Niche-map design

Descriptor choice determines archive value

A quality-diversity archive is only as useful as its niche map. If descriptors are arbitrary, the archive becomes novelty theater. If descriptors map to real operating constraints, the archive becomes an adaptive reserve.

Good descriptors

Good descriptors are measurable, decision-relevant, and stable enough to compare across candidates.

Bad descriptors

Bad descriptors are easy to compute but hard to interpret: arbitrary embedding clusters, aesthetic style labels, internal activation distances without behavioral meaning, or scores that merely duplicate the fitness function.

Niche design workflow

FUNCTION design_niche_map(task_inventory, constraints, policy)
    descriptors <- []
    descriptors.ADD(task_family FROM task_inventory)
    descriptors.ADD(runtime_class FROM constraints.hardware)
    descriptors.ADD(latency_band FROM constraints.service_levels)
    descriptors.ADD(risk_tier FROM policy)
    descriptors.ADD(error_profile FROM evaluation_slices)

    descriptors <- REMOVE_REDUNDANT_DESCRIPTORS(descriptors)
    descriptors <- LIMIT_CARDINALITY(descriptors, policy.max_niches)
    RETURN descriptors
END FUNCTION

Cardinality control

Too many descriptors create empty niches. Too few descriptors erase useful diversity. Start with a small map, then split niches only when they contain multiple meaningfully different candidates.

Review cadence

Review the niche map after major workload shifts. A niche that was useful last quarter may be obsolete; a new task family may need a new descriptor. Treat niche-map changes as evaluation changes because they alter selection pressure.