What ND Can and Cannot Promise

Why this exists

ND is powerful but only safe when its limits are explicit. This page exists to set expectations before you rely on ND in production.

What ND optimizes for

• Speed at scale
• Bounded quality with explicit metrics
• Transparent uncertainty

What ND cannot guarantee

• Bit‑identical replay
• Exact ranking agreement with deterministic search
• Stability without fixed index hashes and parameters

What replaces guarantees

• Quality metrics (instability, margin, entropy)
• Witness checks when enabled
• Declared randomness profiles and bounds