Files
aevolution/docs/engineering/M12_RESERVE_STRATA_EXPERIMENT.md
T
dingo 214991c9a3
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
M12.2b: test exact reserve-strata population candidate
2026-10-16 06:01:42 -04:00

129 lines
6.9 KiB
Markdown

# M12.2b: Exact reserve-strata experiment
## Result (2026-10-16)
The exact-state candidate removes the synthetic baseline's false-extinction
mechanism: **all 83 comparisons match every identified living state, census,
birth/death count and resource compartment at every tick** on MSVC and GCC.
This includes 40 development cases, 40 held-out cases and three original scarcity
runs. All 83 comparison records are identical across platforms. Conservation is
exact. No old model, rejection result, persistent format or golden was replaced.
This completes the isolated M12.2b research segment under
[ADR-0134](../adr/0134-reserve-strata-population-experiment.md). It does **not**
complete M12.2 or authorize automatic tier selection. Exact-state retention can
require one row per living individual and is not yet a cellular population model.
Source: `experiments/resolution/reserve_strata.hpp` and `reserve_strata.cpp`.
Raw results: [MSVC 19.44](results/m12-2b-strata-msvc1944.jsonl) and
[GCC 14.2](results/m12-2b-strata-gcc14.jsonl).
## Representation and predeclared instrument
Age-class averages lose which members can pay upkeep. The candidate instead holds
exact per-member age, energy and matter in run-length encoded contiguous identity
ranges. It merges only identical states. Parent priority stays oldest-first then
original identity; partially reproducing ranges split before their reserves are
debited. Dead working rows retire without renumbering surviving identities.
The original fixture's shared scalar `Rule` predicates are unchanged. An independent
group implementation performs the five phases and checked inventory arithmetic.
No sampler, approximate birth count, age-bin midpoint or resource averaging enters
the candidate. The original identified implementation remains the reference.
A third, compact-individual implementation invokes that reference step and retires
dead rows while retaining original identities and tie order. It is checked against
the same identified outcomes before being used as a cost control.
ADR-0134 declared zero per-tick error before running the experiment. The instrument
checks identified portraits rather than census agreement alone, so charging the
wrong parent at a capacity boundary fails. Controls cover partial reproduction,
same-age unequal reserves, compressed identical members, starvation of a subset,
fragmented identities, empty populations, interval partitioning at unchanged
one-tick quantum, unsupported underfunded breeding and checked-product overflow.
The original 100-founder distribution runs at supplies 240/300/360 for 10,000
ticks. The seeded suite uses 100 founders in blocks of five identical members,
with heterogeneous seeded ages and reserves. Seeds 0-7 are development and 8-15
held out, each at supplies 180/240/300/360/480 for 1,500 ticks. These regimes stress
the research instrument; they do not model M11 spatial competition or diversity.
## Costs
Local Ryzen 7 9800X3D workstation, 8 cores/16 logical processors; Windows x64
MSVC 19.44 Release and Ubuntu WSL2 GCC 14.2 Release. No affinity/frequency lock,
RSS profiling or controlled production runner. One warmup plus three fresh
samples, with strata/reference order alternated. Median milliseconds over
10,000 steps follow; verification and setup are outside the timed loop.
| Supply | MSVC historical | MSVC compact | MSVC strata | GCC historical | GCC compact | GCC strata |
|---|---:|---:|---:|---:|---:|---:|
| 240 | 27.898 | 3.824 | 5.162 | 26.315 | 6.211 | 2.452 |
| 300 | 30.206 | 4.202 | 8.549 | 27.543 | 6.425 | 2.957 |
| 360 | 35.342 | 4.867 | 9.112 | 32.771 | 7.583 | 3.153 |
The candidate beats the historical retained-dead reference on both compilers, but
is slower than compact individuals on MSVC and faster on GCC. Therefore the old
benchmark's apparent aggregation advantage includes dead-row scanning, and these
local observations do not justify a production capacity promise or universal win.
| Supply | Historical peak rows | Compact peak active rows | Strata peak rows |
|---|---:|---:|---:|
| 240 | 2471 | 120 | 120 |
| 300 | 2625 | 131 | 131 |
| 360 | 3156 | 148 | 140 |
Peak state payloads for compact/strata are 4800/4800, 5240/5240 and 5920/5600
bytes respectively. Both carry exact identity, age and reserves; little grouping
is possible in these original heterogeneous histories. Peak group count and
peak individual count may occur at different ticks. Raw files also report vector
capacity; allocator growth differs between MSVC and libstdc++. Payload/capacity
exclude object bytes, per-step temporary vectors/sort scratch, allocator metadata
and RSS. They are not peak working-memory budgets or persisted-size estimates.
## Reproduction and checks
```powershell
./tools/windows-check.ps1 -Preset release -Action check
./build/release/evo-reserve-strata.exe
```
The existing Ubuntu WSL build directories use the checked-in presets with explicit
binary-directory overrides, keeping native and Linux object files separate:
```sh
cmake --preset release -B build/linux-release
cmake --build build/linux-release --target evo-reserve-strata
./build/linux-release/evo-reserve-strata
cmake --preset analysis -B build/linux-analysis
cmake --build build/linux-analysis --target evo-reserve-strata format-check
ctest --test-dir build/linux-analysis -R reserve-strata-qualification --output-on-failure
cmake --preset asan -B build/linux-asan
cmake --build build/linux-asan --target evo-reserve-strata
ctest --test-dir build/linux-asan -R reserve-strata-qualification --output-on-failure
```
Executed: native Release build and **70/70 CTests**; GCC Release experiment and
its CTest; clang-format 19 repository check (now includes experiment headers);
clang-tidy 19 analysis build and its debug CTest; GCC ASan/UBSan experiment CTest.
`--verify` runs the full controls and comparison matrix without timing. JSONL
validation checked 83 comparison and 27 timing rows per platform and exact
cross-platform equality of comparison records. Tests throw under `NDEBUG`.
## Remaining qualification
The candidate is not representable by Cohort v1's age-class totals. There is no
automatic conversion, persisted representation, spatial support or ancestry
codec. Retained synthetic IDs do not supply new individual-parent history.
Mixed genomes, M11 contact/predation/controllers, funded bud placement, mutation,
offline suspension, viable reproductive reserves and migrations remain unsupported.
No runtime adapter silently admits these states.
Next, predeclare paired M11 cellular ensembles and demographic/biomass/diversity
bounds, retain a compact active control, and evaluate whether any supported state
can safely group or approximate without erasing spatial/viability distinctions.
Any new representation needs its own complete public/durable contract. Measure
full-region/conversion and active fallback costs on the target Linux host, with
explicit backpressure. WSL2 workstation evidence is not the R730xd qualification.
M11.9 human product acceptance also remains unrecorded. See the
[delivery plan](M12_DELIVERY_PLAN.md) for dependent gates.