MIZ OKI Media · Pilot guide

The 90-Day Causal Growth Control Pilot

Begin observe-only. Advance to human-approved actions only after evidence quality, economics, and policy boundaries are validated.

Scope — deliberately narrow

• One brand or business unit · limited media and commerce sources
• One performance-incident decision class + one budget-allocation or profit decision class
• Explicit evidence boundaries · explicit human authority · observe-only starting mode
• Measured predicted-versus-realized outcomes · clear expansion criteria

Phase 1 · Days 1–30 · Connect and observe

• Connect the scoped systems; confirm source access
• Validate identities, timestamps, and economics
• Establish canonical event definitions; map critical entities and constraints
• Run observe-only — no recommendations, no actions

Governance checkpoint: evidence-quality review. The pilot does not advance until the evidence base is trusted.

Phase 2 · Days 31–60 · Diagnose and recommend

• Generate causal recommendations for the scoped decision classes
• Evaluate counterfactuals, including the no-action baseline
• Measure and report confidence on every recommendation
• Validate every candidate against configured policy
• Require human approval on everything — no exceptions in this phase

Governance checkpoint: recommendation-quality review — diagnoses scored against ground truth before action authority is discussed.

Phase 3 · Days 61–90 · Approve, act, and learn

• Enable limited governed actions within explicit bounds; preserve rollback paths
• Measure predicted versus realized outcomes
• Produce the executive scorecard
• Recommend an expansion roadmap — or recommend stopping

Governance checkpoint: exit review against the criteria below.

Success metrics

Agreed before day one, measured throughout, reported without editing: evidence completeness and quality · time-to-detection · diagnostic accuracy vs. ground truth · recommendation acceptance rate · policy-check coverage and veto record · prediction accuracy on acted decisions · decision latency. Targets are set with your team during scoping.

Exit criteria — three explicit states

• Expand — the scorecard clears the agreed thresholds; the roadmap names the next decision classes, sources, and authority grants
• Extend — promising but incomplete; a defined extension with named gaps and a firm end date
• Stop — the record does not justify continuing; you keep the scorecard, the decision records, and everything the pilot learned

The pilot's product is evidence about your own decisions — whichever exit you choose, you keep it.