Files
aevolution/docs/engineering/AGGREGATE_BUDGET.md
T
dingo caccc86fc3
Foundation / checks (checks) (push) Successful in 5s
Foundation / checks (client) (push) Successful in 5s
Foundation / checks (deploy) (push) Successful in 5s
Foundation / checks (network) (push) Successful in 5s
Foundation / checks (postgres) (push) Successful in 5s
Foundation / checks (ecs) (push) Successful in 5s
Foundation / checks (tsan) (push) Successful in 4s
Foundation / postgres-tools (push) Successful in 5s
Foundation / server-image (push) Successful in 6s
Foundation / image (push) Successful in 5s
M12.2a: reproduce rejected population fidelity baseline
2026-10-15 22:50:24 -04:00

103 lines
5.6 KiB
Markdown

# M4.6 aggregate budget and sensitivity report
2026-10-15: [M12.2a](M12_POPULATION_QUALIFICATION.md) independently reproduces all
historical non-timing sensitivity/policy results, including eight false-extinction
scenarios and zero qualified widths. The original data below is preserved.
## Decision
M4.6 measurement work is complete; automatic tier selection remains **disabled**.
None of the tested resolutions meets the provisional research gate in ADR-0041.
This is a measured rejection of this model as a production tier policy, not a
promise to simulate every organism individually at world scale.
## Reproduce
```sh
cmake --preset release
cmake --build --preset release
ctest --preset release
./build/release/evo-resolution-benchmark > resolution-budget.jsonl
```
Windows: `tools/windows-check.ps1 -Preset release -Action check`, then
`build/release/evo-resolution-benchmark.exe`. `--verify` runs conservation and
sensitivity without timings and is registered in CTest. Exceptions fail the run.
The executable rejects other arguments. Use a Release build for cost comparisons.
## Method and host
Measured 2026-09-26 on Windows, AMD Ryzen 7 9800X3D, 8 cores/16 logical processors,
MSVC 19.44 x64, CMake Release /O2, single simulation thread, no CPU affinity or
frequency lock. The host is a workstation, not the R730xd. Raw data:
[resolution-budget-msvc1944.jsonl](results/resolution-budget-msvc1944.jsonl).
Base checkout: 55a728a plus the M4.6 working changes. Fixture version 2 identifies
the shared configurable model, not a persistent simulation version.
100 founders, 10,000 ticks, capacity 200; no randomness. Five age widths and energy
supply 240/300/360; only supply varies between scarcity scenarios. Four identical
matter channels remain abundant. A warmup and three fresh measured repetitions
per representation/scenario alternate timing order. Timings include step allocation,
sorting and peak-counter observation, exclude setup/copy, logging, conservation
checks, cohort validation and collapse/advance/materialize transactions. Terminal
censuses are output so the compiler cannot discard the computation.
Peak payload is vector size times row sizeof; peak capacity includes spare vector
storage. Neither includes allocator metadata, per-step eligible-index scratch,
object/stack bytes or RSS. The individual fixture retains every dead row and scans
it, whereas aggregate empty classes are erased. This inflates its cost relative to
a future compact active ECS. These figures compare these fixtures only.
## Baseline costs (supply 300)
Median elapsed milliseconds for all 10,000 ticks. Extinct runs still execute the
remaining empty steps, so their low cost must not be treated as useful throughput.
| Width | Individual ms | Aggregate ms | Peak individual capacity bytes | Peak aggregate capacity bytes | Aggregate survives |
|---|---:|---:|---:|---:|---|
| 10 | 27.107 | 0.250 | 122528 | 360 | yes |
| 25 | 27.256 | 0.054 | 122528 | 160 | no |
| 50 | 29.190 | 0.135 | 122528 | 120 | yes |
| 100 | 27.180 | 0.045 | 122528 | 80 | no |
| 200 | 27.263 | 0.045 | 122528 | 80 | no |
## Sensitivity
Errors are integer parts per thousand: abs(a-b)/max(a,b), rounded down; zero when
both values are zero. Matter is per channel; all four channels are identical here.
The proposed gate requires no false extinction at any tick and final population,
energy and matter error each <=100 ppt in **every** scenario. This threshold was
written before the run; it is a research recommendation, not a biological law.
Maximum errors over the trajectory are also recorded in the raw data and are often
much worse than the terminal errors. No candidate passes even the terminal gate.
| Width | Supply | Individual/aggregate survivors | Population error | Energy error | Matter error | False extinction |
|---|---|---|---:|---:|---:|---|
| 10 | 240 | 108/120 | 100 | 127 | 346 | no |
| 10 | 300 | 75/100 | 250 | 24 | 724 | no |
| 10 | 360 | 130/130 | 0 | 34 | 531 | no |
| 25 | 240 | 108/100 | 74 | 272 | 189 | no |
| 25 | 300 | 75/0 | 1000 | 1000 | 1000 | yes |
| 25 | 360 | 130/130 | 0 | 151 | 657 | no |
| 50 | 240 | 108/0 | 1000 | 1000 | 1000 | yes |
| 50 | 300 | 75/120 | 375 | 43 | 35 | no |
| 50 | 360 | 130/180 | 277 | 217 | 589 | no |
| 100 | 240 | 108/0 | 1000 | 1000 | 1000 | yes |
| 100 | 300 | 75/0 | 1000 | 1000 | 1000 | yes |
| 100 | 360 | 130/0 | 1000 | 1000 | 1000 | yes |
| 200 | 240 | 108/0 | 1000 | 1000 | 1000 | yes |
| 200 | 300 | 75/0 | 1000 | 1000 | 1000 | yes |
| 200 | 360 | 130/0 | 1000 | 1000 | 1000 | yes |
The shared fixture reproduces M4.5's baseline survivor results. Width 10 avoids
extinction but has large matter errors; width 50 fails under reduced energy supply.
Widths 100 and 200 go extinct in every scenario. Conservation passed at every tick
for both representations in all 15 scenarios; that does not imply ecological fidelity.
## Limits and next steps
This is one genome and one deterministic founding distribution. No ensemble,
spatial interactions, oxygen, predation, network load or production threading is
represented. There is no R730xd capacity estimate or Kubernetes resource request.
CPU measurements are wall time on an uncontrolled host and are reports, not CI gates.
Before adopting a tier model: preserve within-class reserve variation, test multiple
genomes and founding distributions, measure trajectory/distribution errors, and
compare an individual store with dead rows retired. Measure conversions, catch-up
and complete region workloads on Linux and the target host. Calibrate acceptable
error budgets with biological/playtest requirements. M5 can independently begin its
small sparse-voxel prototype; no terrain dependency is added to this fixture.