Modern harnesses can reroute a model, load a plugin, change a tool schema, widen a permission, or hand a step to another agent during execution. The declared release and effective runtime can diverge.
Each result stays tied to the Stack Release, workload, repository state, and coverage that produced it. If evidence is missing, the gap remains visible in the release call.
It contains the model, harness, instructions, tools, permissions, routing, context policy, workflow, verification, and runtime that the team approved for this evaluation and intended to place in front of developers.
It follows the model routes, tool schemas, plugins, permission state, subagents, and transitions observed while the session moved from task to accepted result.
A low-cost route cannot claim a success completed by an expensive fallback, and a candidate cannot pass with a tool or subagent that never cleared qualification.
Coverage changes the call. Verified evidence supports attribution, while partial, drifted, and unknown states restrict the claim or keep the workload under evaluation.
Qualification follows the run. An approved label cannot give an altered runtime credit for work it was never cleared to perform under the release policy.
