Writing / Whitepapers
Whitepapers
Long-form technical writing from Bluejay Labs: how our benchmarks are built, how conversations are verified, and what the results say about speech-to-speech models in production-shaped systems.
Published / 01
Whitepaper
Sep 2026
MIVAS Bench: Building RLVR environments for audio models
Multi-agent voice environments across healthcare, legal, and customer support, scored by a conjunctive deterministic verifier over database state, tool calls, and handoffs.
Faraz Siddiqi
Methodology
Aug 2026
MIVAS methodology
The full protocol for the Multi-Industry Voice Agent Simulation Bench: industry packs, digital humans, harnesses, and scoring.
Technical note
Sep 2026
Conjunctive verifier
How MIVAS scores a conversation: three deterministic gates, combined with AND, and partial rewards for a correct-but-longer route.
Industry-pack reports and failure-mode taxonomies are forthcoming.