During private qualification, tasks, prompts, patches, and graders run inside isolated, ephemeral Fullbeam environments. Persisted evidence is bounded to identity, coverage, lineage, cost, and outcome facts needed for the release decision.
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.
During private qualification, repository state, task bodies, prompts, patches, and private graders are confined to ephemeral Fullbeam-managed attempts. GitOps, MDM, or an internal platform keeps deployment authority over every approved Stack Release.
The persisted evidence allowlist can contain component digests, lineage, scores, measured cost, coverage, and redacted outcome facts. Raw source, command text, tool output, provider credentials, hidden graders, and model reasoning are excluded from default evidence. GitHub review features follow the separate data flow documented on the Trust page.
Compare the tools, schemas, permissions, model routes, plugins, and subagents approved for the release with the ones observed during the run.
Partial telemetry, an unresolved model route, or a missing runtime event lowers evidence coverage and withholds confidence until the gap is repaired or accepted explicitly in the decision record.
A missing runtime identity stays unknown, and Fullbeam never turns an observation gap into a compliance claim, a clean attribution, or a qualified release.
