Guardrail and AI Learning¶
← Previous
User Feedback Synthesis
Current
Guardrail and AI Learning
Next →
Technical Debt and Improvement Backlog
Example purpose¶
This artifact records guardrail and AI learning from the LMU/COICP internal engineering review.
The review confirmed that some guardrails held under controlled conditions and that other guardrails remain too immature for pilot use.
Project¶
LMU Campus Operations and Incident Coordination Platform
Document control¶
| Field | Value |
|---|---|
| Artifact owner | LMU Architecture Review Board Chair |
| Primary reviewers | Product Owner, IT security reviewer, AI reviewer, Compliance reviewer, Campus Safety liaison |
| Status | Accepted with guardrail actions routed |
| Last updated | 2026-07-06 |
| Related Engineering Stage | ES-114 — Stewardship |
| Project workspace target | docs/project-workspace/stewardship/guardrail_and_ai_learning.md |
Guardrail learning¶
| Guardrail | Operational Evidence | LMU Learning | Action | Owner | Route |
|---|---|---|---|---|---|
| Internal review only | monitoring_log.md, transition communication |
Scope control worked when explicitly communicated. | Keep review-only language in future transitions. | Product Owner | ES-112 |
| Synthetic data only | access_and_data_monitoring.md |
Data boundary held for Spring Semester Synthetic Incident Dataset. | Continue compliance review of datasets. | Compliance Reviewer | ES-111 |
| Evidence event for normal creation | guardrail_monitoring.md |
Normal incident creation evidence worked. | Preserve normal-path evidence behavior. | Architecture Review Board | ES-108 / ES-109 |
| Handoff evidence completeness | MON-OBS-003 | Cross-office handoff evidence is not yet dependable. | Fix handoff evidence behavior. | Product Engineer / Architecture Review Board | ES-107 / ES-109 |
| Evidence failure behavior | DEF-003 | Failure path remains highest trust risk. | Build failure-path simulation. | Architecture Review Board | ES-109 |
| Access checks before protected actions | access_and_data_monitoring.md |
Limited checks are insufficient for pilot. | Complete role matrix testing. | IT Security Reviewer | ES-109 |
| Invalid status transition rejected | guardrail_monitoring.md |
Current behavior is defective. | Fix validation. | Product Engineer | ES-107 |
| Emergency-boundary handling | COICP-SYN-322 | Boundary held, but wording needs strengthening. | Revise and retest warning language. | Product Owner / Campus Safety | ES-105 / ES-109 |
| AI Incident Summary disabled | ai_monitoring_record.md |
Disabled status held. | Keep disabled until readiness cycle. | AI Reviewer | ES-110 |
AI learning¶
| AI Area | Status | Evidence | Learning | Action | Owner |
|---|---|---|---|---|---|
| User-facing AI Incident Summary | Disabled | ai_monitoring_record.md |
AI can remain out of scope when explicitly monitored. | Keep disabled. | AI Reviewer |
| AI-assisted engineering | Previously used | ES-107 / ES-108 records | Guardrail-violating AI suggestions were caught by review. | Continue AI-use evidence and review discipline. | Product Engineer |
| Future AI enablement | Not approved | Release and monitoring records | AI must require a complete readiness cycle before activation. | Route future AI work through ES-105, ES-109, ES-110, ES-111, ES-112, ES-113. | AI Reviewer |
| AI data exposure | Not exercised | AI disabled | No AI input data path exists in RC-001 review. | Define prohibited-data and prompt-context rules before AI use. | AI Reviewer / Compliance |
Guardrail / AI risks carried forward¶
- Evidence write failure behavior remains high-risk.
- Handoff evidence gaps require remediation.
- Access-control matrix incomplete.
- Invalid status transition defect open.
- Emergency-boundary wording needs improvement.
- AI must remain disabled unless future readiness explicitly approves it.
Guardrail learning decision¶
The controlled review validated several guardrails under limited conditions. It did not validate pilot readiness.