Why novelty matters
A pure leaderboard tends to erase useful specialists. Model breeding works better when the population keeps different kinds of usefulness: the current champion, niche experts, low-cost helpers, and surprising challengers that may become stepping stones later.
The positive goal is quality plus useful diversity. That means the ecology should preserve enough variation to keep learning while still preferring models that prove practical value.
Selection recipe
PROCEDURE select_next_population(population, candidates, niches, capacity)
evaluated <- []
FOR candidate IN candidates
fitness <- RUN_EVALUATION_SUITE(candidate)
novelty <- MEASURE_NOVELTY(candidate, population)
record <- FitnessVector(candidate.id, fitness, novelty)
evaluated.ADD(record)
END FOR
champions <- BEST_PER_NICHE(evaluated, niches)
specialists <- HIGH_VALUE_LOW_COST(evaluated)
challengers <- MOST_NOVEL_ABOVE_MINIMUM_FITNESS(evaluated)
next_population <- UNION(champions, specialists, challengers)
next_population <- TRIM_TO_CAPACITY(next_population, capacity)
IF next_population DOES_NOT_IMPROVE_OR_DIVERSIFY(population)
RETURN NO_OP_WITH_ARCHIVED_EVIDENCE(population, evaluated)
END IF
RETURN next_population
END PROCEDUREFitness terms
| Term | Positive meaning |
|---|---|
| Task utility | The descendant does useful work. |
| Human capability gain | People learn, decide, review, or create better with the ecology. |
| Local privacy | More work can happen on controlled hardware. |
| Efficiency | The same outcome costs less latency, memory, energy, or maintenance. |
| Diversity | The population covers more useful niches. |
| Reuse | The artifact can become a parent, adapter, or teaching example. |
Multi-parent merge pattern
FUNCTION multi_parent_adapter_merge(parents, weights, target_rank)
REQUIRE SUM(weights) == 1
REQUIRE COMPATIBLE_BASE_FAMILY(parents)
merged_delta <- ZERO_DELTA()
FOR parent, weight IN ZIP(parents, weights)
merged_delta <- merged_delta + weight * parent.adapter_delta
END FOR
compressed <- PROJECT_TO_RANK(merged_delta, target_rank)
child <- CREATE_DESCENDANT(base: parents[0].base_model, adapter: compressed)
child.evidence <- EVALUATE(child)
RETURN child
END FUNCTIONPopulation dashboard fields
A useful dashboard should show: genome id, niche, parentage, operator, utility, novelty score, resource profile, lifecycle state, release packet, and retained reason.
VIEW PopulationDashboard
COLUMNS genome_id, niche, lifecycle_state, utility, novelty, latency_ms, memory_mb, retained_reason
FILTERS champion, specialist, challenger, archived, no_op
ACTIONS inspect_genome, inspect_fitness_vector, compare_with_parent, build_release_packet
END VIEWSource reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.