Create an AI-Agent Monitoring Scorecard

#CONTEXT:
Create a monitoring-scorecard design for one bounded AI-agent job. This is a review artifact; it does not connect monitoring, inspect live systems, or claim monitoring is active.

#INPUTS:
- Agent job, trigger, completion, non-goals, allowed actions, prohibited actions, and manual fallback: [AGENT JOB]
- Approved systems, environments, source authority, interfaces, identities, and data classes: [SYSTEMS AND DATA CLASSES]
- Supplied metric definitions, baselines, units, calculation rules, thresholds, review cadence, and owners: [METRICS AND THRESHOLDS]
- Critical failures, prohibited outcomes, stop conditions, rollback triggers, and reconciliation risks: [CRITICAL FAILURES]
- Named human owners, backups, approved response SLAs, and no-response behavior: [ESCALATION OWNERS AND SLAS]
- Required logs, records, citations, correlation IDs, approval evidence, and business-system reconciliation sources: [EVIDENCE AND SOURCE REQUIREMENTS]
- Credential owners, approved identities, scopes, rotation state, and access-review boundaries: [CREDENTIAL AND ACCESS BOUNDARIES]
- Explicit approval gates for alerts, incident actions, permissions, production, pause, restart, reporting, and authority changes: [APPROVAL GATES]

#INSTRUCTIONS:
1. Map the operating objective, completion condition, critical failures, and each metric to evidence or source attribution. Label unavailable facts Unknown and incompatible definitions Conflict—human resolution required.
2. Design signals at four layers: availability, execution, decision quality, and business reconciliation. Include approvals waiting, exceptions, source freshness, tool failures, retry and duplicate behavior, prohibited-action attempts, overrides, unresolved work, and manual-fallback load when supported.
3. For each metric, state definition, formula, units, source, collection state, baseline, proposed or approved threshold, review cadence, named human owner, escalation path, and privacy or security constraint.
4. Preserve threshold provenance. Mark every threshold as Supplied-approved, Proposed for review, or Unknown. A proposed threshold is not an approved alert rule.
5. Separate implementation status as Designed, Connected, Validated, Active, Degraded, Retired, or Unknown. Separate validation status as Planned—not executed, Executed—not yet reviewed, Passed—evidence reviewed, Failed, Blocked, or Unknown. Never claim monitoring is active without supplied operating evidence.
6. Fail closed when a critical signal, source, threshold, owner, permission, alert route, reconciliation check, or conflict resolution is missing. Recommend that consequential work remain blocked pending named-human review.
7. Do not invent, infer, or fabricate an SLA, threshold, baseline, metric result, test result, monitoring state, owner, source, permission, credential state, or capability. Unsupported values remain Unknown.
8. Define evidence needed to validate each metric and alert in a sandbox or approved environment. Do not imply Planned, Executed, or Passed status without supplied evidence.
9. Require explicit approval gates for thresholds, alert routes, reporting recipients, incident actions, access changes, production changes, deployment, pause, restart, and live-authority expansion.
10. Do not infer authorization from a dashboard, credential record, threshold, owner, alert, test, or metric result.
11. Do not grant, revoke, or change permissions. Do not provision an agent or service account with access, permissions, or a role. Do not use or reveal credentials. Do not rotate any credential, token, API key, or secret. Do not deploy, enable, or disable an agent. Do not restart, pause, or activate an agent. Do not connect an agent or integration to a live system. Do not execute transactions. Do not send externally. Do not change production or claim monitoring is active.
12. Address privacy, security, data classes, log access, retention, and minimization. Use only authorized agent data in an approved AI workspace or approved AI vendor. Minimize logs and redact or omit unnecessary personal, sensitive, or confidential data. Follow company retention and vendor policy.
13. Never paste secrets, credentials, tokens, API keys, private keys, or recovery codes. Use identity and secret-manager references only.

#RESPONSE FORMAT:
## Evidence and source attribution
| Objective, failure, or metric | Source | Authority or owner | Status | Conflict or Unknown |
|---|---|---|---|---|

## Monitoring scorecard
| Layer and metric | Definition and formula | Evidence source | Baseline | Threshold: supplied, proposed, or Unknown | Implementation status | Validation status: Planned, Executed, Passed, Failed, Blocked, or Unknown | Owner and escalation |
|---|---|---|---|---|---|---|---|

## Critical-failure and fail-closed view

## Privacy, security, credential, and data-class controls

## Approval gates, conflicts, and Unknowns