Step 3 of 4

Secure Workbench

Task completion with continuous verification

Iron Dome Security

The 3-3-3 System

Every batch of 10 failed prompts is evaluated by 9 experts: 3 from each tier. All work happens in our secure workbench with continuous monitoring to prevent AI contamination.

Live camera and screen sharing required
60-second heartbeat verification
Keystroke pattern analysis
Instant session lock on violations
Active Security Checks
Camera
Face detected ✓
Screen
Recording ✓
Heartbeat
42ms ✓
Keystrokes
Organic ✓
Current Task
Task #2847 • Python Code Generation
12:34 remaining
All activity monitored and recorded. Session auto-locks on security violations.

Security layers

Face detection

Every 60 seconds, we verify the expert's face is in frame. If they leave or someone else appears, the session locks immediately.

• Deepfake detection
• Multiple face detection
• Phone in frame detection
• Lighting quality checks

Keystroke analysis

We analyze typing patterns to detect copy-paste from AI tools. Organic human typing has distinct rhythm; AI pastes are instant.

• Typing speed patterns
• Copy-paste detection
• Burst vs steady typing
• Clipboard monitoring

Screen monitoring

Full screen capture ensures experts only use approved tools. Any attempt to access external AI tools results in instant lock.

• Window title detection
• Browser tab monitoring
• ChatGPT/Claude detection
• Screen recording

UI_OVERLAY_LOCK

When a violation is detected, the workbench immediately locks. The expert can no longer see the prompt or submit work. The task is reassigned to another expert.

Violations that trigger lock
• Face disappears from camera
• Multiple faces detected
• Phone detected in frame
• Copy-paste detected
• ChatGPT/Claude opened
• Screen share disconnected
Immediate consequences
• Prompt removed from view
• Submit button disabled
• Task voided (no payment)
• Quality score impacted
• Account flagged for review
Task reassignment
• Auto-queued for new expert
• Original expert notified
• Incident logged
• Client data protected
• Zero data leakage

How 3-3-3 works

Each batch of 10 failed prompts is distributed to 9 experts across 3 tiers. This multi-tier validation ensures quality while providing comparative data.

3
Tier 1 Experts
Gold Standard

Their corrections define the gold standard. Tier 2/3 answers are compared against them.

Sets the benchmark for quality scoring
3
Tier 2 Experts
Verified

Proven track record. Their answers are quality-checked against Tier 1 for accuracy.

Provides high-quality comparative data
3
Tier 3 Experts
Standard

New experts building their reputation. Provides baseline comparison data.

Establishes difficulty baseline
60s
Heartbeat interval
99.8%
Session completion
<0.1%
AI contamination
100%
Violations caught

Why continuous monitoring matters

Without monitoring, experts could use AI tools to generate corrections, defeating the entire purpose of human verification. Our Iron Dome security ensures every correction is genuinely human.

Without monitoring

  • ✗ Experts use ChatGPT for answers
  • ✗ No way to detect AI contamination
  • ✗ Training data quality unknown
  • ✗ Model learns from AI, not humans

With Iron Dome

  • ✓ 100% human-generated corrections
  • ✓ AI tool usage instantly detected
  • ✓ Guaranteed data authenticity
  • ✓ True human intelligence captured