BrettReifs’ Decision Kit is a Copilot Canvas workbench for incident response, feature prioritization and customer escalation. Jev selects a branch and supplies typed scores; a separate chat model writes the explanation after code computes the review gate.

I inspected the README, request/response logic, provider adapter and Canvas server at commit 416182c6b7e8dfdc6f99b8d689a8c8351cb4aaba. I also ran the repository’s offline suite: 71 tests passed, none failed. I did not run a live Jev request or launch the Copilot Canvas; this was not a security audit or an accuracy evaluation.

A gate computed outside the narrative

  • One decision request asks for a branch (choice), urgency (score) and whether a human must review (noul, a probability).
  • Code takes the lower of branch and urgency confidence, then routes to review below the project’s 0.72 cutoff or when its review probability reaches 0.5.
  • The interface retains the structured decision separately from the chat model’s rationale and next steps.
  • Those cutoffs are configuration choices, not reproduced evidence of calibration or safe automation.

Honest demo mode matters

  • Without a credential, the provider returns explicitly labeled deterministic fixtures—not live Jev output.
  • Live provider failures remain errors rather than quietly turning into a successful-looking fixture.
  • The default live path uses OpenRouter. This capture did not enable that provider or use an account.

The boundary is narrower than the README sounds

  • Response checks reject missing answer types, nonfinite scores and out-of-range probabilities.
  • They do not establish that probability keys match the original option menu, sum to one, or that urgency stays within its intended scale.
  • The LLM cannot overwrite the stored branch through its parsed response, but narrative fidelity is largely a prompt instruction: the parser checks strings, not whether the explanation contradicts the decision.
  • Calling Jev “authoritative” inside this demo means the UI’s source of decision fields—not independently verified truth or permission to act.

A separate delivery-nudge hook and outcome evaluator are also included, but they are not Jev features. Their synthetic regression tests do not demonstrate faster shipping. The useful pattern here is separating decision data, review policy and explanatory prose—then testing each boundary without mistaking a green test suite for good model judgment.