Build the evaluation harness, scorecards, and policy-aware audit plus redaction flow

[KSemenenko]

Problem

Operators need a usable trust layer that combines transcript evaluation, readable scorecards, and safe audit exports with redaction control.

Scope

• Track evaluation harness inputs, transcript scoring, scorecard history, audit exports, and redaction policy behavior
• Cover how quality and trust state appears per agent profile or session
• Keep sensitive data handling explicit

Out of scope

• Low-level telemetry emission that belongs in earlier observability issues
• Toolchain installation and provider health flows

Implementation notes

• Make scorecards and audit exports actionable to operators
• Align redaction with provider secrets, prompt data, tool arguments, and results
• Keep the issue compatible with the official evaluation packages

Definition of Done

• The issue defines the minimum trust surfaces: harness, scorecards, audit, and redaction
• Later implementation can proceed without re-deciding how evaluations surface in the product

Verification

• Review the issue against the feature spec, evaluation-package issue, and approval flows

Dependencies

• Parent epic: [Epic] Add observability, evaluation, and trust controls #20
• Related epic: [Epic] Build the agent catalog and fleet-management experience #16