The conclusions of the experiments, by the question they answer, each with the number that supports it and the page that establishes it. Terms are defined in the Glossary; the status of every experiment is on the experiments index. A statement resting on a result that a planned or running experiment will replace is marked (superseded: E<n> remeasures); until then it is the measurement on record.

In brief

1. Does the estimator reach the exact answer, and with which sampler?

Where an exact likelihood exists, MCEM on the surrogate proposal reaches its maximum. Under cr and dd at 20-100 tips, 27 of 30 cell × proposal rows meet that rule; the surrogate gives up a median 0.00-0.07 nats at 20-50 tips and 0.05-0.32 nats at 100, with rates within 0-7.5% (speciation) and 0-10.5% (extinction) of the exact estimate (E1, unconditioned likelihood, linear link). The reported log-likelihood is the exact one to 10\(^{-12}\) under cr (E1), and on E40's complete real trees to 0.000 (cr) and within 0.01 (dd); where \(\rho\) < 1 the reported value sits 0.01-0.17 nats above DDD's because the two codes model sampling differently (E42). From 25 to 800 tips under cr the MCEM estimate sits within a tenth of the exact MLE's own spread, offset upward by 1-6% of the speciation rate, with the cells from 100 tips clock-limited (S) (superseded: E45 remeasures).

The surrogate is the proposal that keeps its draws. It keeps 200 of 200 effective draws under cr at every size, while thinning falls below 10 by 100-200 tips; per E-step it costs 4.7-7.0 times thinning under cr, but per effective draw it is the cheaper from 100 tips (E4). Over 67 cells of size, turnover and ED effect it falls below 10 in 7, all with a strong negative ED effect, against 54 for thinning (E7) (superseded for the ed and ned cells: E46 remeasures). On the exponential link and the current build, at the generating values over the sampling fraction, it returns all 200 draws under cr and leads by 11-68 times on dd, 22-80 on d, 18-40 on nd and 2.2-12 on edc at every \(\rho\) down to 0.2; on ed, ned and nedc it leads by at most 3.2 times down to \(\rho\) = 0.4 and trails at 0.2, and those models keep fewer than 10 effective draws of 200 on both proposals in nearly every cell (on ned at \(\rho\) = 1, 7.6 on the surrogate) (E6, E6b, which replaces E6's linear-link ranking in the 3 of 20 cells where they differ). For D models the surrogate gives the larger sample but failed its adoption rule, so thinning is their default (E19) (superseded: E46 remeasures, since E6b finds the surrogate 18-80 times ahead on d and nd at the generating values); K models can only use thinning. Between the two thinning variants for ned, the ED-aware one is ahead on 77% of trees, not the 95% the rule required; it is the thinning variant for ED models because it completes where the mean-field one fails (23 trees) and thinning is not their default proposal (E3, decision of 2026-10-19).

Convergence is slow in iterations and noisy per iterate. Per-iterate noise falls about as 1/√draws (1-7% across rates; extinction 2-7 times speciation), and no settling criterion gives a stable scaling (E13). The mc_error rule never stopped while the trace was still moving (0 of 17) and lands about 2% from a 400-iteration reference, for cr and dd (E9). The time of a lineage-level fit is in its starting point: one ned iteration from the box takes 77 s, 78 min and 5.1 h at 100, 200 and 400 tips while its parts take under two minutes (E15, historical); the ladder's nested starts follow from this. On E40's 16 smallest real trees, continued to 1000 iterations, 34 of 72 fits meet the stopping rule: dd 16 of 16, ned 12 of 16, cr 6 of 8, nd and nk 0 of 32 (E41).

