<aside> 🔄 Human review during pilot catches errors, builds training data, and establishes accuracy baselines before reducing oversight.

</aside>

Objective: Define the human oversight process during pilot and early deployment so the team catches errors, builds training data, and establishes accuracy baselines before reducing oversight.

When to use: Before pilot launch. Update thresholds as accuracy data accumulates. Reference this protocol whenever a tier transition is proposed.

Inputs: Identified AI decision types, initial accuracy expectations, risk tolerance from Sponsor, Hierarchy of Agency framework (if available from prior projects).

Position: Sentinel defines the protocol. Architect implements the review workflow. Sponsor approves tier transition criteria.

Review Process

  1. AI processes inputs and produces outputs
  2. Human reviews every output before action
  3. Human marks each output: Correct, Incorrect, or Judgment Call
  4. Incorrect outputs get documented with the correct answer
  5. Documented corrections become training data
  6. Track accuracy by category and over time

Oversight Reduction Thresholds

Tier 3: Full Review Required (starting point)