dsh-deliberation

In multi-step reasoning tasks, a Primary model typically advances along a single trajectory with no controlled second perspective to catch course errors in time. dsh-deliberation adds bounded deliberation to DeepSeek Harness: when the Primary opts in, it calls a single deliberate tool that spawns up to three concurrent children—each taking a distinct role of independent-alternative, trajectory-audit, or reasoning-masked re-read—and returns compact, structured decision-support packets containing evidence, uncertainty flags, and suggested next checks. Children only propose; they never vote, select a winner, or replace the Primary's final judgment.

A second, fully opt-in path provides automatic masked review at Turn boundaries. Once enabled in the profile, Harness forks the current tool-bearing Turn, strips the recorded reasoning chain, and lets a child blind-read only the observable actions and results to surface errors the Primary may have missed. Publishing modes range from updates-only through all to observe-only, so teams can collect shadow-evaluation data before opening the loop.

The plugin suits developers running multi-step inference on DeepSeek Harness who want a controlled audit layer without altering the Primary workflow. The project is an experimental preview: runtime and failure boundaries are tested, but paired benchmark data has not yet established a correctness or cost benefit, so starting with manual deliberate calls before enabling automatic review is the recommended path.