The relaxation default depends on the sampling fraction. A fixed relaxation step of ⅛ converges 107 of 139 diverging mean fields at 1000 sweeps against 77 for "auto" (E10). At the budget the package ships (20 sweeps at \(\rho = 1\), 200 below), ⅛ converges 0 of 192 units at \(\rho\) = 1 against 131 for "auto", at 1.4-3.3 times the E-step time and with no gain in effective sample (ratio 0.987 on 760 contracting trees), and 142 against 131 at \(\rho\) = 0.2; its whole gain on the diverging units is 11 linear-link units at \(\rho\) = 0.2 (E10c). The default is therefore "auto" at \(\rho\) = 1 and a fixed ⅛ below (the author's decision of 2026-10-20, departing from the literal wording of E10c's rule, which would make ⅛ the default everywhere). A mean field that converges only under relaxation leaves an effective sample of about 1 either way. The fit records the relaxation used and whether the mean field converged in its notes, and print() mentions it when it is not the default (decided 2026-10-19).

2. Are lineage-level coefficients recovered and selected?

The ED coefficient is estimable; selecting ned over dd needs a strong effect. At turnover 0.10 the 0.80 selection contour lies between -0.35\(\lambda\) and -0.60\(\lambda\) at 100, 200 and 400 tips, with false positives at 3.3% (E12); at turnover 0.25 and 100 tips it lies between -0.35\(\lambda\) (0.67) and -0.60\(\lambda\) (1.00) (E16). At -0.15\(\lambda\) selection is 0.05-0.55 (E12) and centring ED does not lift it (0.10-0.20; E18). The coefficient's sign is recovered (≥ 0.95 at |effect| ≥ 0.15\(\lambda\)) while its size is attenuated on unfinished fits (0.69, 0.40, 0.19 at -0.60\(\lambda\) from 100 to 400 tips) and comes back with more iterations (0.96, 0.67, 0.72 at -0.35\(\lambda\) in E12b) (E12); on the 7 fits that finished at 200 tips it is 1.02 (E29). Every E12, E16 and E18 cell is clock-limited, so these rates are floors. (superseded: E46 and E47 remeasure)

Specificity holds. On another group's null trees (evesim) ned is selected on 0.05 and 0.00 of trees (E11), and the two simulators agree in distribution (E14). False positives for ned are 2 of 60 null trees in E12 (2 of 20 at 100 tips) and 0 of 15 in E16; for nedc, edc and ed (exact simulator) 0-1 of 20 (E18, E24). The exception is d, at 0.15 (E24) (superseded: E47 remeasures).

Of the single-covariate models, edc meets every half of its rule (ratio 0.96 and 1.02, selection 0.95 and 0.85, no false positive); ed under the exact simulator recovers (0.85, 1.10) and is selected at -0.35 (1.00) but not at +0.35 (0.75); k meets every half at -0.35 and is sampler-limited at +0.35 (E24). d recovers at +0.35 only (E24) (superseded: E46 remeasures).

Turnover and incomplete sampling are where recovery breaks. At turnover 0.5 the thinning-based ned fits of E2 (historical) are biased by 1-4 SD and null siblings select ned on up to 0.9 of trees; at \(\rho\) 0.7 the effective sample falls below 10 outside the -0.15\(\lambda\) cells. In the ladder study the ned and nk cells at \(\rho\) 0.75 and 0.50 are sampler-limited (E27) (superseded: E47 remeasures). Fitted extinction slopes, generated at zero, come back away from zero for ed (-0.25 to -0.73 per unit) and nk (-0.06 to -0.15 per split) (E24, E21).

3. The pendant-age model and the ancestral-splits covariate

nd: the sign of \(\beta_D\) is recovered, its size is not, and no fit has finished. At ±0.35 the ratio is 0.30 and 0.47 on thinning and 0.43 and 0.63 on the surrogate with a 30-minute clock (E19); with 300 iterations and a four-hour clock it is 0.67 (-0.35) and 0.80 (+0.35, sampler-limited), with 0 of 39 fits meeting the stopping rule (E30). At 200 tips the ladder chooses nd on 5 of 20 of its own trees and the pendant-edge pair nd + ned on 9, with \(\beta_D\) at 0.68 (E31). In the pairwise contest nd beats dd on 19 of 19 trees at -0.35 (E30), while in the ladder at 100 tips it is chosen on 0 of 10 (dd on 7) (E23). Every fit in this paragraph maximised the D likelihood before emphasis 4e0b9dd, under which d's \(\beta_D\) returned at 0.70 and 1.46 of its value on complete trees (E44). (superseded: E46 and E47 remeasure)

