Delegation context and evidence interpretation

A specialist starts with a fresh context. A useful handoff identifies the bounded question, exact target and permitted actions, acceptance observable, environment, relevant evidence excerpts and paths, known failed approaches, remaining uncertainty and budget. Credential references should identify protected artifacts, not repeat secrets. The live task schema supplies actual field names and available specialists; examples from another installation do not override it.

Context quality

Distinguish direct observations, inferred mechanisms and untested assumptions. For a proof question, include a competing benign explanation and discriminating control. A code tree, existing reproducer or prior capture may supply enough context without another collection pass. Include exact versions and session prerequisites where behavior depends on them. Summarize large outputs without dropping conditions that would change the conclusion; keep raw artifact links for independent assessment.

Result interpretation

An agent return is testimony to inspect, not proof that its acceptance criterion was met. Look for the actual observable and whether the evidence belongs to the claimed target, identity and run. An empty output, environment fault, exhausted budget, refuted hypothesis and meaningful negative result are different outcomes. Repeated failures with the same mechanism do not justify confidence inflation. Preserve uncertainty and provenance when revising a conclusion.

Startup, dispatch ordering, execution-state mutation, quality gates and engagement close are defined by the selected workflow and injected runtime contracts. This reference does not define another plan, a mandatory role chain or a report sequence.