Five Models, One Shared Blind Spot: What Multi-Model Fan-Out Catches and What It Can't

Short summary

A measurement lab tested five model families on identical evidence digests with deterministic numeric scoring. All 68 numeric citations passed recomputation across models (max 1.68% deviation from rounding), but this only proves copy fidelity, not interpretation quality. The real value of multi-model fan-out was outlier detection — four models independently flagged a data anomaly the human author had missed in their sealed interpretation.

• •68/68 numeric citations passed recomputation across 5 model families — zero hallucinated numbers
• •Clean audit scores measure copy fidelity, not interpretation quality — a floor check, not a ceiling
• •Fan-out's real value is outlier detection: 4 of 5 models caught an anomaly the human author missed

Generated with AI, which can make mistakes.