Status rerun in full under emphasis b21083f and projected submitted 8 November (arrays v6, 3,200 jobs, one a task, queued behind V7; amendments of 2026-11-07 and 2026-11-08), a fit on the 1000-iteration cap read as the estimate; first pass on 182055b (993 of 3,200 fits, no cell–condition read) archived under results/V6-182055b
Series the validation series (replaces E27's \(\rho\) arms, E35(b))
Package build first pass: emphasis 182055b (1.0.0.9000), R 4.5.3; relaunch: b21083f (1.1.0.9000)
Driver sim/V6.R
Data / analysis results/V6/jobs/ on Bre; merged rows and tables in results/V6/; sim/V6-analysis.R

Question

Real clades are incompletely sampled, and the sampling fraction is itself an estimate. Recovery at \(\rho < 1\) has not been measured on a current build.

Design

cells the 100-tip E43 cells: 32
conditions \(\rho\) 0.9, 0.75, 0.5 told the truth; \(\rho\) 0.75 told 0.65 and 0.85
trees 20 per cell and condition (V2's trees at that \(\rho\))
fit the generating model, estimate_rates() defaults, conditioned
jobs 3200

Pre-set rules

  1. Per cell and condition, E46's rule (2) for every speciation coefficient; the smallest \(\rho\) at which it holds is the answer, beside V5's at \(\rho = 1\).
  2. A misstated \(\rho\) is harmless for a coefficient when its median ratio moves by less than 0.1.

Results

Interim, on the 927 fits of 5 November. The design sets no readability condition; the analysis uses E46's rule (1): a cell–condition is read when at least 16 of its 20 fits converged and the median final effective sample is at least 10.

Outcomes: 243 converged, 243 on the 1000-iteration cap, 212 on the 22-hour clock after some iterations, 225 on the clock with no conditioned iteration (iterations = 0), 4 E-step failures. Read: none of 160 cell–conditions; 76 can no longer be read whatever the remaining fits do (more than 4 of 20 unfinished), 84 are pending. No ned, nedc or nk fit has converged (0 of 167).

Descriptive, not a rule reading. Converged fits are on smaller trees than unconverged ones (median 23–41 against 34–68 tips for d, ed, edc, k), so only cr, where every fit converged, is free of that selection. cr median speciation ratio: 1.03 at \(\rho\) 0.9, 1.07 at 0.75, 1.02 at 0.5 (20–22 trees each). Misstated \(\rho\) at 0.75, on the 17 cr trees with all three fits converged: told 0.65 moves the median ratio by +0.16, told 0.85 by −0.13, both beyond the 0.1 of rule 2.

The fits with no conditioned iteration never occur under cr or dd; they make up 79% of ned fits, 60% of nk, 38% of edc, 33% of nd, 31% of nedc, 21% of ed, 14% of d, 6% of k. Their share rises with tree size (1% at 50 tips or fewer, 42% at 76–100, 62–69% above 150; logistic coefficient of log tips +5.8, z = 10) and, at equal size, as the stated \(\rho\) falls. They hold 2,939 of the 3,077 fit-hours spent. They did not stall in a first E-step: the fit spent its clock in the unconditioned start that precedes the conditioned segments (216 of 225 returned that start's trace, at 2400 draws), which iterations does not count (Open issues); since emphasis b21083f the start gets at most a quarter of the clock.

Six of them refitted under b21083f (pro, the series' settings and 22-hour clock, mc_error; sim/stall-verify.R, results/stall-verify/): nk 46 tips, ned 48 and 92, edc 50, nd 51 and 106. Every one now has a conditioned run — 419, 1000, 708, 705, 502 and 513 conditioned iterations over 17–57 segments — where the first pass had none; the start was cut at its 5.5-hour share on four, converged on one, hit the cap on one; five fits ended on the clock and one on the iteration cap, as E46's iteration counts under mc_error predict. Final effective sample 1.8–34 of 2400 on five trees and 277 on ned at 92 tips (\(\rho\) = 0.9), whose \(\beta_N\) (−0.007, generated −0.007) and \(\beta_{ED}\) (−0.063, generated −0.096) are the two lineage-level speciation coefficients the six fits recover; \(\beta_{ED}\) of edc at 50 tips comes back at 0.21 for 0.18. \(\gamma_0\) stays within 0.1 of the lower edge of the box on five of the six (−6.0 to −6.9 for generated −1.6 or −3.0), so the capped start lets these fits run but does not, by itself, recover extinction on them.