Jev MCP — blakestone-x

There are two projects with this name. This one is Python and installs with uvx from a pinned git tag; the tools are jev_ask, jev_classify, jev_score, jev_check, jev_match, jev_screen and jev_health, and the envelope carries probabilities, a band, token usage and latency. It is not the TypeScript server with the nine coding-loop tools.

One state, one label. The worked example sends five ticket texts in a single state and gets one label back for the combined set, so send one request per ticket when you need a label per ticket. Candidate matching is the tool that scales: jev_match takes up to 2,000 candidates and bills ceil(n / window) + 1 requests.

From a field-service company's production data. Clean labels at confidence 0.8 to 1.0 agreed 95% of the time; one-line label descriptions reached 81.0% at 1,689 tokens per request and rich descriptions 84.5% at 4,745 tokens. Reversing criteria order flipped 32 of 200 choices, with flipped confidence averaging 0.42.

The provider publishes none, and the README says so. Observed on one early-access key, small requests ran at 160 per second with flat latency, while a sustained 40 requests per second at 1,700 tokens each drew 429s after three minutes. Treat the limit as tokens per minute, keep bulk jobs resumable, and back off for seconds rather than milliseconds.

Seven: jev_ask for your own Choice, Score and Noul map, jev_classify for one label from a closed list, jev_score for an ordered rubric, jev_check for independent yes/no propositions, jev_match for matching a query to candidates with abstention, jev_screen for filtering untrusted instruction-like text, and jev_health for listing models and measuring round-trip latency.

ok, the model id, an answers object with the choice, confidence, a full probability map and a band, a usage block with input tokens, output tokens fixed at zero and an estimated cost, and a latency in milliseconds. The docs report about 130 ms per judgment.

jev_match accepts a window from 20 to 254 and up to 2,000 candidates by default, billing ceil(n / window) + 1 requests, one per window plus a finalist call. The request char budget defaults to 120,000 via JEV_MCP_MAX_REQUEST_CHARS, and the match deadline is 45 seconds via JEV_MCP_MATCH_DEADLINE_S.