Reducing LLM Hallucination from 27% to <1%:
Multi-Agent Council Architecture

Measured 27% hallucination rate in production before any intervention. Single-model confidence gating reduced it to 18% — then hit a floor. Getting below 1% and holding it required a different architecture.

The Measurement Problem

27% is not an estimate. It is a measured hallucination rate from a labeled evaluation set of 4,200 production queries run through the model at baseline — before any mitigation. The evaluation methodology:

The 27% figure is consistent with published research on similar model families under domain-specific queries. User satisfaction lags hallucination events by 3–6 months — by the time the signal surfaces, the failure class has compounded.

Timeline: 27% → <1%

The Council Architecture

Single-model confidence gating reduces hallucination but hits a floor around 5-8%. Below that floor, the model's uncertainty is no longer well-calibrated — it confidently hallucinates. Breaking through 5% requires architectural diversity, not threshold tuning.

The multi-agent council runs each query through independently prompted agents with distinct system contexts. Three voting lenses per response:

The Bayesian prior injection (Month 4) was the inflection point from 6.2% → 2.1% → <1%. Per query class, prior distributions encode known hallucination failure modes for that domain. Council vote weights shift based on prior calibration — agents with better historical calibration on a given domain class carry higher vote weight.

Why <1% Is Not the Same as 0%

<1% sustained means the golden set evaluation rate is below 1% on every weekly measurement. It does not mean zero hallucinations in production — it means the rate is below the threshold at which downstream systems treat the output as ground truth without verification. For the specific use cases in this deployment (regulatory document analysis, model risk reporting), <1% is the threshold at which human review is triggered less than once per 100 queries.

The 0 P0 across the internal replay corpus figure is separate: it measures hallucinations that reached a downstream decision without human review interception. Zero divergence across the internal replay corpus means the council + confidence gate architecture caught every near-miss before it escaped.