Cheap probabilistic triage for test/build/lint/typecheck failures. Use before broad reasoning when the relationship or scope of a failure is unclear. Returns structured likelihood signals; does not diagnose or fix the problem.

Cheap comparison of two unsuccessful implementation attempts. Use when repeated work may be spending frontier inference on the same failed strategy. Returns stuck/progress signals and an advisory control signal; does not choose the next fix.

Cheap requirement/evidence coverage check before an expensive full-task review. Use on large tasks with explicit requirements. Returns per-requirement likelihoods and which items warrant review. Never certifies correctness, safety, or merge readiness.

Cheap ranking of repository context candidates before deeper frontier inspection. Use when discovery produced many plausible files or symbols. Ranks only; never deletes or permanently excludes candidates.

Cheap normalization of review findings from tests, linters, humans, or frontier reviewers. Use to triage a large finding set before expensive reasoning. Returns likelihood signals only. Does not confirm that a vulnerability or defect exists.

Cheap change-risk triage to decide whether a more expensive frontier review appears warranted. Returns sensitivity likelihoods only. These are not security conclusions and do not certify safety.

Generic bounded judgment primitive. Ask one evidence-grounded yes/no proposition per question and receive calibrated probabilities. Rejects generative requests such as writing code, designing architecture, or fixing bugs. Jev evaluates supplied evidence; it does not invent it.