Status complete (runs E25c and E25d); historical: a method-development step; not a current result (replaced by E28, which refits the three families at 300 iterations under the build that reports the parsimonious choice)
Package build not recorded in the results; the restart followed the retry fix of the E23 amendment
Driver / analysis sim/E25-real.R (--box study for E25c, --box auto for E25d)
Data results/E25c/fits.csv, paths.csv (study box); results/E25d/fits.csv, paths.csv (auto_bounds()); results/E25/jobs/ (an earlier run, Cotingidae and Icteridae only)

Question

E5 fitted Corvidae, Cotingidae and Icteridae with dd, ed and ned on the linear link, and E17 bootstrapped the ned-over-dd margins. The biology paper ends on these three clades and needs the workflow a user runs, on the link every lineage-level experiment uses: the one-call ladder and the covariate path.

Design

setting value
trees the same three, tip edges extended to one present as in E5
\(\rho\) 0.968 (Corvidae), 1 (Cotingidae), 0.971 (Icteridae)
ladder estimate_rates(model = "all"), exponential link: cr → dd → {nd, ned, nk}, each started from the model it nests
box E25c: the box of sim/box.R from the observed tree (the rule of every simulation experiment); E25d: auto_bounds()
draws 300
iterations max_iter 100
clock 60 min per fit
ranking compare_models() (minimum AIC at this build)
path covariate_path() over \(N\), \(D\), ED, \(K\) for speciation and extinction, once on the observed tree (exact) and once on 3000 surrogate draws at the ladder's dd estimate, the 150 of largest weight entering (the E20 setting)

Recorded per family and model: AIC, log-likelihood, effective sample, gap, stop reason, proposal, iterations, coefficients; per family the winner, its margin, the ranking-at-risk flag, and both paths' entry orders and BIC and AIC sets.

Why two boxes: on Corvidae (crown 30.5 Ma) auto_bounds() for nd ran more than five hours without finishing, because its per-lineage feasibility simulations at the explosive corners of the box ran to the lineage cap, each at quadratic cost. The forward simulator also retried clades that overflowed max_lin as if extinct; the retry fix (E23 amendment) is why the run was restarted on 2026-10-01. E25c was run with the study box to get past it; E25d with auto_bounds() once the fix was in. The README's E25 row names --box study; the supplement's index points at results/E25/jobs/, which holds two families only.

Pre-set rule

None. "An illustration, read against the floors of the simulation chapters"; every coefficient is reported with the effective sample and the stop reason of the fit it comes from.

Results

Ladder, E25d (auto_bounds()), with E25c in parentheses where it differs:

family tips AIC cr dd nd ned nk winner margin ESS of winner (of 300)
Corvidae 120 767.4 757.3 753.0 739.4 727.4 nk 12.0 (11.1) 76 (84)
Cotingidae 65 412.0 405.5 409.2 409.5 407.5 dd 2.0 (2.2) 299
Icteridae 102 535.1 508.3 516.7 503.4 510.0 ned 4.9 (5.1) 109 (23)

Coefficients (E25d): Corvidae nk \(\beta_K = +0.148\) per split, ned \(\beta_{ED} = -0.157\) per ED unit; Icteridae ned \(\beta_{ED} = -0.236\); Cotingidae nk \(\beta_K = +0.077\). Stop reasons: Corvidae nk converged at 12 iterations (E25c: 85), ned at 33; Icteridae ned converged at 67 (E25c: on max_iter, ESS 23); every nd fit on max_iter, with ESS 71 (Corvidae), 101 (Cotingidae) and 6.5 (Icteridae, sampler-limited). The ranking-at-risk flag is false on all three. Ladder wall-clock: 4043, 1595 and 2621 s (Corvidae, Cotingidae, Icteridae) with auto_bounds(); 1787, 747 and 1994 s with the study box.

Path (identical sets in E25c and E25d):

family speciation, observed tree: entry order BIC set BIC set on 3000 draws (ESS of the draw set) extinction BIC set: observed / draws
Corvidae D > ED > N > K D+ED D+ED (2959) empty / empty
Cotingidae N > ED > D > K N N (2992) empty / empty
Icteridae N > ED > D > K N+ED N+ED (2867) empty / N

The observed tree holds no extinction events, so the observed-tree extinction path is empty by construction.

E25 trees
Figure 1. The three trees fitted here (the same as E5's), tips shaded by ED. Corvidae's ED runs from 3.4 to 30.5 and its K from 0 to 21 splits; this is the tree on which the ladder names K and the path keeps D and ED.

Conclusion

No rule. On the exponential-link ladder the three families return three different answers: nk on Corvidae by 12.0 AIC over ned, ned on Icteridae by 4.9, and dd on Cotingidae by 2.0 over nk. Corvidae, the clade E5 read as "a clean case for ED", is taken by the ancestral-splits model, while its path keeps D and ED, not K: the ladder and the path name different covariates on the same tree, the pattern E26 then finds across families and E21 predicts (K enters the path through ED with the sign reversed). The two boxes change no winner; they change Icteridae's ned fit from capped at ESS 23 to converged at ESS 109. Corvidae's dd rung ended on max_iter and Icteridae's nd is sampler-limited. The numbers the papers use for these families are E28's, refitted at 300 iterations under the build that reports the parsimonious choice.