Skip to main content
Jeff is a distinct Decision backend for the self-hosted logan-markewich/jeff service. It uses the System One request contract, while retaining Jeff-specific deployment, rounding, error, and telemetry semantics.

​
Configure the service

Set the operator-owned endpoint and optional bearer key:
Then select the bundled preset:
The preset targets /v1/systemone and requests gliformer-large-v1. It does not start Jeff, download model weights, or choose a compute device.

​
Alternative typed-decision lab host

The separately maintained typed-decision-lab can host the same model through its pinned jeff-torch adapter. Run that repository under Python 3.12 with its jeff extra and point JEFF_API_URL at the lab server. The adapter composes Jeff’s engine directly; it does not proxy or launch another Jeff service. The lab fixes Jeff’s temperature to 1.0 and question isolation to all, and reports both facts through /stats. Its jeff-v1 response profile preserves the TypeSafe-compatible Jeff answer shape without fabricating Laya action signals.

​
Required Score deployment setting

Run Jeff with:
Jeff’s upstream default temperature scales the displayed probabilities while leaving Score based on the raw distribution. That representation is internally inconsistent and Polyglot rejects it. Before advertising Score support for a deployment, verify GET /stats returns temperature: 1.0; Polyglot deliberately does not call /stats for every decision. If the deployment keeps another temperature, use Choice and Noul independently and mark Score unsupported in its application model record.

​
Operational behavior

  • Authentication is optional. Authorization is omitted when JEFF_API_KEY is empty.
  • Independently rounded four-decimal distributions are retained without fabricated normalization.
  • x-typesafe-request-id becomes the provider request ID.
  • x-jeff-server-ms and x-jeff-batcher-ms remain available on DecisionResponse::responseData(); they are not end-to-end duration.
  • HTTP 429 and queue-full 529 responses are retryable. The standard Retry-After seconds header drives the current retry policy; raw millisecond and timing headers remain on the HTTP response.
Jeff model output is not Jev output. Capability declarations describe native wire semantics, not calibration, accuracy, or application fitness; evaluate the deployed model on representative data with the provider evaluation workflow.