Files
aevolution/docs/adr/0138-cellular-population-reference-instrument.md
dingo bee57e5829
Foundation / checks (deploy) (push) Successful in 5s
Foundation / checks (ecs) (push) Successful in 2m47s
Foundation / checks (tsan) (push) Successful in 13m32s
Foundation / checks (client) (push) Failing after 17m9s
Foundation / checks (network) (push) Successful in 19m58s
Foundation / postgres-tools (push) Canceled after 0s
Foundation / server-image (push) Canceled after 0s
Foundation / image (push) Canceled after 0s
Foundation / checks (postgres) (push) Canceled after 20m34s
Foundation / checks (checks) (push) Canceled after 20m49s
M12.2c: qualify the current cellular reference instrument
2026-10-17 11:44:00 -04:00

81 lines
4.1 KiB
Markdown

# ADR-0138: Current cellular population reference instrument
## Status
Accepted for isolated M12.2c qualification tooling (2026-10-17). No runtime
aggregation, biological rule, public API, persisted version or golden changes.
M11.9 is closed for now by the maintainer with authorization to continue M12.
## Context
M12.2b repairs a synthetic averaging failure but cannot qualify spatial cellular
ecology. The current implemented baseline is rules 8: schema/development 4,
funded skilled/twin budding, hunting wildlife and finite contact-fed deposits.
ADR-0135 colonies and ADR-0136 infall/photosynthesis remain proposals. A later
population model must be compared against these actual supported cellular rules.
## Decision
Add a headless executable using the existing CellularLives owner as the exact
active reference. It generates player fixture intentions and records accepted
wildlife intentions once. Compare canonical batch encode/decode/replay in parallel
at 1/16/64-tick partitions, all at the existing 50 ms quantum. Partition 1 compares
full canonical state every tick; the other partitions compare at every batch
boundary. Restore checkpoints at completed boundaries and keep all identities,
inventory, archive, mutation eligibility and presence state unchanged.
Declare the instrument before measuring:
- Seeds 0/1 are development and 2/3 held out. Each uses rich and scarce finite
deposits, four equally funded player-origin founders and the rules-8 owner.
- The CTest suite uses 512 ticks per case. The full research run uses 8192;
report observed births, mutations, deaths and suspension rather than assuming
every regime exercises all paths. Report zero-coverage limits explicitly.
- Player fixtures first feed in place, then steer/ask to bud at declared intervals;
earned mutation selections use only the real offer and funded transitions.
Offline intervals freeze controlled lives, while wildlife continues through
the shared intention interface. This is an experiment driver, not a new AI rule.
- Zero tolerance for canonical state, conserved channel totals, demographic or
genome-identity divergence at compared boundaries. No false extinction inferred
from empty local observations. Distinct genome IDs and distinct capability
profiles are reported separately; neither is a species count.
- Record active population, snapshot bytes and local stepping/replay costs, with
verification, codec/restore and biology costs distinguished. These are
workstation observations, not accepted target-host fallback budgets.
No adapter feeds current cellular state into the synthetic strata record. Its
missing spatial/reproductive/deposit state makes such an admission unsupported.
This instrument is the exact control for later paired active/cohort qualification;
passing replay does not establish ecological approximation fidelity or scaling.
## Alternatives considered
Extending the old synthetic fixture with cellular labels would leave its biology
unrepresentative. Treating checkpoint replay as proof of aggregation would confuse
state closure with fidelity. Enabling a runtime selector before either a supported
model or target-host budgets exists remains prohibited by ADR-0041.
## Benefits
The reference follows the current ratified cellular behavior, exercises archive
and offline closure, and is independent of renderer, network, database and clocks
inside the owner. It provides reproducible inputs and observations for future work.
## Costs
Canonical checks, cloning and replay deliberately add measurement overhead.
Short CTests do not establish long unattended ecology or all demographic paths.
## Risks
The driver and owner are deterministic but the reference is not an independently
derived biological model. No approximate cohort, migration, global extinction or
production cost envelope is supplied. Changes to trial rules require a separately
identified reference baseline and preserved earlier evidence.
## Consequences
Run NDEBUG tests, clang-format/tidy and Linux ASan/UBSan. Publish measured coverage
and limits. Keep M12.2 cellular/cohort and target-host qualification open until
actual paired models meet accepted bounds; no automatic tier transition is enabled.