**Pedalgo-Style Human Review Team Setup (Sweden Focus)**

A Pedalgo-style system performs **triage only**. The AI assigns a risk score that reorders the human review queue. All final decisions, including whether any action is taken, remain exclusively with trained human reviewers. No automated enforcement, public naming, or vigilante measures are permitted.

### Core Principles for Any Setup
- AI output is a risk signal, never a determination of guilt or offense.
- All cases use a single universal base rate; no demographic or group weighting is applied.
- Every escalation follows lawful channels only (Swedish Police, NCMEC CyberTipline, IWF, or DSA trusted flagger routes).
- Synthetic examples are used exclusively for training and documentation.

### Required Training for Reviewers
Reviewers must complete a structured program before accessing any queue:

- Swedish criminal law on child sexual abuse material and grooming (Polisen/Barnahus materials)
- Digital Services Act (DSA) obligations and trusted flagger procedures
- Recognition of synthetic versus real content
- Trauma-informed review practices and secondary trauma prevention
- Inter-rater calibration exercises using synthetic cases
- Whistleblower rights and internal reporting channels

Minimum initial training: 40 hours classroom + 20 hours supervised queue time. Annual refresher training of 16 hours is required.

### Inter-Rater Reliability
Teams maintain reliability through:

- Weekly calibration sessions on synthetic cases
- Blind double-review of 10–15 % of the queue
- Cohen’s kappa target ≥ 0.75 on escalation decisions
- Disagreement resolution by a senior reviewer (never by AI score)

Low reliability triggers additional training before the reviewer returns to the queue.

### Logging Requirements
Every reviewed item must generate an auditable record containing:

- Unique case ID (no personal data in logs)
- AI triage score and features used (for transparency)
- Reviewer ID and timestamp
- Decision and rationale (standardized categories only)
- Escalation path chosen (if any)
- Retention period compliant with Swedish data protection rules

Logs are stored separately from any platform user data and are subject to internal audit and, where required, supervisory authority inspection.

### Handling High Volume of False Positives
High false-positive rates are expected in triage systems. Mitigation measures include:

- Tiered review: only highest-scoring items enter the human queue
- Time-boxed review (e.g., 90 seconds per item) to prevent reviewer fatigue
- Regular threshold tuning based on reviewer feedback, never on individual cases
- Clear “no action” outcome category with minimal logging
- Rotation of reviewers to maintain decision quality

### Escalation Paths
1. **Swedish Police** – direct report via established child protection units or Barnahus cooperation.
2. **NCMEC CyberTipline** – for material hosted outside Sweden or requiring international coordination.
3. **IWF** – for confirmed child sexual abuse material URLs.
4. **DSA trusted flagger channel** – when the platform holds trusted flagger status.

Escalation occurs only after human confirmation. The AI score is never cited as the sole justification.

### Whistleblower Protections
Reviewers are protected under the Swedish Whistleblower Act (2021:890). Organizations must:

- Provide an internal reporting channel independent of line management
- Guarantee anonymity options and prohibition of retaliation
- Include explicit protection language in reviewer contracts
- Offer access to external support (e.g., trade union or Barnombudsmannen resources)

### Sample SOP Structure
**SOP-PED-001: Triage Review Workflow**

1. AI generates daily prioritized queue (score only).
2. Reviewer opens next item in queue.
3. Reviewer assesses content against standardized criteria.
4. Decision options: No action / Further internal review / Escalate to authority.
5. All decisions logged within 24 hours.
6. Immediate escalation triggers (see red flags) bypass normal queue order.
7. Weekly team review of reliability metrics and threshold performance.

### Red Flags Requiring Immediate Escalation
- Clear depiction of a child in sexual context (synthetic or real)
- Explicit grooming language directed at a minor
- Threats of imminent harm to a child
- Evidence of ongoing live-streaming or real-time interaction
- Any case involving a Swedish minor where location data suggests immediate risk

In all such cases the reviewer escalates directly to Swedish Police or NCMEC without completing the normal review cycle.

This framework keeps the AI strictly in a prioritization role while ensuring lawful, auditable, and trauma-aware human decision-making.