The paired path had the same defect as the unpaired one: it subtracted two
step counts and handed the differences to the signed-rank test, so a pair
holding a run the wall clock stopped at step 12 entered as a difference
neither run supports. Twenty seeds where the first arm was still clean at
step 12 and the second violated at step 5 in six of them read sign -1 and
p 0.0011, pointing at the arm that never violated.
A pair is now scored the way the unpaired comparison scores one and tested by
the exact sign test over the pairs whose order censoring determines, which is
what the log-rank stratified by seed reduces to here. The signed-rank goes
with the differences it needed: a magnitude-based paired test wants a
difference from every pair, and the arms censor on different clocks. The
median difference stays, over the pairs where both runs violated, and says so.
stepTimes threw the censoring flag away and handed the rank-sum a plain
number per run, so a run the wall clock stopped at step 12 was ranked as one
that violated at step 12. That was defensible while every clean run sat at
the budget, the largest value any run could take, and it stopped being
defensible when a clean run started being censored where it stopped.
Twenty runs clean at step 12 against twenty violations at step 100 read a12
0.000 and p 4.683e-10 from the rank-sum, in the same report as a log-rank
reading p 1.0000. The pairwise comparison is now the Gehan test over the
observations themselves, and the report says how many run pairs censoring
left with no order between them, which is how much of the effect size is the
null value rather than an observation.
Defects per thousand actions divided by every dispatched step, so a
spec whose setup logs in inflated the denominator by however many steps
that took. It is the same error the run gate had, and it does not cancel
between arms.
A model run is separable because only an llm-selected action stamps
next_action.source. A seeded run is not: its setup returns through the
same entry with no marker, and 11261 dispatched steps across the 169
recorded runs carry no source at all, so excluding on it blind would
report every seeded run as having explored nothing. The seeded arm
counts as before and a test pins that.
A run stops at whichever comes first, the step budget or --duration, so
a clean run that reached the wall clock exited with fewer steps than the
budget and was still credited with the whole of it. The model arm pays a
network call and a screenshot per step, so it reaches the wall sooner and
was handed exposure it never had.
Nothing checked that two arms shared a budget either. Thirty identical
clean runs under budgets of 400 and 100 read a12 0.000 and p 1.685e-14
from the rank-sum while the log-rank in the same report read p 1.0000.
groupArms already refused this within one arm.
The claims the old convention left in comments and report lines are
corrected rather than left standing beside the new behaviour.
--paired contrasts two arms running the same seeds seed by seed with the wilcoxon signed-rank test rather than treating them as two independent samples, reporting the per-seed differences, the sign, a12 within pairs and the seeds usable in one arm only. --question names the family holm corrected within, and the family size is recorded next to the p-values rather than left to the reader to reconstruct.
an obligation that never discharges is reported when the run ends, and timing it by the step that armed it recorded a liveness failure flushed at the budget as a violation found on step 1. the survival analysis now measures the detected step, falling back to the origin for campaigns written before the field existed, and says how many events that moved. the report gains the first and third quartiles beside the median.
A host asleep mid-run tested nothing, and charging that sleep to an arm reports
it slower for a reason unrelated to the arm. The legend also claimed wall clock
while the number was monotonic. Campaigns written before the split are still
read through the old field name so their run hours do not silently zero.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Defects per thousand actions counted every step, including steps that chose
nothing and steps whose action was never dispatched. The inflation is
policy-dependent, so it does not cancel between arms: on the fixture campaign
the model arm's yield was reported at 60.3 per thousand against a true 120.7,
because half its steps did nothing.
A runs.jsonl without the count is refused by name and line rather than read as
zero actions, which would report every per-action rate wrongly. The report also
carries steps beside actions now, so the gap is visible rather than folded into
a denominator.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Steps to first violation with clean runs right-censored at the budget, since
per-run yield is a binary at 11 to 45 percent and separating two arms on it
would need roughly 80 runs per arm. Kaplan-Meier, log-rank, Wilcoxon rank-sum
with Vargha-Delaney A12, Holm within each family.
A hand-rolled log-rank that is subtly wrong is a silent-wrong-number generator
and would be believed, so every statistic is validated against a published
worked example with the source named in the test: R survdiff on aml, Freireich
6-MP, Hollander and Wolfe 1973 for the rank sum, printed p.adjust output for
Holm. Two could not be: the k>2 log-rank, guarded by calibration instead, and
the tie-corrected variance, checked against an exact permutation variance.
Failed and timed-out runs are excluded as missing data and counted by reason,
never treated as censored observations, which would bias the result.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
pj