Lineage experiments

Lineage turns experiments into memory

If a candidate improves performance but its ancestry is unknown, the system cannot reproduce the success. Lineage experiments compare family relationships so the ecology learns which parent traits and operators produce useful descendants.

Comparisons

Family scorecard

FUNCTION family_scorecard(family, suites)
    rows <- []
    FOR package IN family.packages
        evidence <- EVALUATE(package, suites)
        rows.APPEND({
            package_id: package.id,
            parent_ids: package.parents,
            operator: package.operator,
            scores: evidence.scores,
            costs: evidence.costs,
            regressions: evidence.regressions
        })
    END FOR
    RETURN rows
END FUNCTION

Detecting heritable failure

If multiple descendants from the same parent fail the same slice, the failure may be inherited. Mark the parent family with a known limitation so future candidate generators can compensate or avoid the line.

Detecting operator bias

If a mutation operator repeatedly improves public benchmarks while harming calibration or rare slices, treat that as an operator-level failure. The operator may need tighter budgets, additional gates, or retirement.

Archive use

Lineage experiments should include archived ancestors when feasible. Sometimes an older model family is less accurate on average but more robust on a slice that the new family forgot.