Files
aevolution/docs/adr/0028-funded-reproduction.md
2026-10-08 12:08:53 -04:00

7.4 KiB

[Reviewed]

ADR-0028: Funded reproduction and retained lineage identity

Status

Adopted for the M2.5 synthetic community fixture. No genome content, mutation, player account, control grant, spatial placement or gestation model is selected. Lineage identity here is an opaque parent reference, not the M3.1 genome record.

Context

M2.1-M2.4 conserve inventory through capture, growth, feeding, terminal events and bounded decomposition. No operation creates an actor, so every community decays monotonically and M2.6's long-run ecology fixture cannot observe anything but extinction. Reproduction is the missing transfer in ENERGY_MODEL's table.

Offspring must be funded. A child credited from nothing would break conservation; a child credited by the host would make founder admission implicit. ADR-0024's ledger fixes its reservoir set at construction and exposes no append operation, so actor creation cannot allocate storage mid-run without reopening that contract.

Decision

Extend FeedingCommunity::advance with a fourth bounded span of immutable ReproduceIntent{sequence, parent, investment} records. Like FeedIntent these are trusted co-located fixture intentions, not client commands or autonomous decisions. Admit at most max_consumers per tick, require nonzero unique host sequences, sort by sequence and permit at most one intent per parent per tick.

ReproductionConfig{capacity, construction_per_thousand, minimum_endowment} is fixed at create. Capacity reserves that many additional consumer reservoirs after the configured consumers and before waste. Configured consumers plus capacity must not exceed max_consumers. Capacity defaults to zero, so existing fixtures keep an identical reservoir layout, waste index, action sequence and golden results. Reproduction is intent-driven; a tick without intents is a M2.4 tick.

Process state becomes unborn, active or corpse. Reserved slots begin unborn: they hold an empty reservoir, never act, are never valid feeding targets, are not corpses and are not decomposed. active(id) is false for unborn and corpse alike; corpse(id) and the ancestry queries reject a reserved slot with a distinct unborn_actor error rather than reporting an empty corpse. Reproduction allocates the lowest-numbered unborn slot. A slot never returns to unborn, so IDs are never reused and a dead child's identity persists exactly as ADR-0027 requires.

The intent names the withdrawal from the parent reservoir, mirroring feeding's bite semantics. A fixed construction_per_thousand fraction splits it with the same overflow-safe quotient decomposition: the quotient endows the child, the energy remainder goes to heat and the matter remainder goes to waste. Construction overhead is real inventory, not rounding loss. A value of 1000 endows everything; 0 sends the whole investment to heat and waste and therefore always fails the endowment floor unless that floor is zero. No fractional entitlement is carried between reproductions.

Reject the whole tick when the parent is not an active consumer, when the investment is zero or exceeds the parent's balance, when no unborn slot remains, or when the resulting endowment does not meet minimum_endowment on every channel. Capacity exhaustion is backpressure: it rejects, it never grants a free or reduced offspring. The producer cannot reproduce in this increment; capture is its funding path and conflating the two would make an abiotic source a parent.

A child duplicates its parent's consumer configuration verbatim. With no genome, offspring differ from their parents only in inventory. This is duplication, not heredity, and M3 replaces it with developmental capabilities.

Record an immutable parent and origin tick per process. Founders record parent zero and origin zero. Both survive the death of the child and of the parent, and remain readable after decomposition empties the reservoir.

Tick phases: terminal facts -> producer activity -> ordered feeding -> reproduction -> decomposition -> atomic publish. Reproduction after feeding lets a consumer fund offspring with what it ingested this tick; this is a deliberate fixture choice, not a biological claim about digestion latency. Terminal facts run first, so a process terminated this tick cannot reproduce. Intents validate parents against the state entering the reproduction phase, so a child born this tick cannot itself be a parent, feed, be fed upon or be decomposed until the next tick.

All phases run on the existing private candidate. Any rejected intent rolls back reproduction, feeding, producer progress, decomposition and diagnostics together. Receipts report the sequence, allocated child ID, endowment and overhead, and are published by noexcept move with the rest of the committed state.

Alternatives Considered

Appending reservoirs on birth would reopen ADR-0024's fixed opening totals and shift the waste index mid-run, invalidating inspection spans and existing goldens for no benefit at fixture scale. Binary fission would force a decision on parent continuity and control identity that REPRODUCTION_MODEL explicitly leaves open. Charging overhead on top of the investment needs a second bounds check and can overflow a parent that could otherwise afford the child. Eggs, cohorts and broadcast spores require gestation timers and statistical allocation that this increment has no fixture for. Reusing dead IDs would save slots and destroy ancestry. Host-credited offspring would break conservation outright.

Benefits

The energy cycle closes: nutrients fund growth, growth funds feeding, feeding funds offspring, death and decomposition return matter. Offspring are provably paid for, capacity is explicit and bounded, ancestry outlives its organisms, and M2.6 can run a seeded long-run ecosystem that is not a guaranteed decay.

Costs

A fixed pre-allocated slot budget is not a population model; a community cannot outgrow its configured capacity and has no aggregation path until M4. Lowest-slot allocation is deterministic but arbitrary. Whole-community copies per tick remain the correctness baseline. Ancestry is a parent reference and an origin tick, not heredity: nothing is inherited, because nothing heritable exists before M3.

Risks

Do not read this as a reproductive strategy, a fecundity rule or a funded founder admission for a persistent world. Construction overhead, the endowment floor and the post-feeding phase position are synthetic parameters that must be reviewed before any biological tuning, and must not become gameplay by being measured. No account owns a process; a parent reference grants no control and no wildlife ownership. Capacity rejection must never be softened into a reduced endowment.

Consequences

Test hand-calculated endowment and overhead splits, the endowment floor, per channel insufficiency with abundant energy, capacity exhaustion, unborn slots rejected as targets and as corpses, terminated and unknown parents, duplicate and zero sequences, more than one intent per parent per tick, whole-tick rollback including producer progress and decomposition, order permutation equality across seeds, absent ID reuse, ancestry surviving both deaths, construction endpoints, u64 extremes and independently recomputed conservation. Default zero capacity must leave every M2.4 result unchanged. A later process snapshot must serialize states, parents, origins and capacity together; M1's codec serializes none of it. No database, renderer, network or deployment change follows from this record.