RFC: Agent Outcome Observability and Assurance Metrics

[dinpd]

Summary

• adds a focused community RFC for measuring agent outcomes, constraint compliance, control effectiveness and friction, execution discipline, and evidence quality
• defines independently adjudicated positive/negative control semantics (correct_enablement, correct_intervention, unnecessary_intervention, and missed_control)
• makes eligibility, exclusions, indeterminate states, numerators, denominators, profile compatibility, privacy, and trace-sampling independence explicit
• links the companion RFC from the proposals index without redefining the Agent Action Boundary Evidence lifecycle

Acceptance validation

• python -m pytest — 99 passed
• parsed both fenced JSON examples
• verified all required RFC sections and local links
• git diff --check
• reviewed current primary references, including OpenTelemetry GenAI semantic conventions

Security review

Documentation-only change. No dependency, authorization, or runtime behavior changes. The RFC itself includes tenant isolation, telemetry leakage, metric gaming, self-confirming controls, evidence selection, high-cardinality, and historical-authority risks. A scoped credential-pattern scan found no secrets.

Closes #58