Introducing Sanad 0.1
A research preview investigating whether open-weight models can improve software work by exchanging direction while they are still reasoning.
Coordination while an answer is forming
We are introducing Sanad 0.1, a research preview focused on software development. The question is whether several open-weight models can exchange useful direction before they finish generating an answer.
In a coding task, an early mistake about an API, repository convention, or invariant can shape the entire solution. Selecting a better answer at the end leaves that early reasoning untouched. We want to investigate whether selective coordination can influence the path before the mistake becomes a completed change.
Phase-Locked Consensus Decoding
Our proposed mechanism keeps models active in a synchronized reasoning rhythm. A consensus layer examines partial trajectories and feeds back a steering signal while generation continues. The research is about the timing and usefulness of that influence.
Agreement needs careful treatment. Models can share a mistaken assumption, progress at different speeds, or disagree for useful reasons. A coordination rule has to preserve productive differences while limiting the cost of synchronization.
Testing the value of coordination
Software gives the study observable feedback: tests, compiler output, diffs, and repository constraints. We plan to evaluate the mechanism along four dimensions.
- Functional correctness on code generation and repair tasks, including comparison with repeated sampling from one model.
- Repository fidelity: whether the change solves the task while respecting local conventions and avoiding unrelated edits.
- The contribution of coordination through ablations of model diversity, steering cadence, consensus, and the number of participating models.
- System cost: latency, synchronization overhead, and the compute required for any improvement.
What this preview establishes
This release sets out the hypothesis, architecture, and evaluation plan. We have not yet established a benchmark improvement for Sanad 0.1. Parallel inference and correlated errors are central constraints for the study.
Approved users can explore the preview through the Sanad clients. We will use the evaluation to determine where coordinated reasoning helps software work and where its additional cost is difficult to justify.