7.2 KiB
[Reviewed]
ADR-0029: Deterministic ecosystem driver and bounded episodes
Status
Adopted for the M2.6 headless measurement harness, and amended in M3.6 so a strategy that does not reproduce is run rather than refused (see the founder strategy paragraph below). No other part of the driver changed. No gameplay AI, perception, behavior model, senescence biology, metabolic rate or population abstraction is selected. The simulation contracts of ADR-0024 through ADR-0028 are unchanged.
Context
M2.2 through M2.5 close the transfer cycle, but every intent is supplied by the caller. Nothing decides what happens, so a long run needs a decision source before extinction or diversity can be measured at all.
Two structural limits surfaced while scoping this increment. First, ADR-0028
retains identities permanently and caps slots at max_consumers, so one community
supports a bounded number of consumer lifetimes and is not a population: a single
million-tick community would spend almost all of those ticks after slot exhaustion
in a static state. Second, no maintenance cost exists, so a consumer that is never
eaten never dies and a run has no endogenous end. ENERGY_MODEL leaves baseline
metabolism and starvation rules open, so inventing them here would settle an open
biological question inside a measurement harness.
Decision
Add EcosystemEpisode, which privately owns one FeedingCommunity and one seeded
Random stream. It holds no inventory, performs no transfers and calls no public
mutator other than advance. It reads committed public state and emits intents.
Measure many bounded episodes rather than one long community. Each episode ends at extinction, defined here as no living consumer, or at a configured tick limit. Because births consume permanent slots and every consumer has a finite lifespan, extinction is reachable in every episode. The long-run figure is the total ticks advanced across seeded episodes, and each episode reports its own outcome.
Mortality has exactly two sources, both already contracted. Predation: the driver bites a target's entire remaining balance when that fits the intake cap, which ADR-0026 already turns into terminal depletion. Lifespan: the driver emits an ADR-0027 terminal fact once a consumer's age reaches its configured lifespan. Lifespan is a host scheduling rule that bounds the episode, not senescence, aging biology or a death cause. No maintenance, starvation or metabolic cost is added.
A founder strategy fixes opening inventory, intake cap, assimilation, an investment
bundle, a reproduction threshold and a lifespan. A strategy with nothing to invest
does not reproduce, and the driver runs it rather than refusing it: the empty
bundle would otherwise clear both reproduction bounds trivially and request a birth
transferring nothing, so the driver proposes no birth for it at all. Amended for
M3.6, which derives strategies from genomes: a mutation may legitimately drop the
reproduction capability, and refusing such a strategy here would have made the
strategies a run can measure a subset chosen by this guard rather than by the draw.
Only configurations this ADR previously rejected outright behave differently, so
the M2.6 fingerprint 2cd4a15f9fbe9949 is unchanged. ADR-0028 makes a child duplicate
its parent, so a founder strategy identifies one lineage for the whole episode and
competing strategies are the only source of variation. Lineage membership is
recovered by walking retained parent references to a founder.
Per tick the driver emits, in this order: terminal facts for every active consumer
that has reached its lifespan; then feed intents; then reproduce intents. It must
never submit a batch the community rejects, so it applies three self-restrictions.
A consumer terminated this tick neither feeds nor is fed upon nor reproduces. Each
target is claimed by at most one feeder per tick, assigned in ascending consumer
ID, which removes contention rejection. A consumer that is a feeding target this
tick does not reproduce this tick, which removes the interaction between a reduced
balance and the later reproduction check. Reproduce intents are truncated to the
number of free slots in ascending parent ID. Bites are per-channel
min(intake cap, target balance) and targets with no inventory are skipped.
A rejected advance is therefore a driver defect, not an ecological outcome, and
is reported as community_failure rather than being retried or absorbed. Target
selection draws from the seeded stream; nothing else consults it.
Each episode reports seed, ticks advanced, births, predation and lifespan deaths, extinction tick, founder and surviving lineage counts, cumulative decomposition, conserved totals and a state fingerprint. Conserved totals must be identical at every tick. The report is data for review, not a pass threshold.
Alternatives Considered
A maintenance cost with a starvation rule would give endogenous mortality, but
ENERGY_MODEL leaves metabolism and starvation explicitly open and a measurement
harness is the wrong place to settle them. Reusing retired slots would destroy the
ancestry ADR-0028 just established. Raising max_consumers delays exhaustion
without removing it and would widen per-tick intent bounds with no evidence. One
continuous million-tick community measures a static post-exhaustion state. Letting
the driver retry a rejected batch would hide driver defects as ecology.
Benefits
The full cycle runs unattended and deterministically: nutrients fund growth, growth funds feeding, feeding funds offspring, lifespan and predation produce corpses, and decomposition returns matter. Extinction and lineage survival become measurable across seeds, conservation is checked every tick over long horizons, and the driver is separable from the contracts it exercises.
Costs
Episode length is bounded by slots times lifespan, so this measures many short histories rather than one long one. Strategy constants, lifespans and the targeting rule are chosen, not derived. One-feeder-per-target and the reproduction exclusion are scheduling conveniences that bias contention away from the ordering ADR-0026 deliberately exposes. Whole-community copies per tick remain the baseline.
Risks
Do not read episode statistics as population dynamics, life-history theory or balance. Lifespan is not senescence, targeting is not perception, and a surviving lineage count is not speciation or fitness. Conservation holding across a million ticks demonstrates arithmetic, not ecological plausibility, and the explicit risk of this increment is passing every invariant while modelling nothing interesting. Strategy constants must not become gameplay by having been measured.
Consequences
Test per-tick conservation and nonnegative balances across many seeds, identical reports for one seed on repeated runs, independence of episodes, extinction reached and recorded, lineage attribution through retained ancestry, both mortality paths, slot exhaustion, tick-limit termination, a single-founder episode, an episode whose founders cannot reproduce, and that the driver never provokes a rejected batch. Cross-platform fingerprints must agree. No snapshot format, database, renderer, deployment or simulation contract changes. Serialized recovery of episode state is not provided; M4 aggregation remains the path to real populations.