Orchestrator-Only (No Direct Data)

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

An orchestrator analyzes request-failure counts without directly reading customer payloads. A teaching bounded local data role aggregates logs and returns time bucket, error code and count for orchestration.

Mechanism

Define orchestrator scope and data-role permissions: approved execution roles handle reads/mutations, with no mutation needed here. Send aggregation/field contracts without raw logs. The role returns minimal schema from a controlled environment. Inspect identifiers/secrets before handoff and retain run identity/errors/coverage; summaries do not imply full raw-data coverage.

Bad example

Have the orchestrator read all raw logs first and redact later, while specialist roles receive unrestricted access.

Good example

The orchestrator defines time-bucket/error counts only. An authorized local role reads bounded logs and returns three fields plus coverage/errors, keeping payloads/credentials outside model context. Inspect schema before synthesis; delegation grants no extra data authority.

Why the change matters

Division plus minimal handoff places access responsibility at a defined boundary and limits orchestration exposure. Redacting after raw-data ingestion cannot provide that boundary.

Observable expectation

Teaching orchestrator inputs contain only aggregates/scope and no raw-log reads. Role requests remain authorized; an email field makes handoff invalid. Independently reconcile counts in the data role without filling unread files with zero.

Limits

No frozen source is confirmed. Roles alone do not isolate tools/credentials; implement access controls. Aggregates can leak small populations. This architecture example authorizes neither agent spawning nor customer-log access now.