Reproduction network v2

The two rings

The hub is the original benchmark — a measurement that editorial has approved and made public. Reproductions attach to that hub in two ways. The inner ring shows community reproductions: operators who ran the same triple (model, hardware, runtime) on similar or identical hardware, confirming the result is repeatable rather than a one-off. The outer ring shows independent operators who ran the measurement on distinct hardware — different GPU generation, different OS, different runtime version — and still landed inside the tolerance band.

Both rings matter, for different reasons. Community reproduction protects against measurement error: if three operators on the same rig family land within ten percent, the original number is probably right. Independent reproduction protects against generalization error: if four operators on four different hardware classes land within tolerance, the result is robust to setup details, not just a one-machine artifact.

The emerald arrow is the confidence-tier upgrade path. It starts at the hub and sweeps through the rings to show the mechanical rule: two independent operators on distinct hardware moves the row from moderate to high; four reproductions across distinct hardware moves it to very- high. These thresholds are deliberately stricter than the older “independently-reproduced” promotion bar because confidence and reproduction-count are different measurements.

Operator names in the diagram are placeholders. Real attribution lives on the benchmark detail page and on /operators. The diagram’s purpose is to explain the structure — who is in which ring and why — not to credit individuals.