jev-corelisted

# Jev core (engine reference) **Engine:** `jev <mode> <input.json> [out.json]` (installed by `install.sh`; or `node <repo>/bin/jev.mjs`) — needs `TYPESAFE_API_KEY`. Exit 0 = every item passed its gate, 1 = at least one failed, 2 = error. Examples for every mode: `bin/examples/` in the pack. Jev (TypeSafe System One, pinned `jev-1.13.0`) returns calibrated probabilities, not text. The engine asks **several phrasings** of each yes/no question and reports `mean [95% interval]` across them. A narrow interval = a stable judgment, **not** proof of correctness. | Mode | Input items | Use for | |---|---|---| | `claims` | `{id, claim, evidence}` | Is a statement backed by the evidence? | | `resolve` | `{id, finding, fixed, original}` | Does a fix resolve a finding? `original` is the control | | `ac` | `{id, criterion, evidence}` | Is one acceptance checkbox met? | | `tests` | `{id, behavior, test}` | Does a test prove the user-visible behavior? | | `noul` | `{id, state, question, variants[]}` | Any custom yes/no (give 3+ variants) | | `score` | `{id, state, dimensions{name:{question, levels[], weight}}}` | Rank options on weighted rubrics | | `rerank` | `{id, query, candidates[{id,text}]}` | Order sources/tickets/files by relevance | | `choice` | `{id, state, question, options{k:desc}}` | Route to one of N buckets | ## Calibrated gates (jev-1.13.0, measured on paired good/bad examples) Evidence gates pass at mean ≥ 0.50 and lower bound ≥ 0.40. Clear good cases land 0.55–0.90, clea