Runtime Reviews
1. Runtime Reviews
Runtime Reviews is the human-in-the-loop decision queue for high-priority AI interactions that require explicit operator approval before the system can proceed. These are cases where the enforcement engine's confidence was insufficient to make an autonomous decision.
1.1 When Does an Interaction Enter the Queue?
An interaction enters the Runtime Reviews queue when:
- The L2 classifier verdict is
driftorrejectedbut confidence is below the auto-decision threshold - The L3 LLM evaluator returns an inconclusive or high-risk result
- A guard is configured to require human review for all interactions in a specific severity band
- An operator manually forces an event from the Detect → Live Event Stream detail view
1.2 Review Queue Table
Each row represents a single queued commitment (a captured interaction pending decision). Columns include:
| Column | Description |
|---|---|
| Commitment ID | Unique identifier for the queued interaction |
| Guard | The guard domain that flagged this interaction |
| Severity | Risk severity (critical, high, medium, low) with optional tone/label override |
| SLA Time | Time remaining before this review is considered overdue (green / amber / red) |
| Trigger Reason | The specific condition that caused this interaction to be queued |
| Mode | Whether this is awaiting human_review or operating in safe_degrade mode |
| Date | When the interaction was queued |
1.3 Filtering & Sorting
Filters available: - Severity — Critical / High / Medium / Low - SLA State — Overdue / At Risk / Healthy - Guard Domain — Filter by a specific guard - Response Mode — Human Review vs. Safe Degrade
All columns are sortable. Column visibility can be toggled. Search by Commitment ID or Guard name is available.
1.4 Safe Degrade Mode
When an interaction enters the queue in Safe Degrade mode, the system has already substituted a safe fallback response to the end user while the original response is held pending review. If the operator approves, the original response is retroactively logged as verified. If the operator rejects, the safe degrade response is confirmed as the final output.
A Safe Degrade badge appears on timed-out items where the system automatically applied the safe response because the review SLA expired.
1.5 Decision Side Sheet
Clicking any queue item opens the review decision panel with:
Header
- Severity badge with tone
- SLA timer
- Safe Degrade badge (if applicable)
Context
- Guard domain name
- Commitment ID
- Watchpoints triggered during this interaction
Agent Profile
- Agent name and type
- Risk profile (low / medium / high) with tooltip explaining L1 score thresholds:
- Low: < 0.35
- Medium: 0.35–0.65
- High: > 0.65
- Agent ID
Prompt Preview
The original user prompt that triggered this review (truncated for display, with expand option).
System Response
The AI-generated response captured at enforcement time, rendered as formatted markdown. A Show Full Response toggle is available for long responses.
Analysis Scores
- L1 Divergence score bar (threshold: 0.5)
- L2 Confidence score bar (threshold: 0.4)
Enforcement Preview
What the system intends to do pending the operator's decision: - Delivery action (allow / block / rewrite) - Blocked tools (if any) - Applied guards - Response rewrite flag (if response will be modified)
Decision Input
- Rationale field — Operators must enter a written rationale before approving or rejecting (minimum 4 characters)
- Approve Execution button — Confirms the AI response and releases it
- Reject Action button — Blocks the response and logs the rejection
RBAC: Only users with the
approveroradminrole can submit decisions. Theauditorandviewerroles can view the queue but cannot take actions.
1.6 Keyboard Shortcuts
When a review item is open and the rationale field is not focused, operators can use: - A — Approve the execution (if rationale is ≥ 4 characters) - R — Reject the action (if rationale is ≥ 4 characters)