Facts and policy
Check whether the system used approved information and followed the operating policy behind the answer or action.
Evals & Controls
Test the operation before it carries live work.
Turn expected behavior, edge cases, policy, tool use, approvals, and completed outcomes into shared quality gates across every Zoft product.
Test the system
Quality gates that follow the work across products.
Evaluate the full path from incoming request through agent decisions, tools, browser actions, human review, and the final record.
Tools and permissions
Verify that each agent selected an allowed tool, supplied valid arguments, and stayed inside its access boundary.
Handoffs and approvals
Test whether work reached the right agent or person with the context needed to continue safely.
Completed outcome
Evaluate the final business result—not only whether an individual model response sounded plausible.
Questions before release
Evals & Controls FAQ
What does Zoft evaluate?
An eval can check facts, policy adherence, tool selection, action arguments, handoffs, format, stop conditions, and whether the operation completed the intended outcome. The exact suite depends on the responsibility being released.
Do evals replace human approval?
No. Evals test repeatable behavior. Human approval remains appropriate before sensitive, financial, protected, irreversible, or customer-facing actions where policy requires a person.
Can the same controls apply across Zoft products?
Yes. The operating design can apply permissions, approvals, stop conditions, eval gates, and run records across workflows, crews, browser execution, voice, and chat.
What happens when a case fails?
The failed path should remain blocked from release while the team inspects the trace, evidence, tool behavior, and generated output. After the design changes, the relevant cases run again before approval.