Files
aevolution/docs/engineering/M12_EXACT_CELLULAR_COHORTS.md
dingo 2b074b9ae1
Foundation / checks (deploy) (push) Successful in 5s
Foundation / checks (ecs) (push) Successful in 2m19s
Foundation / postgres-tools (push) Canceled after 0s
Foundation / server-image (push) Canceled after 0s
Foundation / image (push) Canceled after 0s
Foundation / checks (tsan) (push) Canceled after 6m24s
Foundation / checks (network) (push) Canceled after 6m33s
Foundation / checks (postgres) (push) Canceled after 6m34s
Foundation / checks (checks) (push) Canceled after 6m44s
Foundation / checks (client) (push) Canceled after 6m44s
M12.2: select and qualify exact cellular local cohorts
2026-10-17 12:04:39 -04:00

58 lines
3.1 KiB
Markdown

# M12.2d exact cellular cohort qualification
Decision: [ADR-0139](../adr/0139-exact-cellular-local-cohorts.md), accepted under
the maintainer's delegated M12.2 decision authority. The rules-8 model retains
every member exactly. Regional approximation remains disabled. No compression,
speedup, deployment or production capacity is claimed.
## Executed evidence, 2026-10-17
Full Release runs on MSVC 19.44 and GCC 14.2 use the current reference's four
seeds (0/1 development, 2/3 held out), rich/scarce finite deposits and conversion
oscillation widths 1/16/64. Each of 24 cases compares complete canonical
biological snapshots at every one of 8192 ticks: **196,608 paired ticks** per
platform, with zero state error. All non-timing paired JSON fields agree across
compilers. The original reference's setup, rules and goldens are unchanged.
Across these repetitions: **2354 collapsed members, 2354 materialized members,
6380 cohort member-ticks, 12 offspring, 48 applied mutations and 57 deaths**.
Counts include repeated reference scenarios at three conversion widths, not
independent biological replicates. Some rich seeds produce no offspring; the
test requires actual conversion coverage in seed 1 and across the full matrix.
The short CTest also runs that long seed-1 case, preventing a vacuous pass with
no cohorts. Controlled founder collapse, missing intentions, duplicate collapse
and repeated materialization refusals preserve state. Per-tick checks validate
tags against live unowned active cells and preserve total inventory.
Raw output: [MSVC](results/m12-2d-cohorts-msvc1944.jsonl),
[GCC](results/m12-2d-cohorts-gcc14.jsonl).
| Living cells | MSVC p95 step (ms) | GCC/WSL p95 step (ms) | Snapshot bytes |
|---|---:|---:|---:|
| 16 | 0.045 | 0.058 | 21,108 |
| 64 | 0.274 | 0.200 | 71,076 |
| 128 | 0.773 | 0.777 | 137,700 |
| 256 | 2.359 | 2.326 | 270,948 |
These are 32 local warm stepping observations per size, fully funded controlled
founders and finite contact food. They include the complete biological owner
step but exclude network, storage/publication, rendering, initialization and
the qualification's byte comparisons. They are not a mature wildlife stress
workload, long-term tail distribution or the Linux deployment host. Refusing a
257th admission leaves the existing 256 alive and the snapshot unchanged.
The hard ceiling is retained; production headroom remains unqualified.
## Limits and next work
Local exact-state fidelity passes the declared matrix; useful aggregation and
target-host fallback capacity remain open parts of M12.2. The largest observed
cohort in these ecological histories is small. Full predation, twin-bud,
multi-member cohort, capacity wildlife and blocked-resume matrices need stronger
coverage before production enablement. No arbitrary approximation may fill that
gap. Bins do not exist; widths change conversion timing, not biological quantum.
M12.3 can implement the exact owner's durable conversion contract without
enabling a live selector. Database publication, process loss and deployment
qualification must be reported separately from codec/replay success. No World0
state is accessed.