jev vs perplexity decisions api

Start with the Request, Not the Name

Both APIs take bounded yes/no, choice and ordered-score questions. Perplexity names its model pplx-decider-v1-27b and uses /v1/decisions. TypeSafe publishes the Jev primitives under its own endpoint. Before swapping a client URL, test authentication, invalid questions, response fields and errors. Shared question names are a thin form of compatibility.

Perplexity describes a Qwen3.8-27B fine-tune and provides a model card and inference example. The weights are available, but they are Perplexity’s model, with its own revision and license to inspect before local serving.

Compare One Support-Routing Task

An invoice download fails because the customer cannot sign in. Under the first-blocker rule, account access owns the ticket. That expected label tests your rubric, not either model’s performance. Keep other tickets that mention billing while the actual blocker sits elsewhere.

Published Differences and Unknowns

Perplexity lists a 128-question limit, 262,144 input tokens, image size rules and 429 behavior. None of those limits carry over to Jev. The model card compares eleven benchmarks and shows wins on both sides. It is a Perplexity-run panel, so use the task rows to design tests rather than to rank the models for every application.

How to Test a Switch from Jev

The benchmark method gives you the paired record. Check probabilities on your ticket population using calibration testing, then set the review threshold from that result.

1. Pin both model versions, endpoints, questions and evidence projection.
2. Run contract fixtures for Choice, Score, Noul, invalid state, timeout and rate limit.
3. Score both against the same untouched human labels; keep every failure in the denominator.
4. Fit separate review thresholds and compare false actions at equal automation coverage.
5. Measure full provider latency and cost including retries and human review before rollout.