Productivity should compound
Each model, agent endpoint, or skill should add more useful work than operating complexity. Solver gives new capabilities the same eligibility, ranking, lifecycle, and evidence model so teams can grow the system without rebuilding its controls.Limits come before work
Data, cost, time, and permitted actions belong in the run request. Solver does not infer broader authority from the wording of a task.Eligibility comes before ranking
An option must be ready and allowed before Solver compares it with anything else. An unavailable or disallowed arm is not a fallback.Identity stays explicit
For deployment-bound model runs, the requested model, provider-reported model identity, serving provider, provider connection, and commercial owner remain separate facts.Access paths stay explicit
Your provider key and model access provided by Millwork are different commercial and security boundaries. Deployment-bound model provenance records which path was used.Output and evidence have different jobs
The result carries useful content. The receipt explains the run without copying the task or output. Keeping them separate makes each safer and more useful.Action remains bounded
Solver filters for eligibility before selection. After a check, it may fall back or create a repair proposal only when the run allowed that response. A skill proposal never changes the saved skill until an authorized person approves it.Availability is reported, not implied
Provider names and examples are not promises. The live source profiles and model catalog show what an organization can connect and run now.See current availability
Check the private-preview boundary before designing an integration.