Auspexi

Hallucination Controls: Runtime Gating, Abstention, and Evidence

Auspexi • Updated:

High‑reliability AI is less about “eliminating” hallucinations and more about controlling their impact at decision time. Our platform implements practical guardrails that prefer a correct abstention over a confident error, with auditable evidence for every change.

TL;DR: Optimize a 3‑state objective (correct, abstain, wrong), gate responses on information sufficiency, verify outputs, and ship decisions with signed evidence and SLOs.

The 3‑State Objective

Coverage–Precision Tradeoff with Abstention Precision Coverage No abstention: more coverage → more wrong answers With abstention & gating: tuned operating point Chosen OP No abstention Abstention + gating Operating point
Figure: Abstention + gating yields a tunable operating point with higher precision at the same coverage.

Information‑Sufficiency Gating

Shadow Evaluation and SLOs

Grounded Inputs Without Raw Data

Synthetic outputs are calibrated against anchor bundles (aggregates) or ZKP‑protected seeds. No raw row transfer; provenance includes anchor_hash.

Falsifiable Test (Production‑Practical)

  1. Fix latency.
  2. Clamp coverage with rejection sampling.
  3. Measure wrong‑answer and re‑ask rates on a fresh‑news holdout, pre/post gating.

FAQ

For deeper details, see our Whitepaper: Hallucination Controls and FAQ.