This Phase B experiment isolates the prevalence of one required capability from the bundled R2 stress regimes. It asks how much candidate information a local scheduler needs as eligible workers become rare.
For each seed, every cell reuses the same worker capacities, zones, task
arrivals, and unit-work task stream. Every task requires rare-capability.
Deterministically ordered, nested worker subsets provide it at 50%, 20%, 10%,
5%, 1%, and 0.1% prevalence. Churn, bursts, availability lag, and load lag are
all zero; offered load stays at one task per tick. This avoids conflating
capability rarity with the earlier composite regimes or deliberate saturation.
The sweep compares one-random, power-two, power-three,
capability-power-two, and global-least-loaded on identical projected traces.
It retains five raw seeds per cell and reports the same bounded Student-t 95%
uncertainty summaries used by the repeated-seed scale study.
python3 -m randomness_lab.r2_capability_rarity \
--workers 1000 \
--fractions 0.5,0.2,0.1,0.05,0.01,0.001 \
--seeds 41,42,43,44,45 \
--ticks 100 \
--arrivals 1 \
--drain-ticks 250 \
--output results/experiments/r2/capability-rarity-seeds41-45.json
Interpret completion, latency, capability-mismatch failures, and metadata probes together. Capability-aware sampling can pay for additional metadata and still fail under other factors not present here. These results do not establish optimality under churn, stale observations, correlated regional loss, heterogeneous requirements, or saturation; those remain separate issue #84 sweeps.
The first five-seed reference result and guarded interpretation are retained in
../../results/experiments/r2/capability-rarity-seeds41-45.md.