Four of five runs conserve perfectly; three are ecologically worthless

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>
This commit is contained in:
dingo
2026-09-24 21:57:00 -04:00
co-authored by Claude Opus 5
parent 14e0c1af5e
commit b4c3da56e9
11 changed files with 1155 additions and 11 deletions
+80
View File
@@ -1242,3 +1242,83 @@ parse, so the larger test edits in this increment were written to a scratchpad f
and run from there. That is the third distinct way heredocs have mangled content in
this project, after backslash escapes and `/tmp` resolving differently between
shells.
## M4.5 resolution comparison
Executed on this Windows host (2026-09-24): Release configure, build and `ctest`
38/38; the analysis preset (clang-tidy 19, warnings as errors) built clean; the ASan
preset configured, built and ran 38/38; `clang-format` 19.1.5 reported no
violations. All four checks ran and passed on the committed tree. No Clang or GCC
compile happens on this host, so Linux CI remains the first such result.
The increment adds no library code apart from one defect fix described below. One
rule is written once and applied at both resolutions; the individual run gives it a
member at a time and the aggregate run gives it a class average and multiplies by
the count. The aggregate run sub-steps at one tick, so interval-length error is not
confounded with averaging error, and interval-length error is therefore unmeasured.
The control is the part that makes the rest mean anything. With one member per class
a class average is that member, so the two runs must agree exactly; the test checks
population and every channel at every tick over forty ticks, and checks the class
count stays inside the sixty-four-class bound. Introducing a one-unit drift into the
aggregate upkeep is caught by it, so the instrument was verified rather than assumed.
Without this control a measured divergence would be indistinguishable from a
transcription mistake in one of the two runs.
Measured, and reported rather than smoothed into a tolerance. Class widths ten and
fifty: the aggregate population survives ten thousand ticks, with population gaps of
250 and 375 parts per thousand and energy gaps of 24 and 43. Class widths
twenty-five, one hundred and two hundred: the aggregate population reaches zero while
the individual run sustains seventy-five members. That is a false extinction caused
by resolution alone. The numbers are in
`docs/engineering/results/resolution-equivalence-msvc1944.jsonl`.
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, and the measurement refuted it; the assertion was removed rather than the
widths being chosen to make it hold, and the non-monotonicity is now asserted
instead, because it is surprising and because the warning written around it could
otherwise become wrong without failing. Five widths locate the effect and do not map
it; no mechanism beyond whole-class starvation is claimed.
Conservation was exact at every tick on both sides of all five runs, including the
three whose populations died out. Four of five conserve perfectly and three are
ecologically worthless, which is the clearest available demonstration of
ECOLOGY_MODEL's point that conservation is insufficient ecological evidence.
Two flaws in my own experiment were found and fixed before any number was believed.
The aggregate model originally shifted class bounds every tick and opened newborn
classes at the full declared width, which produced overlapping classes: the model
being measured was not one a cohort record could represent, and the first numbers
measured were therefore of a model the library could not hold. Newborns now extend
the youngest class down while it is narrower than the width, keeping the structure
contiguous and ascending, and every step of the aggregate run is now checked to
construct a valid `Cohort`. The earlier figures were discarded.
The comparison also found a real defect in M4.3. `plan_advance` checked affordability
with arrivals added before losses were taken, but emitted the ledger legs with the
outflows first. The ledger applies legs in sequence, so an interval affordable as a
whole underflowed leg by leg and was refused at commit with `ledger_failure`: a
fifty-tick interval shedding 10,480 energy against 6,950 held while drawing 13,080
in reproduced it exactly. Arrivals are now the first leg, matching the order the
arithmetic was validated in. The fix was confirmed load-bearing by restoring the old
order, which fails both suites, and a dedicated regression in `evo-population-tests`
covers an interval affordable only with its arrivals. One existing test had seeded
its escrow assuming the old order and was corrected, with its comment; that comment
had asserted the old ordering was what made the interval fundable, which was true
and is now true of the opposite order.
Coverage limits. One genome, so the mixed-genome diversity case ECOLOGY_MODEL also
asks for is untested. One model, one scarce axis, one initial distribution and one
capacity, so every number describes this fixture. The control runs over tens of
ticks because one class per age exhausts the class bound, while the measurement runs
over ten thousand: the instrument is checked at a shorter horizon than it is used at.
The asserted envelope bounds are round numbers above the measurement with room for
platform variation, so they are a regression guard and not a theoretical limit.
Interval-length error, crash behaviour at conversion boundaries beyond one process,
distribution across owners and persistence are all out of scope here.
The proof golden `d61ad5bf0c928c76`, the ecosystem fingerprint `2cd4a15f9fbe9949`,
the ensemble fingerprint `5ec59d390909b526` and the frozen genome fixture are
unchanged. No deployment manifest or image pin changed. Linux validation is pending;
no Gitea result is claimed since `b01a824`.
@@ -0,0 +1,6 @@
{"experiment":"resolution-equivalence-v1","model":"fixture-v1","compiler":"MSVC 1944","build":"release","commit":"m4.5","founders":100,"ticks":10000,"scarce_axis":"energy","lifespan":400,"capacity":200,"note":"Aggregate and individual runs apply one rule at two resolutions. The width-1 control matches exactly; these are the coarse measurements. Not an equivalence claim. See docs/adr/0040-resolution-equivalence.md"}
{"class_width":10,"individual_living":75,"aggregate_living":100,"individual_energy":6275,"aggregate_energy":6120,"living_gap_per_thousand":250,"energy_gap_per_thousand":24,"aggregate_survived":true}
{"class_width":25,"individual_living":75,"aggregate_living":0,"individual_energy":6275,"aggregate_energy":0,"living_gap_per_thousand":1000,"energy_gap_per_thousand":1000,"aggregate_survived":false}
{"class_width":50,"individual_living":75,"aggregate_living":120,"individual_energy":6275,"aggregate_energy":6000,"living_gap_per_thousand":375,"energy_gap_per_thousand":43,"aggregate_survived":true}
{"class_width":100,"individual_living":75,"aggregate_living":0,"individual_energy":6275,"aggregate_energy":0,"living_gap_per_thousand":1000,"energy_gap_per_thousand":1000,"aggregate_survived":false}
{"class_width":200,"individual_living":75,"aggregate_living":0,"individual_energy":6275,"aggregate_energy":0,"living_gap_per_thousand":1000,"energy_gap_per_thousand":1000,"aggregate_survived":false}