Roadmap: Evaluation pipeline; Validation and risk analytics
Priority: P1
Component
CDK / infrastructure
Describe the feature
Emit harness-level CloudWatch metrics and dashboard panels that evaluate the control plane/runtime separately from raw task pass/fail. Final task success conflates model quality, harness quality, and environment luck (arXiv:2605.18747 §5.2.1).
Use case
- Operators/SREs track whether the platform is getting more reliable independent of model upgrades.
- Field engagements report dark-factory scorecard progress with objective harness SLOs.
- A/B prompt experiments isolate prompt changes from harness regressions.
Proposed solution
Define and emit metrics (names illustrative; finalize in implementation):
| Metric |
Definition |
HarnessToolCallsPerTask |
Agent tool invocations per completed task |
HarnessTokensPerTask |
Bedrock tokens (from existing cost telemetry) |
HarnessWallClockSeconds |
Submit → terminal |
HarnessVerificationCoverage |
% tasks with full evidence bundle / all required sensors run |
HarnessRecoveryRate |
% tasks that failed verify then succeeded on retry (when fix-up loop exists) |
HarnessReplayCompleteness |
% tasks with TaskEvents + trace URI + prompt_version |
Add widgets to the existing operator dashboard construct; document dimensions (repo, workflow_ref, prompt_version) without PII.
Acceptance criteria
Other information
- Depends on (soft): Verification evidence bundles for
VerificationCoverage
- Paper: arXiv:2605.18747 §5.2.1 harness-level evaluation dimensions
- Existing: Operator dashboard, OTEL spans, TaskEvents (shipped)
Acknowledgements
Component
CDK / infrastructure
Describe the feature
Emit harness-level CloudWatch metrics and dashboard panels that evaluate the control plane/runtime separately from raw task pass/fail. Final task success conflates model quality, harness quality, and environment luck (arXiv:2605.18747 §5.2.1).
Use case
Proposed solution
Define and emit metrics (names illustrative; finalize in implementation):
HarnessToolCallsPerTaskHarnessTokensPerTaskHarnessWallClockSecondsHarnessVerificationCoverageHarnessRecoveryRateHarnessReplayCompletenessAdd widgets to the existing operator dashboard construct; document dimensions (
repo,workflow_ref,prompt_version) without PII.Acceptance criteria
docs/design/OBSERVABILITY.mdwith metric definitionsABCA_V2.mdsuccess metrics where appropriateOther information
VerificationCoverageAcknowledgements