The two runs allocated contested breeding slots in different orders: the
individual run by member insertion order, the aggregate run by stored class order,
which is youngest first. When eligible parents outnumber free slots that decides
which parents reproduce, for a reason that has nothing to do with averaging. The
control never reached the capacity boundary and so could not see it, which left
ADR-0040's central claim unsupported rather than merely imprecise.
The allocation order is part of the stated rule now, oldest first, honoured at both
resolutions: classes partition ages, so the oldest class holds the oldest members,
and ordering inside a class stays something only the individual run can see, which
is a resolution effect rather than an implementation difference.
A second control sits at the capacity boundary, twelve eligible parents contending
for one free slot, and both controls compare the whole distribution of members
rather than censuses. That change matters by itself: whoever breeds hands the same
investment to a child either way, so every total agrees while the distribution does
not, and a census-only control cannot see a different parent having bred. The
probe diverging only by tick 44 is that effect.
Attribution was verified by mutation rather than argued. Restoring the individual
run's insertion order fails the new control; restoring the aggregate run's
youngest-first order fails it; the earlier one-unit drift in aggregate upkeep still
fails the first. Since either allocation order alone fails, the reported divergence
is attributable to averaging.
My first attempt at the control was wrong and the same sweep caught it. I gave the
contested fixture ages descending with the index, reasoning that this made
insertion order differ from age order; it does the opposite, and the
individual-side mutation passed. Ages ascend now, so oldest-first disagrees with
both the insertion order and the stored class order. Without the sweep the new
control would have looked like it established attribution while covering half of
it.
Also added the depletion check the ADR's attribution argument relies on: a shared
store that ran out would make intake order-dependent too. I had written that the
fixture asserts this before it did, which is now true.
The measurements were regenerated and are unchanged. Saying so plainly rather than
as confirmation: the finding was about what the numbers can be attributed to, not
about their values.
Release 38/38, analysis, ASan 38/38 and clang-format all ran and passed here.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
M4.5 closes M4 by measuring what aggregation costs, and adds no library code
except one defect fix the measurement found.
One rule is written once and applied at both resolutions. The control is what
makes the rest mean anything: with one member per class a class average is that
member, so the two runs must agree exactly, and they do, tick by tick, in every
channel. A deliberate one-unit drift in the aggregate rule is caught by it, so the
instrument is verified rather than assumed. Without it a measured gap would be
indistinguishable from a transcription mistake in one of the two runs.
The result is reported, not smoothed into a tolerance. At class widths ten and
fifty the aggregate population survives ten thousand ticks within 375 parts per
thousand on population and 43 on energy. At twenty-five, one hundred and two
hundred it dies out entirely while the individual run sustains seventy-five
members: a false extinction caused by resolution alone, because a whole class
starves the moment its average cannot pay upkeep, where individuals starve a few
at a time and the survivors keep the shared per-capita share up.
The collapse is not monotone in class width. Twenty-five destroys the population
and fifty sustains it. An earlier version of the test asserted that coarser is
never better; the measurement refuted it, so the assertion went rather than the
widths being chosen to make it hold. The non-monotonicity is asserted now, because
it is surprising and because the warning written around it could otherwise become
wrong without failing. A tier policy cannot assume degradation worsens smoothly as
classes widen, and M4.6 inherits that.
Conservation was exact at every tick on both sides of all five runs, including the
three that died out. That is the clearest available demonstration of the thing
ECOLOGY_MODEL keeps saying: conservation is not ecological evidence.
Two flaws in my own experiment were found before any number was believed. The
aggregate model had been producing overlapping classes, so the model measured was
not one a cohort could represent; newborns now extend the youngest class while it
is narrower than the width, and every step is checked to construct a valid Cohort.
The first figures were discarded.
The comparison also found a real M4.3 defect, which is what a validation milestone
is for. plan_advance checked affordability with arrivals added before losses were
taken, but emitted the ledger legs outflows-first, so an interval affordable as a
whole underflowed leg by leg and was refused at commit. A fifty-tick interval
shedding 10,480 against 6,950 held while drawing 13,080 in reproduced it. Arrivals
are the first leg now, matching the order the arithmetic was checked in, with a
regression and a corrected escrow expectation in an existing test.
Release 38/38, analysis, ASan 38/38 and clang-format all ran and passed here.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>