Scout-owned per-run context, in two regions. Top-level keys are stamped by the runner at run start. Always present: harness_prompt_version (id of the harness prompt build the run was given), report_channel (which report tools the run held: none, emit, edit, or both), skill_origin (canonical or custom), github_guidance (whether the run got the GitHub evidence section), and business_knowledge_maintained (whether the run got the business-knowledge section: the product flag is on and the team's knowledge base looks maintained) — the provenance set that says which instructions the run actually got, so runs are only compared against runs of the same shape. Present only when the run departed from a default: model, runtime_adapter, and reasoning_effort (routing overrode the agent-server default), network_access (full when the scout's config lifted the trusted-domain network restriction for this run), write_scopes (the extra write access the run's token carried, when the scout was granted any), and triggered_by (manual or workflow when the run was fired off-schedule; absent means the run came from the coordinator's schedule). The nested derived object is the harness's own map of boolean run dimensions, computed server-side at finalize: has_emit_report, has_edit_report, has_self_improvement, has_chart, has_self_validation, has_structured_output, and has_not_in_use_closeout (the run wrote a not-in-use: scratchpad entry and produced no report, finding, or structured output). Use derived to answer 'what kind of run was this?' instead of parsing the summary prose. Note the flags describe the reports the run authored as they stand now, so charts attached to someone else's report via an edit are not counted. A missing derived object is unknown, not all-false: the run predates the field, never finalized, or its stamp failed.