Summarize adaptive refits — summarize

Build a thin per-refit diagnostics summary from the adaptive round log. This is a pure view over round_log and does not recompute posterior quantities or stop metrics.

Usage

summarize_refits(state, last_n = NULL, include_optional = TRUE)

Arguments

state: An adaptive_state or list containing adaptive logs.
last_n: Optional positive integer; return only the last n rows.
include_optional: Logical; include optional diagnostic columns.

Value

A tibble with one row per refit (canonical round_log schema).

Details

The round log is the canonical stop-audit trail. This summary is a direct view over round_log with no recomputation.

Key fields include:

identity: refit_id, round_id_at_refit, step_id_at_refit
run scale: total_pairs_done, new_pairs_since_last_refit, n_unique_pairs_seen
candidate health: proposed_pairs_mode, starve_rate_since_last_refit, fallback_rate_since_last_refit, fallback_used_mode, starvation_reason_mode
identifiability/quota adaptation: global_identified, global_identified_reliability_min, global_identified_rank_corr_min, long_quota_raw, long_quota_effective, long_quota_removed, realloc_to_mid, realloc_to_local
diagnostics/stopping: diagnostics_pass, divergences, max_rhat, min_ess_bulk, ess_bulk_required, reliability_EAP, rho_theta, delta_sd_theta, rho_rank, stop_decision, stop_reason
report-only uncertainty metrics: ci95_theta_width_*, near_tie_adj_*, cov_trace_theta, top20_boundary_entropy_*, nn_diff_sd_*

Examples

# These summaries work on either an adaptive_state or a plain list of logs.
logs <- list(
  round_log = tibble::tibble(
    refit_id = 1:2,
    round_id_at_refit = c(1L, 2L),
    step_id_at_refit = c(10L, 20L),
    new_pairs_since_last_refit = c(50L, 50L),
    total_pairs_done = c(50L, 100L),
    divergences = c(0L, 0L),
    max_rhat = c(1.01, 1.00),
    min_ess_bulk = c(800, 900),
    stop_decision = c(NA, TRUE),
    stop_reason = c(NA_character_, "btl_converged")
  )
)

# Full per-refit view:
summarize_refits(logs)
#> # A tibble: 2 × 10
#>   refit_id round_id_at_refit step_id_at_refit new_pairs_since_last_refit
#>      <int>             <int>            <int>                      <int>
#> 1        1                 1               10                         50
#> 2        2                 2               20                         50
#> # ℹ 6 more variables: total_pairs_done <int>, divergences <int>,
#> #   max_rhat <dbl>, min_ess_bulk <dbl>, stop_decision <lgl>, stop_reason <chr>

# Only the most recent refit row:
summarize_refits(logs, last_n = 1)
#> # A tibble: 1 × 10
#>   refit_id round_id_at_refit step_id_at_refit new_pairs_since_last_refit
#>      <int>             <int>            <int>                      <int>
#> 1        2                 2               20                         50
#> # ℹ 6 more variables: total_pairs_done <int>, divergences <int>,
#> #   max_rhat <dbl>, min_ess_bulk <dbl>, stop_decision <lgl>, stop_reason <chr>

# Drop optional diagnostics if you want a compact core summary:
summarize_refits(logs, include_optional = FALSE)
#> # A tibble: 2 × 10
#>   refit_id round_id_at_refit step_id_at_refit total_pairs_done
#>      <int>             <int>            <int>            <int>
#> 1        1                 1               10               50
#> 2        2                 2               20              100
#> # ℹ 6 more variables: new_pairs_since_last_refit <int>, divergences <int>,
#> #   max_rhat <dbl>, min_ess_bulk <dbl>, stop_decision <lgl>, stop_reason <chr>