The engineering meaning of inheritance
In model breeding, inheritance is not a mystical transfer of essence. It is the preservation of traceable properties from parent artifacts to descendant artifacts. Those properties may include base architecture, tokenizer, adapters, training data lineage, evaluation history, safety restrictions, runtime requirements, and known failure modes.
Inheritance channels
| Channel | Inherited property | Example |
|---|---|---|
| Parameter lineage | weights, adapters, quantization state | a LoRA descendant from a shared base |
| Data lineage | examples, teacher outputs, labels, filters | a distilled specialist from a larger teacher |
| Contract lineage | input and output schemas | a replacement summarizer that preserves JSON output |
| Evaluation lineage | benchmark suites and thresholds | a candidate compared against the same release gate |
| Operational lineage | owners, rollback, deployment tier | a canary descendant with the same rollback target |
Variation operators
Variation should be declared before execution. This makes the candidate reproducible and keeps selection honest.
RECORD VariationOperator
name
operator_version
allowed_parent_types
allowed_parameter_budget
allowed_data_sources
sandbox_profile
expected_outputs
risk_tier
END RECORDParentage depth
A descendant can have one parent, multiple parents, or a non-model seed origin. The more parents it has, the more provenance and compatibility checks it needs. Multi-parent merging and adapter fusion can be valuable, but they also increase the chance of license conflicts, tokenizer mismatch, and destructive interference.
Trait records
Borrow the biological word trait only when it maps to a measurable property. A trait can be latency class, calibration quality, reasoning style, language coverage, refusal behavior, memory footprint, or tool-use permission. Each trait should have evidence.
RECORD DescendantTrait
trait_name
measured_value
measurement_suite
measured_at_utc
confidence
inherited_from_parent_id OPTIONAL
changed_by_operator_id OPTIONAL
END RECORDWhy rejected descendants matter
Rejected candidates teach the ecology where the search failed. Record why they failed: hard gate, insufficient gain, excessive cost, evaluator regression, unsafe behavior, compatibility failure, or duplicate niche. Rejected lineage records prevent repeated expensive mistakes.
Source 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.