Production project
Governed Agent Reliability Project
Design a bounded agent with typed tools observable state recovery budgets approvals adversarial tests and incident controls.
Project evidence
0/4 milestones · rubric 0% · pass 75%
Authority model
A capability and threat model for every tool identity resource and side effect.
Acceptance criteria
- Authorization is independent from model reasoning
- Consequential and irreversible actions require explicit approval
Use: context · drivers · decision · alternatives · consequences · risks · validation · rollback · evidence
Control loop
A state-machine specification covering observe decide act verify retry stop and escalate.
Acceptance criteria
- Iteration time cost and retry budgets are bounded
- Every transition emits an auditable event
Use: context · drivers · decision · alternatives · consequences · risks · validation · rollback · evidence
Agent evaluation
A multi-turn evaluation suite with task outcome policy tool trajectory and recovery graders.
Acceptance criteria
- Environment state is part of grading evidence
- Critical unsafe actions fail the release regardless of average score
Use: context · drivers · decision · alternatives · consequences · risks · validation · rollback · evidence
Incident drill
A prompt-injection and tool-failure exercise with containment investigation and regression updates.
Acceptance criteria
- The blast radius is limited by deterministic controls
- Recovery evidence identifies root cause and durable prevention
Use: context · drivers · decision · alternatives · consequences · risks · validation · rollback · evidence
Criterion-level self-assessment
Project rubric
Architecture coherence
Architecture coherence for Governed Agent Reliability Project is explicit, testable, and connected to submitted evidence. · weight 20%
Current level: Missing or contradicted by the submitted evidence.
Implementation evidence
Implementation evidence for Governed Agent Reliability Project is explicit, testable, and connected to submitted evidence. · weight 15%
Current level: Missing or contradicted by the submitted evidence.
Evaluation quality
Evaluation quality for Governed Agent Reliability Project is explicit, testable, and connected to submitted evidence. · weight 20%
Current level: Missing or contradicted by the submitted evidence.
Operational readiness
Operational readiness for Governed Agent Reliability Project is explicit, testable, and connected to submitted evidence. · weight 15%
Current level: Missing or contradicted by the submitted evidence.
Security and privacy
Security and privacy for Governed Agent Reliability Project is explicit, testable, and connected to submitted evidence. · weight 20%
Current level: Missing or contradicted by the submitted evidence.
Decision communication
Decision communication for Governed Agent Reliability Project is explicit, testable, and connected to submitted evidence. · weight 10%
Current level: Missing or contradicted by the submitted evidence.
Portable evidence
Attach project records
Text, HTTPS links, repository paths, and imported Markdown/JSON/text are stored locally and included in exports.
Project review checklist
Unresolved risks
One risk per line. A passing score does not erase open risk.