Skip to main content

T2. Scale and Noise: the credibility benchmark

Vector search looks great in a demo and degrades in production: after months of use the pool fills with near-duplicate and contradictory records, and the right memory stops making the top results. We tested this deliberately, and we report where it breaks.

Recall holds, then crowds​

We buried 8 known facts under up to 1,000 realistic distractors, 25% of them adversarial near-misses (same wording, wrong value). Retrieval recall stays perfect through 500 records. At 1,000, under that hostile mix, the most-contested fact is pushed out of the top 5. Latency stays flat throughout (13 to 42 ms).

Recall understates the real cost​

The number that matters is whether the answer is correct. Even when the true fact was still retrieved, the block also carried several contradictory wrong values, and the model could not tell which was authoritative: it either picked a wrong value or refused to answer. Answer fidelity fell from 0.94 to 0.42. This is the classic vector-store failure at scale, and pure semantic retrieval is subject to it.

The architectural fix: trust in ranking​

The fix is provenance. When contradictions arrive from lower-trust sources (bulk imports, unvetted content) while the real decision is self-authored, trust can rank the authoritative memory first. We measured two settings:

  • Down-weighting provisional memories helps but is partial (0.42 to 0.65): a down-ranked contradiction is still injected and still confuses the model.
  • Excluding provisional memories from the decision set restores full fidelity (0.42 to 0.94).

Honest scope. This recovers the common case, where a contradiction is lower-trust than the authoritative fact. It does not resolve contradictions that are equally trusted and equally recent; those require temporal belief revision (see T3) or human review. Trust-in-ranking is a validated capability currently behind a feature flag and not yet enabled in production.