Files
aevolution/docs/adr/0035-genome-ecology-ensembles.md
2026-10-08 12:08:53 -04:00

9.9 KiB

[Reviewed]

ADR-0035: Genome-derived strategies over an environment gradient

Status

Adopted for M3.6, which closes M3. Nothing mutates during an episode: autonomous divergence would settle NPC mutation, which OPEN_QUESTIONS leaves open until M12. No simulation contract changes, and ADR-0029's episode driver is untouched.

Context

Every piece of M3 exists and none of them meet. Genomes validate, phenotypes derive, budgets are charged, candidates are drawn and lineages are retained, but the M2 fixtures still run on hand-written constants. M3.6 is where a genome finally decides how something behaves in an ecosystem, measured across a gradient rather than at one point.

The named risk is tuning one seed until it looks good. A single run of a seeded ecosystem says almost nothing: the question is whether a difference between strategies survives being measured again on seeds nobody chose.

Decision

Derive a StrategyConfig from a genome rather than writing one. Intake, assimilation and reproduction investment all come from ADR-0032's budget, so a strategy is what the genome already implies and not a second description of the same organism. A parent reproduces once it holds what it would invest, which avoids inventing a separate margin.

Run those strategies over a grid of resource levels and temperatures, many seeds per cell, on ADR-0029's existing episode driver, whose behaviour for every configuration it already accepted is unchanged, so the M2.6 fingerprint still holds and the ecology under test is the one already validated.

One guard in that driver had to be amended rather than worked around. It refused a strategy with nothing to invest, which is exactly what a genome without the reproduction capability derives, so a draw that happened to produce one failed the whole run. Filtering such variants out before sampling would have been worse than the failure: the measured set would then be the fertile subset, chosen by a guard, and a strategy's inability to reproduce is a result rather than a reason to exclude it. The driver now runs such a strategy and proposes no birth for it. ADR-0029 is amended accordingly.

The two axes reach a run by different routes, and implementation showed the routes are not symmetric. Founders are derived fully grown, so their construction transition is empty and ADR-0031's per-channel available cannot move intake, assimilation or reproduction investment. The resource level therefore does not enter strategy derivation at all; it reaches a run only as the nutrient pool an episode starts with, which bounds producer growth. Temperature is the opposite: it enters only the derivation, through thermal shortfall, and reaches the ecology through the scheduling rule below. Saying the resource level shapes a strategy would have been false, so this ADR says instead exactly where each axis acts.

Temperature has to be made consequential or the axis is decoration: ADR-0031 derives thermal shortfall but deliberately lets it change no budget. The harness therefore shortens a strategy's scheduled lifespan in proportion to its shortfall. That is an extension of the lifespan rule ADR-0029 already owns, which is a fixture scheduling device and explicitly not biology. It is not a mortality model, it lives in the harness rather than in any contract, and a genome carrying thermal tolerance simply keeps more of its scheduled life at the extremes.

Founders are pre-diverged variants of one ancestor, produced by ADR-0033's candidate draw before any episode starts and admitted to ADR-0034's archive with their parent links. Every organism's birth and death is recorded as it happens, so the diversity a run reports comes from retained records rather than from counters the harness kept for itself. Nothing mutates mid-episode, because an organism diverging on its own is exactly the NPC-mutation question that remains open.

An organism alive when an episode stops was not observed to die. Observation ended, which is a fact about the run and not about the organism, so no death is recorded for it: ADR-0034 makes a recorded death terminal, and there is nothing here to be terminal about. Such an organism stays living in the archive and is reported as a survivor whose watched time is a lower bound on a lifetime this run never saw the end of. Completed lifetimes and censored observations are summed separately and never added together, because a mean over both would understate every survivor.

The report is distributional. Per strategy and per cell it carries episodes, births, survivals, lifetime and mean thermal shortfall, so a reader sees a spread rather than an outcome. To make overfitting visible rather than merely warned against, the seeds of every cell are split in half and the two halves reported separately: a difference that appears in one half and not the other is an artifact of the seeds, and the tests assert the halves agree.

