Design-vintage (late 2025). The "Active" status column below predates the governance spine and is not a current-state claim. Current design of record for bandit arms, causal bidding and pacing: docs/design/AD_CONTROL_PLANE_BANDIT_CAUSAL_BIDDING_DESIGN.md (PROPOSED, 2026-09-02); the pacing rule's executable home is services/service-policy-engine/pacing_veto.py, wired into the policy engine's evaluate() behind PACING_VETO — a flag that ships off (source literal "false", production env absent), so the veto is dark and nothing on this page is a current-state claim.

Ad Control Plane Strategy: Bandits, Causal Bidding & Pacing Safeguards

Strategic Context

As of late 2025, the MIZ OKI Ad Control Plane is pivoting from simple heuristics to a structured learning framework that balances adaptive online learning (RL/Bandits) with rigorous causal measurement and safety.

This strategy governs the interaction between:

• Cell 3 (KG Brain): Causal inference and strategy selection.
• Cell 11 (Decide ADC): Decision engines (Bandits/RL).
• Cell 15 (Act ADC): Execution and safety guardrails.

The Deployment Checklist (Governing Policy)

All automated bidding implementations must adhere to this 5-step safety progression:

1. Simple Bandit Policies (Phase 1)

• Mechanism: Use Contextual Bandits (e.g., Thompson Sampling, LinUCB) for creative/segment allocation.
• Why: Provides safe, interpretable exploration with theoretical regret bounds.
• Implementation: Current FunnelOptimizer -> DecideADC implementation.

2. Gated RL with Offline Evaluation (Phase 2)

• Mechanism: Deep RL policies must be evaluated offline ("shadow runs") before controlling live budgets.
• Constraint: RL agents cannot directly control production bids without passing a "Safety Gate" (Actionable by ReasonADC).

3. Causal Uplift Experiments (Phase 3)

• Mechanism: Run randomized controlled trials (RCTs) or geo-splits to estimate true incremental lift.
• Integration: Feed uplift results into CausalService to calibrate the "Reward" signal for Bandits/RL.
• Goal: Optimize for Incremental ROAS, not just attribution-based ROAS.

4. Uplift-to-Bid Translation

• Formula: Bid = Base_Bid * f(Uplift_Probability)
• Requirement: Bids must be shaped by the causal probability of conversion (p(conv|ad) - p(conv|no_ad)).

5. Pacing Safeguards (The "Kill Switch")

• Mechanism: Strict hourly/daily budget pacing controllers.
• Rule: If spend velocity > 200% of target for > 15 mins, trigger Halt.
• Responsibility: ActADC (Cell 15) must enforce these hard constraints regardless of upstream RL commands.

Architecture Alignment

Implementation Priority

1. Verify Pacing in ActADC: Ensure ActADC has hard-coded budget velocity checks.
2. Refine Uplift Signal: Update CausalService to output an explicit uplift_multiplier for bidding.
3. Shadow Mode RL: Configure DecideADC to support a "shadow" mode for testing deeper RL policies without execution.