setup
#| include: false
knitr::opts_chunk$set(collapse = TRUE, comment = "#>", eval = FALSE)
```

This note indexes the validation evidence kept in the repository under `docs/reports/`, and
states what each item supports and does not support.

| Topic | Report | Supports |
|-------|--------|----------|
| Calibration of MBCO and model bootstrap | `CALIBRATION-2026-10-14.md` | Nominal MBCO coverage off the double null; conservative at it; model bootstrap by design |
| MBCO near the double null | `MBCO-NULL-T5-FEASIBILITY-2026-10-14.md` | A calibrated cutoff was not found; the interval stays conservative |
| MBCO optimizer | `MBCO-OPTIMIZER-REVIEW-2026-10-14.md` | Exact constraint elimination; optimizers agree to about 1e-11 |
| Coverage claims in the manuscript | `PMED-COVERAGE-CLAIMS-CHECK-2026-10-14.md` | Percentile intervals carry Monte Carlo inflation |
| Exact Gaussian reference | `I2-EXACT-REFERENCE-2026-10-15.md`, `AUDIT-exact-gaussian-hopper-2026-10-20.md` | Exact-Gaussian run passed its original six-cell rule |
| Gaussian candidate gate | `I6-GATE-2026-10-19.md` | Candidate d failed the frozen coverage-equivalence and width gates |
| Wasserstein repairs | `VALIDATE-wasserstein-*-2026-10-20.md` | Boundary correction, selection, joint contrast and observed inference checks |
| Ecosystem compatibility | `tasks/plan-mediationverse-compat.md` | Suite passes on medfit 0.3.2 (CRAN) and 0.12.0 (r-universe), 832 tests each; 0.4.1 fixed a missing-value guard that medfit 0.10+ bypassed |
| Boundary point risk | `REPORT-boundary-point-risk-2026-10-21.md` | Negative result; see below |

## Boundary behavior in one paragraph

For the binary-outcome Wasserstein proportion mediated, the plug-in bias at an exactly zero
path is of order $n^{-1/4}$. The frozen bias gate (|bias|/SD below .10) cannot be met by any
estimator with values in [0, 1] in the zero-direct and weak-direct cells at n = 500. A
two-point floor on maximum RMSE is about .09 (direct) and .07 (indirect) at n = 500. Five
reference estimators failed the author's RMSE tolerances. The supported uncertainty statement
is the Fieller interval, which passed coverage in every archived binary cell.

## What this record does not show

It does not show that no estimator could meet the tolerances; the floor lies below them. It
does not clear the author's scientific hold on the boundary diagnostics, and it does not
authorize a CRAN or manuscript submission.

## Reproducing

Scripts live in `scripts/` (`check-boundary-*.R`, `check-gate-*.R`, `pilot-reference-estimators.R`).
Set `VECLIB_MAXIMUM_THREADS=1` and `OPENBLAS_NUM_THREADS=1` before running forked simulations, or
the workers stall on BLAS threads.