The question behind the project
Can carefully orchestrated models produce better engineering decisions than a single, unchallenged answer? Verdict explores this through independent reasoning, adversarial review, and explicit evidence requirements.
What we are exploring
- Whether diversity of approaches matters more than the number of agents.
- How blind evaluation changes model preference and adjudication.
- When execution-based checks add more value than agreement alone.
- Whether failure memory and adaptive routing improve repair while controlling cost.
Where it stands
Verdict is a local experimental project with an evolving multi-model workbench and a broader coding-system roadmap. A proposed evidence check is not a completed test; model confidence is not a calibrated probability. We have not published a benchmark establishing superiority over other systems.
What comes next
The research direction is to compare configurations on defined engineering tasks, retain run traces, and measure correctness, cost, and failure modes. Planned capabilities are evaluated separately from demonstrated results.
Research preview · Under active development · Not a public production service
Talk about the research