D and ED share the pendant edge. On the path, the mixed generator returns D alone on 0.80 and both on 0.15, so D and ED are not separable at 100 tips (E20); at 400 tips D replaces ED as the first covariate on ned trees (E22) (superseded: E47 remeasures). Within a tree their rank correlation is 0.57-0.74 (E39).

K: recovered, selected on most of its own trees, short of 0.80. By fitting, \(\beta_K\) comes back at 0.95 and 0.92 of its value at ∓0.35 with selection over dd on 0.15 and 0.60 (E21, E21a, clock-limited). In the ladder nk is chosen on 7 of 10 trees at 100 tips (E23) and on 11-15 of 20 in the six readable conditions of E27, where \(\beta_K\) returns at 0.65-0.76 at 50-200 tips, 1.04 at turnover 0.25 and 1.30 at 0.40. The two pages differ on the ratio; the designs differ in box, clock and in whether the fit is the ladder's. On the path K is admitted on at most 1 of 20 trees of other generators, and on nk trees ED enters first on every tree, carrying the K signal with the opposite sign (E21, E21b). (superseded: E46 and E47 remeasure)

4. Model choice: the ladder and the path

Monte Carlo bias does not decide AIC rankings. Correcting the Jensen gaps moves rankings by 0.01-0.25 AIC; it would reverse 1 of 101 wrong choices, a tie at 0.003 AIC (E8). The package warns when the chosen fit has an effective sample below 10 (from emphasis f30a791).

The reading matters. At 200 tips the minimum-AIC reading admits nd on 3 of 12 null trees, the parsimonious reading (fewest parameters within 2 AIC) on 0 (E23); across E27's cells the parsimonious reading keeps lineage-level false positives at 0-2 of 20 while the minimum fails at the reference and 50-tip cells (7 and 6 of 20) (E27). (superseded: E47 remeasures) compare_models() reports the parsimonious choice (from emphasis f30a791).

The ladder recovers the clade-level answer and nk, not ned or nd. At 100 tips the generator is chosen on 0.80 of cr and 0.90 of dd trees (E23). Null trees keep dd on at least 16 of 20 at the reference, family-like, 50%-sampling and 200-tip cells, and move to cr at turnover 0.25 and 0.40 (8 and 6 of 20 keep dd) (E27). ned is chosen on at most 2 of 20 of its own trees in any condition although its coefficient is recovered at 0.91-0.98 in three cells (E27). (superseded: E46 and E47 remeasure)

The dgLARS path selects the null and N, and not the lineage-level covariates. At the generating parameters the null is chosen on 0.85-1.00 of cr trees from 50 to 400 tips; on dd trees N is in the set on 0.85-1.00 but N alone on only 0.40-0.70; ned's N + ED is never chosen on 0.80 (E20, E22). Taken at a fitted estimate, as a user would, dd's exact set falls to 0.35 and 0.25 and absent covariates enter on 0.45-0.55, so the path is not usable as a route by its rule (E33). On E40's 170 real trees the path's BIC set equals the ladder's choice on 16 (Real phylogenies). (superseded: E47 remeasures)

5. The extinction side, conditioning and the information a tree carries

At 200 tips and turnover 0.10 extinction is not estimable, by any estimator. The fitted extinction rate is a median 26 times below its generating 0.05, with the ED slope on its box limit, on finished fits as on the others (E29); widening the box moves the estimates with it for a median log-likelihood gain of 0.01, and an extinction slope on ED is returned on the box limit whatever its generating value (E32). The exact bd_ML puts extinction at zero on 37 of 40 of the same trees; dd_ML puts it 2.4-3.2 times too high (E32).

