jev-harnesslisted

# JevHarness Turn the user's intended outcome into a runnable, measurable decision workflow: ordinary code prepares observations and enforces actions; Jev supplies structured judgments; an optional reflection model proposes improved pipelines offline. This is task-level pipeline optimization, not model-weight training. The deployed pipeline can retain Jev while removing the reflection model. ## Establish the task before building Read the user's request and existing authorized project material first. Reuse answers already given. Ask focused questions about missing information that would change the implementation or evaluation. Do not start a new task's implementation, generate its questions/criteria, access credentials, or run a paid experiment until enough of the following contract is known. Read-only inspection that helps identify those gaps can proceed. - **Outcome and scope:** What should improve, for whom, and what counts as success? Is this one decision per input or a multi-step episode? What is outside scope? - **Inputs and outputs:** Obtain representative input, desired output/action, available fields, legal actions, and failure behavior. Ask for a concrete example when a verbal goal leaves the semantics ambiguous. - **Evidence:** Which examples, logs, labels, simulator, or real outcome signals exist? How may they be accessed and sent to providers? Which information is available at decision time? - **Evaluation:** Who or what can score a result, on what timescale,