Module crowdless_nsga2 

The regression sentinel for Nsga2: the same genetic algorithm with crowding distance deleted.

NSGA-II’s contribution is not that it uses a genetic algorithm on a multi-objective problem — it is how it ranks a population: non-domination rank first, crowding distance second. The second half is the entire diversity mechanism. Rank alone cannot separate two mutually non-dominated solutions, and once a run has converged nearly the whole population is rank 0, so from that moment on crowding distance is doing all of the deciding — in the parent tournament and in the survival truncation alike.

A claim like that is only worth asserting if something would notice when it stopped being true, so this module ships the ablation the oracle measures against: NSGA-II with the crowded-comparison operator’s second key replaced by a constant.

It is an ablation, not a second algorithm

Everything except crowding is the real sampler’s (crates/atune_core/src/sampler/nsga2/mod.rs), on purpose — including the operators themselves, which are called rather than re-implemented:

Piece	Here	Nsga2
initial population	random_unit_vector, one uniform draw per gene	same function
parent selection	binary_tournament on (rank, **0**)	binary_tournament on (rank, crowding)
crossover / mutation	crossover / mutate, η_c = η_m = 20, rate 1/d	same functions, same constants
survival	pool parents + children, fast_non_dominated_sort, fill front by front	same
splitting-front truncation	by pool order	by decreasing crowding distance
admission guard	evaluated-is-the-bred-config, finite objectives	same
RNG	ChaCha8Rng seeded seed_for_name(sampler_seed, "nsga2")	same

The RNG tag is deliberately the same string, so for a given (study seed, trial number) both samplers consume the identical draw stream: the sentinel is a paired comparison in which crowding distance is the only moving part. That is what makes “the crowding-blind variant loses the diversity margin” a statement about crowding distance rather than about two unrelated searchers.

Note what the ablation is not: it is not a weak searcher. It keeps elitism, non-dominated sorting and the whole variation machinery, so it still converges toward the front — it simply has no pressure to stay spread along it. That is precisely the failure the diversity axis exists to catch, and precisely the one a hypervolume-only gate can miss.

Why it is here and not in the facade

It is a benchmark fixture. atune ships one NSGA-II; shipping a second, deliberately weakened one as public API would be a maintenance liability and an invitation to use it. atune_bench is publish = false, which is exactly where an ablation belongs — the same reasoning that put IsotropicEs and DeKnobs::degenerate here.

Determinism

Identical to the real sampler’s: every draw comes from a ChaCha8Rng keyed on the trial’s sampler_seed, never ambient entropy, so a single-worker run with a fixed study seed is fully reproducible (docs/design/09-implementation.md §5, promise 3). Like every population-based sampler it is history-dependent, so it does not get promise 2 under parallelism — and, being a bench fixture, it never persists state at all (state returns None).

Structs

CrowdlessNsga2

NSGA-II with crowding distance deleted: the diversity ablation.