September 2026 Summaries

TrustedRouter presents Trev, Zev, and Lev as decision-model configurations using a shared API that returns typed, probability-based answers for tasks such as ticket routing, refund detection, and urgency classification, while allowing users to enforce zero-data-retention routing. In AnyEval’s public JevBench snapshot of 231 tasks, Jev led with 199 passes and the lowest recorded model cost, followed closely by Zev with 194 passes, while Trev and Lev scored 187 and 178 respectively; Zev and Lev recorded faster average evaluation times than Jev, whereas Trev was substantially slower. The results are presented with important limitations, including overlapping confidence intervals, benchmark-specific costs and latency, incomplete provider identity data, and the lack of evidence about probability calibration or production performance. For privacy-sensitive use cases, users can pin Zev to Baseten, Trev to Cerebras, or Lev to Parasail and require a ZDR privacy floor, though provider policies may still permit operational metadata retention and should be checked before sending sensitive information. Jev may offer enterprise ZDR directly through TypeSafe, but that arrangement is not configured for TrustedRouter’s Jev route, while gateway attestation and provider retention policies represent separate privacy and security boundaries.

TrustedRouter has introduced an alpha decision API designed for production workflows in which LLM outputs drive classifications, routing, scoring, or other conditional actions rather than requiring prose responses. The `/api/alpha/decide` endpoint accepts a model, JSON state, and named boolean, multiple-choice, or scaled-score questions, returning validated answers with probabilities in a single call. The announcement highlights Jev from TypeSafe AI, along with eight compatible models including the default open-weight `trustedrouter/trev-1.0` and Mercury 2-based `mev-1.0`, and reports benchmark results, latency, failover arrangements, and per-decision costs for a small support-ticket evaluation. All models use a common request and response format, allowing model selection through configuration while preserving pinned model versions. Unlike conventional chat-model JSON prompting, the service validates response options, probability ranges, and distributions, while confidence scores allow users to set automated-action thresholds and send uncertain cases for human review.