# Validation materials This directory contains reproducible validation code, party-blocked sensitivity analyses, posterior predictive checks, subgroup balance diagnostics, and figure-generation scripts for the party-position panel. All analyses preserve the production release (`v0`, production run `run_2026-06-12_09-34-03`). Existing posterior draws and completed sensitivity fits are reused wherever possible. The party-blocked expert validation tests prediction for parties whose expert evidence is entirely excluded from estimation. ## Inputs - `data/releases/party_2d_election_year_panel_v0.csv.gz` - `data/releases/party_2d_annual_model_output_v0.csv.gz` - production posterior run `run_2026-06-12_09-34-03` - completed no-V-Party sensitivity output `party_positions_2026-06-05_14-57-17.csv` - model-ready input files in `data/` Large model runs and their chains remain outside Git. This package contains the scripts, manifests, diagnostics, grouped summaries, plotting data, and figures needed to inspect and reproduce the published validation results. Two large row-level intermediate tables used during post-estimation are intentionally excluded; the retained grouped summaries and scripts document all reported calculations. ## Computational notes The validation fit retains the production warmup length and four-chain design but uses 1,000 retained iterations per chain (4,000 draws total), rather than 2,000 for the production release. The scripts record immutable train/test identifiers and input hashes before fitting and use existing posterior draws for descriptive and posterior-predictive analyses where applicable. ## Reproduction Entry Points - `plot_scale_examples.R`: historically anchored party paths and scale-reference cases generated from the released election-year panel. - `run_postprocessing.R`: country coverage, scale examples, V-Party sensitivity and partially pooled source-support diagnostics. - `validate_uncertainty.jl`: production-chain predictive coverage, calibration and source/item/country/decade breakdowns. - `prepare_blocked_validation.jl`: deterministic party-blocked train/test construction. - `run_blocked_validation.sh`: guarded low-priority fit with disk, duplicate-process, toolchain and logging checks. - `monitor_long_run.sh`: non-invasive status, child-process, disk and recent-log monitoring. - `finalize_blocked_validation.sh`: guarded post-estimation, held-out summaries and compact diagnostic-manifest collection after the fit succeeds. - `summarize_blocked_validation.jl`: held-out prediction records and grouped performance summaries after the fit. - `plot_workflow.R`: data-processing workflow schematic.