Neither conditioning nor the emulator is the cause. Conditioning on survival brings the ned fit's extinction rate within a factor of 2 of 0.05 on 1 of 20 trees, and on 0 of 20 under the corrected emulator (E34, E34r). The package's survival emulator labelled clades that pass the lineage cap as extinct; with them labelled survived and 2000 training clades it meets its accuracy rule in all four exact cells, and emphasis d3dee92 ships that (E36, E36c).

For the lineage-level models the survival factor has no adequate emulator. Against a Monte Carlo reference with a standard error of 0.01 in \(\log P\) at 200 points a model, the package's emulator has a 90th-percentile error of 0.5 (k) to 2.6 (nk) in \(\log P\), against the rule's 0.05, at 2000, 5000 and 20 000 training clades alike, and a retraining standard deviation of 0.03-0.12 against 0.02. Conditioned fits of d, nd, ed, edc, ned, nedc, k and nk therefore use a direct Monte Carlo estimate of \(P(S)\) at each parameter value; cr and dd use the exact functions, which the same reference matches (median error 0.007 and 0.008) (E43).

The information map is of the trees. With the exact constant-rate likelihood, extinction is detected (zero excluded on ≥ 80% of trees) never up to 1600 tips at turnover 0.10, from 1600 at 0.25, from 400 at 0.40 and from 100 at 0.50-0.60; it is bounded within a factor of 2 only from 400-1600 tips at turnover ≥ 0.40. Under exact diversity dependence it is detected on every tree from 100 tips at every turnover (E38, on trees kept within 0.6-1.6 times the target size; its cr cells up to 800 tips at turnover up to 0.40 are superseded: E45 remeasures). The study's 200-tip, turnover-0.10 trees sit where cr detection is 5%; on E40's 16 smallest real trees 0 of 16 exclude zero extinction (E41).

Where trees hold more extinction, a K slope appears in one direction. At turnover 0.25 the negative K slope returns at 0.75 of its value with the right sign on 18 of 20 and the zero slope at +0.004, while the positive arm is sampler-limited, so by the rule the slope is not estimable (E35); at turnover 0.10 the negative slope returned at 0.36 (E32). The extinction path does not detect an ED driver; it admits D instead on 0.50-0.70 of trees where extinction rises with ED (E33) (superseded: E47 remeasures).

A sampling fraction off by 0.1 does not change the chosen model (same choice on ≥ 18 of 20 trees), and changes \(\beta_N\) by a median 19% when told 0.82 instead of 0.92 (5 of 20 trees outside the rule's 0.75-1.25 band) and \(\beta_K\) by 5-8%; the fitted extinction rate does not register the change (E35, part b; its nd rung maximised the D likelihood before 4e0b9dd).

6. On complete trees, and the exact estimate on real ones

With nothing hidden, every model's likelihood returns its generating coefficients once the D likelihood holds D constant between events (E44; before the change d returned \(\beta_D\) at 0.70 and 1.46 of its value). Raw ED shows a per-tree ridge between \(\beta_0\) and \(\beta_{ED}\) that centring removes. On E40's real trees, MCEM reaches the exact maximum for cr and dd (E42), and the exact dd maximum puts extinction at zero as the fits do.

7. Real trees

The workflow runs on deposited family trees. On 46 of 47 Condamine family trees the cr fit is within 0.53 nats of bd_ML (E5); the 47th, Macropodidae, is refused for a tip spread of 19% (the limit is 15%). Under the exponential-link ladder Corvidae goes to nk (by 12.0 over ned), Icteridae to ned and Cotingidae to dd (E25, E28).