Alternatives Considered

Mutating during an episode would exercise more of M3 and would settle NPC mutation by implication, which ADR-0034 was careful not to do. Teaching the episode driver about genomes would change a tested fixture and break a pinned fingerprint for no gain, since a derived StrategyConfig is the same shape the driver already takes. Letting thermal shortfall reduce a budget would put a mortality rule inside a contract that says it is not one. Reporting one number per cell would hide exactly the variance the named risk is about. Treating the end of an episode as a death would have made every reported lifetime a mixture of two different facts and would have written terminal records the archive cannot retract. Excluding non-reproducing variants before sampling would have kept the driver untouched at the cost of measuring only the strategies that a guard already approved of. Deriving lifespan from body size would be an allometric claim this project has no basis for; it stays a harness constant applied to every strategy alike.

Benefits

A genome now determines behaviour end to end, through code that was already tested at every step. Both axes are consequential, so strategies can be compared rather than merely listed. Split halves make an overfitted conclusion fail a test instead of surviving in a report, and the archive gives the diversity figures a provenance.

Costs

The grid is small and the seed count modest, so the report describes this fixture and nothing wider. The thermal penalty, the lifespan base and the grid bounds are chosen constants. Recording every birth and death costs a scan of the slots each tick. Lifetimes are censored: an episode that ends with survivors reports a lower bound for each of them, so lived_ticks and observed_ticks answer different questions and a reader has to choose which one it is asking. Across a run the archive therefore ends holding living carriers, which is accurate about what was observed and is not a claim that anything persists. Founders diverge once, before the run, so the ensemble compares a fixed set of strategies rather than an evolving population.

Risks

Do not read a cell's numbers as fitness, adaptation or balance: a strategy that produces more offspring here is doing so under constants this ADR chose. The thermal penalty is a scheduling rule and is not a claim that heat kills.

The resource axis is emphatically not monotonic, and no test asserts that it is. Sweeping the fixture, organisms rise to a peak near 8000 femtograms and fall away on both sides. The mechanism is a property of ADR-0026's feeding rules, not of ecology: a target may be bitten by only one consumer per tick, so the community's income is capped at one producer bite however much is in the pool. Scarcity empties the producer, consumers then bite each other, and that concentrates mass into individuals that reach the reproduction threshold; abundance spreads the same income across more feeders, and none of them accumulates enough to fund a birth. Reading "more resource supports more life" out of a cell would be reading past the rules that produced it.

The fixture's reproduction investment sits at the genome schema's floor of one per thousand. That is forced, not chosen for taste: intake scales with surface and investment with body mass, so the ticks needed to fund a birth grow as the cube root of mass and exceed any plausible lifespan at every larger fraction. At the previous constants no organism could earn a birth at all, every birth came from the opening endowment, and both axes measured nothing. A cost model in which reproduction is affordable only at the schema floor is a finding about ADR-0031 and ADR-0032 worth revisiting, not a fixture detail. Agreement between seed halves shows a measurement is reproducible, not that it is ecologically meaningful, and conservation holding across the grid remains arithmetic rather than plausibility. No organism mutates, no population is modelled, and nothing here is persistent.

Consequences

Test that a derived strategy matches the genome's own budget, that conservation holds at every tick of every cell, that the archive ends holding exactly the founders and every organism that lived, that lineage status reflects the carriers, that ending an episode records no death and leaves survivors living in the archive, that a strategy which cannot reproduce is measured rather than refused, that a half's thermal shortfall is a total across its episodes rather than one reading, that the two seed halves agree within a stated band for every strategy and cell, that each axis alone changes outcomes rather than leaving cells identical, that a thermally tolerant strategy keeps more of its life at the extremes than an intolerant one, and that a repeated run reproduces the report exactly. The gradient test asserts only that an axis moves the outcome, never a direction, because the direction is not a property of this model. evo-sim --ensemble reports every cell with its two halves separate and never merges them. No genome schema, mutation catalog, archive, ledger, episode driver, renderer or deployment changes.