Approximate vector indexes are fast but can silently forget results. SIEVE is the deterministic index with guaranteed zero false negatives: where missing a result is a loss, it finds them all and certifies completeness.
For range and k-NN in L2, SIEVE proves that no neighbor within the radius was left out. Not '99% recall': proven absence — 'I searched everything and none is left', the sentence compliance and forensics buy.
SIEVE ∘ CLAMP = certified aggregate over a certified neighborhood. Proved on real data (327/327, a 2.26 °C bracket = ε + search). Completeness enables attribution, and attribution enables audit.
The claim 'faster than brute force in RAM for L2' was refuted and reported — SIEVE's strength is the guarantee (deterministic zero-FN), not raw speed. We say what we measure.