qiluoli-usc’s Copilot Discussion lets an existing assistant conduct a continuing research conversation through Microsoft 365 Copilot’s website. It saves questions, answers and timing records. Its supported browser path targets Windows and Chrome.

The default path uses the executor

The host assistant supplies each exact question. Local code selects the observed message field and send button, then uses Jev Native’s browser layer to act.

  • The default worker declares decisionModel: false.
  • It checks the owned tab and approved draft before sending.
  • It records the runtime source fingerprint and action receipts.
  • Browser-action time sits inside the larger discussion timing; adding both would count it twice.

Using Jev’s execution code therefore says little about whether a Jev model improved the research.

The model comparison stopped before an answer

A separate experimental adapter adds model.choose to select the same permitted controls. The comparison fixes the questions and executor, with an ABBA order and two planned samples per mode. Those are protocol settings, not completed results.

The project’s benchmark note reports two synthetic attempts with one Jev call each. Its second diagnostic returned BLOCKED at 0.34 confidence. The author says no input or send followed, and the real Copilot comparison was abandoned. This failure was reported by the project, not reproduced with live inference here.

The adapter’s inspected guard requires the configured model, expected control and confidence checks. A blocked answer raises a failure. Typed output is not a correct decision.

Keep the speed claim unresolved

There are insufficient successful model-enabled samples to compare speed. Copilot’s generation, network delays, polling and saving also contribute to elapsed time. The failed adapter run says little about Jev’s quality on other tasks.

Verification: at 318048c09fa7ff2b6b7bb14da4fb3a32e72d924f, six offline JavaScript tests and three Python decision-guard tests passed; one Windows-only Python test was skipped. Browser transport and model replies were mocked. No live Copilot session, Jev inference or security audit was performed.