Purpose
A model-breeding system needs clear records. A Genome records how a descendant was created. A FitnessVector records how that descendant performed. Keeping those records separate makes the ecology explainable, reproducible, and easier to improve.
Genome schema
STRUCT Genome
genome_id: stable identifier
base_model_id: immutable parent or champion identifier
base_family: architecture and tokenizer family
compatible_tensor_schema: checksum of tensor names, shapes, and dtypes
delta_adapters: list of AdapterSpec
merge_parents: list of parent genome identifiers
operator: adapter_merge | model_soup | distillation | quantization | pruning | router_specialization
random_seed: deterministic variation seed
source_provenance: checksum or release packet reference
mutation_log: human-readable operator trace
artifact_uri: registry path or local package path
created_at_utc: timestamp
END STRUCTAdapter specification
STRUCT AdapterSpec
adapter_id: stable identifier
adapter_type: lora | sparse_delta | low_rank_delta | representation_intervention
rank: integer
density: decimal
scale: decimal
target_layers: list of layer selectors
compatibility_digest: checksum of expected base identity
END STRUCTFitnessVector schema
STRUCT FitnessVector
genome_id: reference to Genome
evaluation_scope: named benchmark, workflow, or niche
utility: task success, accuracy, acceptance, or business outcome
calibration: confidence quality and abstention behavior
latency_ms: measured response time
memory_mb: measured working set
energy_units: estimated or measured energy
novelty_score: distance from existing population behavior
reuse_score: likelihood that the descendant becomes a useful stepping stone
human_capability_gain: measured improvement in user or reviewer capability
maintenance_burden: expected care cost
composite_score: benefit minus cost after weights
evidence_packet: release packet or eval sidecar path
END STRUCTCompatibility checkpoint
Weight-space breeding only works when the parent artifacts line up. If the base family, tokenizer, tensor layout, or adapter target map differs, use distillation or feature-level transfer instead of direct parameter mixing.
FUNCTION choose_breeding_operator(parents)
IF ALL_SAME(parents.base_family)
AND ALL_SAME(parents.tokenizer_digest)
AND ALL_SAME(parents.tensor_schema_digest)
RETURN PARAMETER_LEVEL_MERGE
END IF
IF SHARED_INPUT_OUTPUT_CONTRACT(parents)
RETURN DISTILLATION_OR_ROUTER_SPECIALIZATION
END IF
RETURN NO_OP_WITH_SCOPE_NOTE
END FUNCTIONNovelty archive record
A novelty archive prevents the ecology from collapsing onto one overused champion.
STRUCT NoveltyRecord
genome_id: reference to Genome
niche_id: capability niche
behavior_sketch: embedding, top-k distribution sketch, or eval outcome fingerprint
nearest_neighbor_distance: decimal
novelty_score: decimal
retained_reason: champion | specialist | diverse_challenger | teaching_example
END STRUCTSource 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.
Implementation directivesModelBreeder Site and Theory: Controlled Evolution and Positive ImplementationImplementation directive · 25.7 KB
Implementation directivesControlled Evolution of AI Models: Site Copy, Schemas, and Breeding LoopImplementation directive · 25.7 KB
Browser and edge runtimeArchitectural Advancements for Zero-Dependency In-Browser Large Language ModelsEmerging implementation architecture · 54.0 KB