TypeSafe Triage Guard

Watch out

The repo describes itself as R&D and a worked example, not a product, built to test what the primitives are good at outside the docs' own demos - so read it as a pattern catalogue rather than something to deploy. Most of the commands shown use --mock, and the mock is a keyword-overlap heuristic with no semantic understanding: the README says it misses a phrase like pick a lock on someone else's front door entirely because there is no shared vocabulary, and that its Score and Choice outputs get noisy on near-zero-overlap inputs, so its raw severity numbers should not be read as meaningful. Passing tests therefore demonstrate composition logic - thresholds, weighting, short-circuiting - and not model quality. The real model path needs an early-access key; a live-smoke CI job does run all three CLIs against Jev on every push using a repository secret, and one unedited JSON output per CLI is included. Beyond those single samples there is no evaluation set, no accuracy figure, no false-positive rate, no latency and no cost per pipeline documented anywhere. Weights and policy tables are the intended tuning surface, but no defaults table or threshold list is published. This README copy is cut off mid-JSON in the alert example, so the later sections, including the referenced write-up on the calibration objective, were not visible. Copy reviewed against the repository README on 2026-10-04, not from running it.