V1 — on complete trees, is each model's likelihood the likelihood of its simulator?
| Status | complete 2026-10-31 (2560 of 2560 trees); registered rule met in 29 of 64 cells; pooled over each cell's trees, 138 of 150 coefficients within [0.9, 1.1] |
| Series | the validation series (replaces E44) |
| Package build | emphasis 182055b (1.0.0.9000), R 4.5.3 |
| Driver | sim/V1.R, runner sim/V-run.R, library sim/V-lib.R |
| Data | results/V1/jobs/<cell>-k<kk>.csv, results/V1/build.txt |
Question
A complete tree — every lineage kept, the extinct ones included — hides nothing, so maximising its complete-data likelihood tests whether that likelihood is the one of the process the simulator runs, with no Monte Carlo involved. E44 asked this with a 60–160-tip window on the trees, which biased the ed cells; V1 asks it again on the current build, with no window, at two sizes and two turnovers.
Design
| models | cr, dd, d, nd, ed, edc, ned, nedc, k, nk, exponential link |
| cells | every reachable E43 cell at 100 and 200 tips: 64 cells |
| trees | 40 a cell, complete, \(\rho = 1\), crown age 15, kept when both crown lineages survive; no size window |
| fit | complete-data log-likelihood maximised by Subplex from the generating values and a perturbed start, restarted until a restart gains less than \(10^{-6}\) |
| jobs | 2560, seconds each |
Pre-set rule
Per cell, the median ratio of every speciation coefficient to its generating value lies in [0.9, 1.1], its bootstrap 95% interval contains 1, and a lineage-level coefficient has the right sign on at least 95% of trees. A cell that fails at both sizes is a likelihood that is not its simulator's; one that fails at 100 tips and passes at 200 is a size effect.
Results
The registered rule holds in 29 of 64 cells (sim/V1-analysis.R, results/V1/coefficients.csv).
crpasses at both sizes and turnovers;edcin 7 of 8 cells;kat \(-0.35\) in all;nkat turnover 0.10.- The cells that fail split two ways. Single-covariate models miss by 5–10% on the intercept or the slope (
d,ed), or on the sign share ofkat \(+0.35\) (0.78–0.93). The \(N\) models at turnover 0.40 (dd,nd,ned,nedc) read \(\beta_N\) at 1.1–1.4 of its value and the intercept 5–10% high. - The per-tree rule reads a 100- or 200-tip maximum, which carries small-sample error, on trees selected for survival and for staying under the lineage cap. A pooled estimate per cell (one parameter vector for the cell's 40 trees,
sim/V1-pooled.R, a diagnostic beside the rule, not a change to it) separates the two: small-sample error averages out of it, selection does not.
Pooled estimate (diagnostic, results/V1/pooled.csv). 138 of the 150 coefficients lie within [0.9, 1.1] of their generating values, and almost no clade was dropped at the lineage cap (a share of 0 in the \(N\)-model cells at turnover 0.40, 0.02 elsewhere), so the cap does not explain the failures and most per-tree failures are small-sample error. Outside [0.9, 1.1]: \(\beta_N\) of dd at turnover 0.40 (1.17, 1.15), of nd (1.21, 0.86), ned (0.85) and nedc (1.14, 1.15); \(\beta_D\) of d and nd (1.10–1.16) in three cells; \(\beta_{ED}\) of ned (0.85, 1.16). These are what is left once 40 trees are pooled, on trees kept only when both crown lineages survive.