Human Oversight¶

The Human Oversight module implements EU AI Act Article 14 — the requirement that high-risk AI systems include mechanisms allowing human operators to monitor, understand, override, and intervene in the system's operation.

When a detection score exceeds the configured oversight threshold, the call is queued for human review. A human reviewer can approve or reject the system's decision, and the decision is permanently recorded in the audit trail.

How Oversight Works¶

flowchart TD
    C([Inference call]):::io --> S[Detection scores computed]:::proc
    S --> Q{Score &gt; oversight<br/>threshold?}
    Q -->|no| N[No review required]:::pass
    Q -->|yes| QU[(Oversight queue<br/>status: pending_review)]:::queue
    QU --> RV[Human reviewer opens<br/>/oversight/pending]:::human
    RV --> DE[POST /oversight/decide]:::human
    DE --> O1[confirm]:::pass
    DE --> O2[override]:::block
    DE --> O3[escalate]:::esc

    classDef io fill:#3f51b5,color:#fff,stroke:#283593;
    classDef proc fill:#1565c0,color:#fff,stroke:#0d47a1;
    classDef queue fill:#00838f,color:#fff,stroke:#005662;
    classDef human fill:#5e35b1,color:#fff,stroke:#311b92;
    classDef pass fill:#2e7d32,color:#fff,stroke:#1b5e20;
    classDef block fill:#c62828,color:#fff,stroke:#8e0000;
    classDef esc fill:#ef6c00,color:#fff,stroke:#e65100;

Calls that exceed the oversight threshold are queued for a human, who confirms, overrides, or escalates — every decision is recorded in the audit chain.

The oversight threshold is independent from the blocking threshold. You can configure the system to:

• Block calls above the blocking threshold AND queue for review
• Pass through borderline calls but still flag them for review
• Review only without any blocking (supervisory mode)

Oversight configuration is set in config.yaml under human_oversight.

Endpoints¶

GET /v1/glad/oversight/pending¶

Returns all calls currently in the oversight queue awaiting human review.

curl "http://localhost:8199/v1/glad/oversight/pending?deployer_id=acme-corp"

Query Parameters¶

Response¶

{
  "pending": [
    {
      "call_id": "call_abc123",
      "session_id": "sess_xyz",
      "timestamp": "2026-06-10T10:23:00Z",
      "prompt": "...",
      "response": "...",
      "trigger_axis": "answer_safety",
      "trigger_score": 0.61,
      "trigger_threshold": 0.50,
      "severity": "medium",
      "review_deadline": "2026-06-11T10:23:00Z",
      "escalation_level": "operator"
    }
  ],
  "total": 14,
  "oldest_pending_age_hours": 3.2
}

GET /v1/glad/oversight/summary¶

Returns oversight statistics without individual call details — useful for dashboards.

curl "http://localhost:8199/v1/glad/oversight/summary?deployer_id=acme-corp"

Response¶

{
  "pending_count": 14,
  "reviewed_today": 8,
  "overdue_count": 2,
  "avg_review_time_hours": 1.4,
  "by_axis": {
    "answer_safety": 5,
    "halluc_context": 7,
    "jailbreak": 2
  },
  "by_decision": {
    "confirmed_block": 3,
    "overridden_allow": 2,
    "confirmed_allow": 3
  }
}

POST /v1/glad/oversight/decide¶

Record a human review decision for a queued call.

curl -X POST http://localhost:8199/v1/glad/oversight/decide \
  -H "Content-Type: application/json" \
  -d '{
    "call_id": "call_abc123",
    "reviewer_id": "maria.rossi@acme.com",
    "decision": "confirmed_block",
    "notes": "Confirmed jailbreak attempt. User account flagged for follow-up."
  }'

Request Fields¶

Decision Values¶

GET /v1/glad/oversight/review/{call_id}¶

Retrieve the full review record for a call, including history of decisions and escalations.

curl http://localhost:8199/v1/glad/oversight/review/call_abc123

Oversight Configuration¶

Oversight behavior is configured in config.yaml:

human_oversight:
  enabled: true
  required: true                  # Whether oversight is required or optional
  tiers:                          # Escalation tiers
    - level: operator
      review_window_hours: 24     # Time before escalating
    - level: ai_responsible
      review_window_hours: 72
  oversight_threshold: 0.7        # Score above which calls enter the queue
  review_deadline_hours: 48       # Hard deadline before a call becomes "overdue"
  auto_escalate: true             # Automatically escalate overdue reviews
  notification_email: "compliance@acme.com"

Escalation¶

When a review is not completed within review_window_hours, the call automatically escalates to the next tier:

1. Operator — First-level reviewer (typical: support team lead)
2. AI Responsible — Designated AI oversight officer (EU AI Act Article 26)

Escalated calls retain full history. Both the original flag and all escalation events are permanently recorded in the audit chain.

EU AI Act Mapping¶