M12.2b: test exact reserve-strata population candidate
Foundation / checks (deploy) (push) Successful in 5s
Foundation / checks (ecs) (push) Successful in 2m26s
Foundation / checks (tsan) (push) Successful in 11m44s
Foundation / checks (client) (push) Successful in 19m23s
Foundation / checks (network) (push) Successful in 19m45s
Foundation / checks (checks) (push) Successful in 45m39s
Foundation / checks (postgres) (push) Successful in 45m56s
Foundation / postgres-tools (push) Successful in 23s
Foundation / server-image (push) Successful in 31s
Foundation / image (push) Successful in 3m35s

This commit is contained in:
dingo
2026-10-16 06:01:42 -04:00
parent b3550298f7
commit 214991c9a3
15 changed files with 1031 additions and 5 deletions
@@ -0,0 +1,79 @@
# ADR-0134: Exact reserve-strata population experiment
## Status
Accepted for the isolated M12.2b research instrument (2026-10-16). No runtime
tier policy, biological rule, public API, persistent version or golden changes.
## Context
ADR-0041 and M12.2a reproduce false extinctions when an age class pays upkeep
from its mean reserve. Narrower age bins do not reliably repair this. Before
approximating M11 cellular ecology, test whether retaining exact age/reserve
variation removes that mechanism in the original synthetic model.
## Decision
Add an experiment using exact per-member age/energy/matter, run-length encoded
only across contiguous identities with identical state. Retain monotonic identity
ranges and oldest-first, then identity-first birth priority. Partial reproduction
splits a range; equal newborns can share a range. Dead members leave the working
representation; no averaging, resampling, age reset or death-rounding occurs.
Keep the existing failed fixture, benchmark and goldens unchanged.
Predeclare the acceptance instrument before measurement:
- Every tick must exactly match identified living portraits, population, reserve,
cumulative births/deaths, environment, detritus and conserved totals against the
retained individual reference. No terminal-only tolerance. Any mismatch fails.
- Revisit supplies 240/300/360 for 10,000 ticks from the original 100 founders.
- Seeds 0-7 are development and 8-15 held out. For each seed, test supplies
180/240/300/360/480 for 1,500 ticks. Generate blocks of five identical founders
with seeded heterogeneous ages/reserves, so both compression and variation occur.
These are declared synthetic regimes, not M11 calibration or diversity evidence.
- Check contested partial births with same-age unequal reserves, empty/dead input,
fragmented identities, conservation and interval partitioning without enlarging
the model's one-tick integration quantum.
- Report local time, peak active rows, state payload and vector capacity alongside
error. Timing runs exclude verification and setup; no portable timing assertion.
Report the historical reference's retained-dead-row bias explicitly. Include an
identified compact-individual cost control that retires dead rows while keeping
birth priority, and check its identified outcomes against the same reference.
The candidate representation is **not** Cohort v1: that contract retains only
age-class totals and cannot recover the exact reserve distribution. No converter
is provided that silently upgrades v1. This experiment retains identified state;
it must not be described as statistically sampled ancestry or a production model.
## Alternatives considered
Changing the old fixture would invalidate the rejection evidence. Exact individual
rows provide a useful control but cannot demonstrate any grouping benefit. Reserve
bins that still mix viable and starving members need an error/viability model;
test the exact sufficient-state control first. Directly wiring the old aggregate
into the cellular owner remains prohibited by ADR-0041.
## Benefits
Tests a specific repair with an independent implementation of interval accounting
and identified outcomes. Shared scalar Rule predicates keep the synthetic biology
constant while the representation changes.
## Costs
Heterogeneous states may require one row per living member. Identity order can
prevent merging otherwise identical states. Per-tick splitting and sorting cost
time and temporary memory; report rather than assume a speedup.
## Risks
This single-genome, nonspatial fixture lacks M11 controller, contact, predation,
mutation, budding placement and suspension. Exact synthetic fidelity does not
qualify current cellular workloads or establish a target-host fallback budget.
## Consequences
Add release-safe CTest coverage and run formatting, analysis and ASan/UBSan for
the experiment. Publish results with limitations. M12.2 remains open until M11
paired ensembles, supported-state admission and target-host cost/error evidence
are accepted; M12.3-12.5 remain dependent on that qualification.
+2
View File
@@ -1,5 +1,7 @@
# Architecture decision records
- [ADR-0134: Exact reserve-strata population experiment](0134-reserve-strata-population-experiment.md) - M12.2b isolated exact-state candidate and predeclared instrument; no runtime tier qualification.
- [ADR-0133: Ecological identity and time evidence](0133-ecological-identity-and-time-evidence.md) - M12.1 engineering evidence, viability/time boundaries and explicit deferrals; no runtime changes or qualified aggregation.
- [ADR-0100: Primitive lives and persistent control](0100-primitive-lives-and-persistent-control.md) - accepted M11.2 direction: finite spawn pools, equal full-grown founders, new cellular checkpoint and atomic life publication. 11.2a adds in-memory funded creation; durability/integration follow.