Across families, a lineage-level model is chosen in about half the usable trees. For the 29 families inside the calibrated range, refitted with 300 iterations, 24 are usable and the parsimonious choice is cr 4, dd 5, nd 1, ned 6, nk 8 (E28, with the nd rung of E48). On the merged table of 46 families (E28's rows for those 29, E26's for the other 17), 33 are usable; dd leads cr by more than 2 AIC in 32, \(\beta_N < 0\) in all 46, and by minimum AIC 20 of the 33 choose a lineage-level model (nk 12, ned 6, nd 2: the honeyeaters and the petrels) and 9 a diversity-dependent one (E26, E48). The parsimonious choice comes first because it is the rule registered before the data, and MCEM log-likelihoods carry Monte Carlo error and a downward bias, so AIC differences below 2 are within the estimator's noise; the minimum-AIC split is the sensitivity reading beside it. Where both are readable (24 families) nk has a lower AIC than ned in 18, by a median 3.4. K raises speciation in 29 of 30 read nk fits (median factor 1.41 per sd), ED lowers it in 29 of 31 read ned fits (median factor 0.69), and D lowers it in 28 of 30 read nd fits off their box edge (median factor 0.69) (E48). Only 24 of 138 lineage-level fits met the stopping rule, and E27 measured K at 0.65-0.76 of its value on such fits, so the coefficients are read as attenuated (E26, E28). Extinction is not read: fitted turnover is below 0.05 in 41 of 46 families (E26, E48).

On real trees K and ED overlap. Their within-tree rank correlation is a median -0.44 on the family trees against -0.18 on reconstructed simulated trees, and stronger the less balanced the tree (-0.51 across families), so a K choice and an ED choice on an unbalanced family tree are not distinct mechanisms (E39). On the 46-family table the K factor rises with Colless imbalance (rank correlation 0.55; 0.34 over the 26 below the four most imbalanced trees), and the Pybus-Harvey \(\gamma\) falls with the dd-over-cr lead (rank correlation -0.88) (E26). E40's coefficients track imbalance in the same direction: over its 170 trees the K factor rises with it (rank correlation 0.53 over 124 read nk fits) and the ED factor falls (-0.57 over 145 read ned fits) (Real phylogenies).

On clades cut from the class trees the choice does not carry across trees. On E40's 170 finished trees a lineage-level model is the parsimonious choice on 125 (nk 67, ned 50, nd 8) with the nd rung refitted under the corrected D likelihood (E48), and the bird clades choose ned or nk on 72 of 77; but a clade and the family tree of the same family agree on the choice on 10 of 29 pairs (3 of the 18 where both chosen fits are read), below the 70% the pairs rule requires, so the class-tree clades are reported and not read (E40, E48). D lowers speciation on 92 of the 109 read nd fits whose \(\beta_D\) is off its box edge (median factor 0.78 per sd), from fits of which 7 of the 217 refitted met the stopping rule (E48). Where both members' fits are read, the ED factor is below 1 on both on 10 of 10 pairs, the K factor above 1 on both on 11 of 11, and the D factor on the same side of 1 on 8 of 9 (Real phylogenies). Continuing the 16 smallest trees' fits to 1000 iterations leaves the choice unchanged on 16 of 16, and moving a box edge out gains more than 1 nat and 2 standard errors on 4 of 63 edge fits, so the box stays (E41).

What is still open

The validation series (V1–V12) remeasures the results above on one build (182055b, and since the amendment of 2026-11-07 b21083f for the lineage-level fits) under one protocol; each page's superseded mark names an E experiment whose remeasurement is now a V experiment (E46 is V5; E47 became V7 and V9). On 7 November: V1, V2 and V4 are complete; E45 is complete for cr and dd; V3, V5, V6, V7 and V9 give no reading for any lineage-level model, whose fits ended on the iteration cap or the 22-hour clock, many with the clock spent in the unconditioned start before any conditioned segment (Open issues); V3, V7, V6 and V9 rerun under the default rule projected, a fit on the 1000-iteration cap read as the estimate (amendment of 8 November, evening). Until they are read, the lineage-level statements in section 2 rest on the superseded measurements.

Questions the pages name and no experiment answers: